Add Batch c8308103-947b-46cb-8b35-3efd04bd836e data

.gitattributes

@@ -2471,3 +2471,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/VLOGGER_[[:space:]]Multimodal[[:space:]]Diffusion[[:space:]]for[[:space:]]Embodied[[:space:]]Avatar[[:space:]]Synthesis/e32a2703-b32a-4a6f-b45c-33f8b72e19be_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/VLog_[[:space:]]Video-Language[[:space:]]Models[[:space:]]by[[:space:]]Generative[[:space:]]Retrieval[[:space:]]of[[:space:]]Narration[[:space:]]Vocabulary/5cce3931-2141-492f-94ca-887e72e0b54b_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/VLsI_[[:space:]]Verbalized[[:space:]]Layers-to-Interactions[[:space:]]from[[:space:]]Large[[:space:]]to[[:space:]]Small[[:space:]]Vision[[:space:]]Language[[:space:]]Models/d257a4d1-091b-48d7-b804-775fddd94e7e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VODiff_[[:space:]]Controlling[[:space:]]Object[[:space:]]Visibility[[:space:]]Order[[:space:]]in[[:space:]]Text-to-Image[[:space:]]Generation/3dac837e-c8db-4c3a-a872-8c96ea7194f5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VSNet_[[:space:]]Focusing[[:space:]]on[[:space:]]the[[:space:]]Linguistic[[:space:]]Characteristics[[:space:]]of[[:space:]]Sign[[:space:]]Language/1135c1a4-3af0-45ac-a497-2c9557104d8c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VTON[[:space:]]360_[[:space:]]High-Fidelity[[:space:]]Virtual[[:space:]]Try-On[[:space:]]from[[:space:]]Any[[:space:]]Viewing[[:space:]]Direction/4af7cfba-2431-4854-85f1-f3f911dcc811_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VTON-HandFit_[[:space:]]Virtual[[:space:]]Try-on[[:space:]]for[[:space:]]Arbitrary[[:space:]]Hand[[:space:]]Pose[[:space:]]Guided[[:space:]]by[[:space:]]Hand[[:space:]]Priors[[:space:]]Embedding/00462254-bb4b-4c24-9d0a-9a8f2054757a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/V^2Dial_[[:space:]]Unification[[:space:]]of[[:space:]]Video[[:space:]]and[[:space:]]Visual[[:space:]]Dialog[[:space:]]via[[:space:]]Multimodal[[:space:]]Experts/a2fbca87-c18a-41dc-864b-f66927e9f633_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Variance-Based[[:space:]]Membership[[:space:]]Inference[[:space:]]Attacks[[:space:]]Against[[:space:]]Large-Scale[[:space:]]Image[[:space:]]Captioning[[:space:]]Models/26bae7f8-8e37-4862-840e-6578a63d1d11_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VasTSD_[[:space:]]Learning[[:space:]]3D[[:space:]]Vascular[[:space:]]Tree-state[[:space:]]Space[[:space:]]Diffusion[[:space:]]Model[[:space:]]for[[:space:]]Angiography[[:space:]]Synthesis/221d762b-beb5-4562-9aa6-3ebe0030c356_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VerbDiff_[[:space:]]Text-Only[[:space:]]Diffusion[[:space:]]Models[[:space:]]with[[:space:]]Enhanced[[:space:]]Interaction[[:space:]]Awareness/ad57dfe8-0121-4da2-a044-841b75f6e661_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViCaS_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Combining[[:space:]]Holistic[[:space:]]and[[:space:]]Pixel-level[[:space:]]Video[[:space:]]Understanding[[:space:]]using[[:space:]]Captions[[:space:]]with[[:space:]]Grounded[[:space:]]Segmentation/ae891128-2b76-44ec-8a3e-060beba26fb9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViKIENet_[[:space:]]Towards[[:space:]]Efficient[[:space:]]3D[[:space:]]Object[[:space:]]Detection[[:space:]]with[[:space:]]Virtual[[:space:]]Key[[:space:]]Instance[[:space:]]Enhanced[[:space:]]Network/c314dea2-42fe-4d4d-8bb4-7010c813e409_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViUniT_[[:space:]]Visual[[:space:]]Unit[[:space:]]Tests[[:space:]]for[[:space:]]More[[:space:]]Robust[[:space:]]Visual[[:space:]]Programming/8d692184-2a4f-4cca-806a-220832f3b327_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vid2Avatar-Pro_[[:space:]]Authentic[[:space:]]Avatar[[:space:]]from[[:space:]]Videos[[:space:]]in[[:space:]]the[[:space:]]Wild[[:space:]]via[[:space:]]Universal[[:space:]]Prior/61a0578e-e851-4106-a096-325f4e7911db_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vid2Sim_[[:space:]]Generalizable,[[:space:]]Video-based[[:space:]]Reconstruction[[:space:]]of[[:space:]]Appearance,[[:space:]]Geometry[[:space:]]and[[:space:]]Physics[[:space:]]for[[:space:]]Mesh-free[[:space:]]Simulation/48cd44e1-bff7-400d-aceb-8a86fa95ccb5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vid2Sim_[[:space:]]Realistic[[:space:]]and[[:space:]]Interactive[[:space:]]Simulation[[:space:]]from[[:space:]]Video[[:space:]]for[[:space:]]Urban[[:space:]]Navigation/a9b54426-b8df-4fc1-983c-717562672544_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VidBot_[[:space:]]Learning[[:space:]]Generalizable[[:space:]]3D[[:space:]]Actions[[:space:]]from[[:space:]]In-the-Wild[[:space:]]2D[[:space:]]Human[[:space:]]Videos[[:space:]]for[[:space:]]Zero-Shot[[:space:]]Robotic[[:space:]]Manipulation/c9488d81-95eb-4b83-bf3b-3ddc8f3b7707_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VidComposition_[[:space:]]Can[[:space:]]MLLMs[[:space:]]Analyze[[:space:]]Compositions[[:space:]]in[[:space:]]Compiled[[:space:]]Videos_/6c9efc65-d2ba-48e3-a8d2-fe7c5c1d7646_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VidHalluc_[[:space:]]Evaluating[[:space:]]Temporal[[:space:]]Hallucinations[[:space:]]in[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Video[[:space:]]Understanding/c6c9e644-f51f-4e6b-aa79-2d374850db15_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VidMuse_[[:space:]]A[[:space:]]Simple[[:space:]]Video-to-Music[[:space:]]Generation[[:space:]]Framework[[:space:]]with[[:space:]]Long-Short-Term[[:space:]]Modeling/e0ce05e4-9f59-4588-812e-59896766b6d3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VidSeg_[[:space:]]Training-free[[:space:]]Video[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]based[[:space:]]on[[:space:]]Diffusion[[:space:]]Models/a8a75be1-b6e5-46b8-a340-187193161579_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VidTwin_[[:space:]]Video[[:space:]]VAE[[:space:]]with[[:space:]]Decoupled[[:space:]]Structure[[:space:]]and[[:space:]]Dynamics/396b7efb-b08c-4897-8d56-4437ed850d42_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video[[:space:]]Depth[[:space:]]Anything_[[:space:]]Consistent[[:space:]]Depth[[:space:]]Estimation[[:space:]]for[[:space:]]Super-Long[[:space:]]Videos/3d97efd2-8c6d-451d-8d75-d43ab643dfda_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video[[:space:]]Depth[[:space:]]without[[:space:]]Video[[:space:]]Models/adc6ac4e-2f14-4dd6-b43a-b385177519a9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video[[:space:]]Language[[:space:]]Model[[:space:]]Pretraining[[:space:]]with[[:space:]]Spatio-temporal[[:space:]]Masking/592a1785-c0f4-44c0-ad31-218ccb7da263_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video[[:space:]]Motion[[:space:]]Transfer[[:space:]]with[[:space:]]Diffusion[[:space:]]Transformers/f0f4362e-9924-4b87-aeff-45869e95fb16_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video[[:space:]]Summarization[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/49bb8211-5bbe-4fd6-a3e5-e9eae1a8e8d9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video-3D[[:space:]]LLM_[[:space:]]Learning[[:space:]]Position-Aware[[:space:]]Video[[:space:]]Representation[[:space:]]for[[:space:]]3D[[:space:]]Scene[[:space:]]Understanding/c7916639-0a81-44c5-a96b-970d900fe738_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video-Bench_[[:space:]]Human-Aligned[[:space:]]Video[[:space:]]Generation[[:space:]]Benchmark/b48e59c4-b36d-49ef-a7ca-02fe49c0d380_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video-ColBERT_[[:space:]]Contextualized[[:space:]]Late[[:space:]]Interaction[[:space:]]for[[:space:]]Text-to-Video[[:space:]]Retrieval/6c2b31d1-930f-426e-b6df-3bce096a9320_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video-Guided[[:space:]]Foley[[:space:]]Sound[[:space:]]Generation[[:space:]]with[[:space:]]Multimodal[[:space:]]Controls/27ba25cd-566f-43d3-badb-af29c2dceb60_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video-MME_[[:space:]]The[[:space:]]First-Ever[[:space:]]Comprehensive[[:space:]]Evaluation[[:space:]]Benchmark[[:space:]]of[[:space:]]Multi-modal[[:space:]]LLMs[[:space:]]in[[:space:]]Video[[:space:]]Analysis/8df19168-ac1f-43c7-9bef-05050364c876_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video-Panda_[[:space:]]Parameter-efficient[[:space:]]Alignment[[:space:]]for[[:space:]]Encoder-free[[:space:]]Video-Language[[:space:]]Models/8e708f5e-f4c2-49cb-b40d-85c403ad2f23_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video-XL_[[:space:]]Extra-Long[[:space:]]Vision[[:space:]]Language[[:space:]]Model[[:space:]]for[[:space:]]Hour-Scale[[:space:]]Video[[:space:]]Understanding/410e6e98-5f80-4e5e-9672-9d5b8fa4130a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoAutoArena_[[:space:]]An[[:space:]]Automated[[:space:]]Arena[[:space:]]for[[:space:]]Evaluating[[:space:]]Large[[:space:]]Multimodal[[:space:]]Models[[:space:]]in[[:space:]]Video[[:space:]]Analysis[[:space:]]through[[:space:]]User[[:space:]]Simulation/5a43a242-bd5d-4961-aed4-ae68b47ff795_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoComp_[[:space:]]Advancing[[:space:]]Fine-Grained[[:space:]]Compositional[[:space:]]and[[:space:]]Temporal[[:space:]]Alignment[[:space:]]in[[:space:]]Video-Text[[:space:]]Models/85b1b586-c77a-4209-812f-b601b6870f0d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoDPO_[[:space:]]Omni-Preference[[:space:]]Alignment[[:space:]]for[[:space:]]Video[[:space:]]Diffusion[[:space:]]Generation/2f70dda9-0787-49af-abce-676cbf00394e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoDirector_[[:space:]]Precise[[:space:]]Video[[:space:]]Editing[[:space:]]via[[:space:]]Text-to-Video[[:space:]]Models/8f2614cf-ca8d-409c-94f9-479e72f1f29f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoEspresso_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Chain-of-Thought[[:space:]]Dataset[[:space:]]for[[:space:]]Fine-Grained[[:space:]]Video[[:space:]]Reasoning[[:space:]]via[[:space:]]Core[[:space:]]Frame[[:space:]]Selection/eff006bb-cde1-4b27-a0ef-89b5fd58945a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoGEM_[[:space:]]Training-free[[:space:]]Action[[:space:]]Grounding[[:space:]]in[[:space:]]Videos/0a3fbcb1-73fe-47da-ac43-25ceb4ef9ba0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoGLaMM[[:space:]]_[[:space:]]A[[:space:]]Large[[:space:]]Multimodal[[:space:]]Model[[:space:]]for[[:space:]]Pixel-Level[[:space:]]Visual[[:space:]]Grounding[[:space:]]in[[:space:]]Videos/a7f6c402-bb8e-45ec-b9bf-5b07a3082f52_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoGigaGAN_[[:space:]]Towards[[:space:]]Detail-rich[[:space:]]Video[[:space:]]Super-Resolution/52ec8084-0006-41dd-a7b0-3a3880698e67_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoGuide_[[:space:]]Improving[[:space:]]Video[[:space:]]Diffusion[[:space:]]Models[[:space:]]without[[:space:]]Training[[:space:]]Through[[:space:]]a[[:space:]]Teacher's[[:space:]]Guide/8fb0f104-e76b-45ee-932c-212b94049d7b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoHandles_[[:space:]]Editing[[:space:]]3D[[:space:]]Object[[:space:]]Compositions[[:space:]]in[[:space:]]Videos[[:space:]]Using[[:space:]]Video[[:space:]]Generative[[:space:]]Priors/2cf37ed5-ea0a-4c05-acd5-0c34fb8789ea_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoICL_[[:space:]]Confidence-based[[:space:]]Iterative[[:space:]]In-context[[:space:]]Learning[[:space:]]for[[:space:]]Out-of-Distribution[[:space:]]Video[[:space:]]Understanding/3dc9da78-cea1-471c-b4db-ee52c6853efe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoMage_[[:space:]]Multi-Subject[[:space:]]and[[:space:]]Motion[[:space:]]Customization[[:space:]]of[[:space:]]Text-to-Video[[:space:]]Diffusion[[:space:]]Models/80815955-c6e2-4c64-a2b6-3555fa9f4dc2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoRefer[[:space:]]Suite_[[:space:]]Advancing[[:space:]]Spatial-Temporal[[:space:]]Object[[:space:]]Understanding[[:space:]]with[[:space:]]Video[[:space:]]LLM/1eb2e581-034c-4440-aa2e-5d6157a422fc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoSPatS_[[:space:]]Video[[:space:]]SPatiotemporal[[:space:]]Splines[[:space:]]for[[:space:]]Disentangled[[:space:]]Occlusion,[[:space:]]Appearance[[:space:]]and[[:space:]]Motion[[:space:]]Modeling[[:space:]]and[[:space:]]Editing/47539518-9a92-424c-830b-8495e559a706_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoScene_[[:space:]]Distilling[[:space:]]Video[[:space:]]Diffusion[[:space:]]Model[[:space:]]to[[:space:]]Generate[[:space:]]3D[[:space:]]Scenes[[:space:]]in[[:space:]]One[[:space:]]Step/3d9c7184-8bc7-4684-b4b3-dee716e5015a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoTree_[[:space:]]Adaptive[[:space:]]Tree-based[[:space:]]Video[[:space:]]Representation[[:space:]]for[[:space:]]LLM[[:space:]]Reasoning[[:space:]]on[[:space:]]Long[[:space:]]Videos/40b3e4b5-66bb-4e61-88c1-0c383821b986_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoWorld_[[:space:]]Exploring[[:space:]]Knowledge[[:space:]]Learning[[:space:]]from[[:space:]]Unlabeled[[:space:]]Videos/d7364fb4-c126-4443-875f-c1f7daf2bf13_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Viewpoint[[:space:]]Rosetta[[:space:]]Stone_[[:space:]]Unlocking[[:space:]]Unpaired[[:space:]]Ego-Exo[[:space:]]Videos[[:space:]]for[[:space:]]View-invariant[[:space:]]Representation[[:space:]]Learning/ae8678aa-d4cb-4d54-aa26-b58e33e1e461_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViiNeuS_[[:space:]]Volumetric[[:space:]]Initialization[[:space:]]for[[:space:]]Implicit[[:space:]]Neural[[:space:]]Surface[[:space:]]Reconstruction[[:space:]]of[[:space:]]Urban[[:space:]]Scenes[[:space:]]with[[:space:]]Limited[[:space:]]Image[[:space:]]Overlap/23124bd5-ecfd-490d-a48b-e03c3ec06af3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VinTAGe_[[:space:]]Joint[[:space:]]Video[[:space:]]and[[:space:]]Text[[:space:]]Conditioning[[:space:]]for[[:space:]]Holistic[[:space:]]Audio[[:space:]]Generation/a74d1a60-3b5d-4c9c-a324-7df50d716e1c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VinaBench_[[:space:]]Benchmark[[:space:]]for[[:space:]]Faithful[[:space:]]and[[:space:]]Consistent[[:space:]]Visual[[:space:]]Narratives/e5948aac-c2d3-4df8-a7f3-82cb99be6728_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision-Guided[[:space:]]Action_[[:space:]]Enhancing[[:space:]]3D[[:space:]]Human[[:space:]]Motion[[:space:]]Prediction[[:space:]]with[[:space:]]Gaze-informed[[:space:]]Affordance[[:space:]]in[[:space:]]3D[[:space:]]Scenes/9b42c369-4ef9-4c45-9192-f6b0a17d8bda_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision-Language[[:space:]]Embodiment[[:space:]]for[[:space:]]Monocular[[:space:]]Depth[[:space:]]Estimation/959b3b25-9f17-45a6-aab1-48ac53aa864f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision-Language[[:space:]]Gradient[[:space:]]Descent-driven[[:space:]]All-in-One[[:space:]]Deep[[:space:]]Unfolding[[:space:]]Networks/bd772a11-bc94-4fb3-8aee-976fd29748e0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision-Language[[:space:]]Model[[:space:]]IP[[:space:]]Protection[[:space:]]via[[:space:]]Prompt-based[[:space:]]Learning/e2d53e64-2ba0-4eec-9a41-e03d1858d841_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision-Language[[:space:]]Models[[:space:]]Do[[:space:]]Not[[:space:]]Understand[[:space:]]Negation/f44bf4ea-b507-4a1f-8e9b-87600bcf741b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VisionArena_[[:space:]]230k[[:space:]]Real[[:space:]]World[[:space:]]User-VLM[[:space:]]Conversations[[:space:]]with[[:space:]]Preference[[:space:]]Labels/f703962d-2329-4c48-bd31-19be8745b28f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VisionPAD_[[:space:]]A[[:space:]]Vision-Centric[[:space:]]Pre-training[[:space:]]Paradigm[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/4e6d5045-e8c8-4646-a828-f050db70849c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VisionZip_[[:space:]]Longer[[:space:]]is[[:space:]]Better[[:space:]]but[[:space:]]Not[[:space:]]Necessary[[:space:]]in[[:space:]]Vision[[:space:]]Language[[:space:]]Models/d6a76b28-eeb8-41ac-bee1-8e5731b2110b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Visual[[:space:]]Agentic[[:space:]]AI[[:space:]]for[[:space:]]Spatial[[:space:]]Reasoning[[:space:]]with[[:space:]]a[[:space:]]Dynamic[[:space:]]API/18209ccb-89a4-4fe8-8dec-31f34f20c734_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Visual[[:space:]]Consensus[[:space:]]Prompting[[:space:]]for[[:space:]]Co-Salient[[:space:]]Object[[:space:]]Detection/3f1a6dd0-b317-40d4-8fff-a7533b787af0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Visual[[:space:]]Lexicon_[[:space:]]Rich[[:space:]]Image[[:space:]]Features[[:space:]]in[[:space:]]Language[[:space:]]Space/18fabbbb-6482-437e-bf79-727cc6d61f01_origin.pdf filter=lfs diff=lfs merge=lfs -text

2025/VODiff_ Controlling Object Visibility Order in Text-to-Image Generation/3dac837e-c8db-4c3a-a872-8c96ea7194f5_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:50ba22015d009bd54a9e03f4e3e884a4c751b8406520b67b7e77ca9367dcfcb6
+size 6456123

2025/VODiff_ Controlling Object Visibility Order in Text-to-Image Generation/full.md

@@ -0,0 +1,459 @@
+# VODiff: Controlling Object Visibility Order in Text-to-Image Generation
+
+Dong Liang $^{1,2}$  Jinyuan Jia $^{1,3,\dagger}$  Yuhao Liu $^{2,\dagger}$  Zhanghan Ke $^{2}$  Hongbo Fu $^{4}$  Rynson W.H. Lau $^{2,\dagger}$ $^{1}$ Tongji University  $^{2}$ City University of Hong Kong  $^{3}$ HKUST(GZ)  $^{4}$ HKUST
+
+sse_liangdong@tongji.edu.cn, jinyuanjia@hkust-gz.edu.cn, yuhaoliu7456@gmail.com, zhanghake2-c@my.cityu.edu.hk, hongbofu@ust.hk, Rynson.Lau@cityu.edu.hk
+
+![](images/605eb3a26f0b7f7ec9bc5363a4f2b3c1fb8f14102a074ea03cac17478070f9e8.jpg)  
+(a) Layout+Order
+
+(b) SD 1.5 [55]  
+![](images/396ae77a63730ad6b504e4a40fbeb46cb8acba828a8beb07b5bc76d08f7ca37b.jpg)  
+"A picture of a table, with a mug and a cup on it, the mug is on a dish."
+
+![](images/90d8390c805022ce98b0af3bd8a069e40e530d985c5c3495f8e28953b43c651b.jpg)  
+(c) GLIGEN [30]
+
+![](images/9156d041384b67c029095cc08a48ac276a6775d1663f900496069dd8f81bacb8.jpg)  
+(d) DenseDiffusion [27]
+
+![](images/3ce347d46bf265f8084f29a7a91ae70c3642042c6d1324a0846eac8aae94d125.jpg)  
+(e) MIGC [83]
+
+![](images/2e1e30381a2348527e12af9930f5f690daeb7d743f2ff622f2309da400d2597f.jpg)  
+(f) VODiff (Ours)
+
+![](images/7d2c4545c2adaf30a0ee372b5f280c38b4cfbe7e01ad5f5c7824507005d7b049.jpg)
+
+![](images/d002f4801e9720ae91a0bb523b77c2a4d91a051492d21f0091802344fc7b52a5.jpg)  
+"A suitcase with a banana in front of it and a dog behind it."
+
+![](images/74686a116326c0a5784c05a7b5fd42838d1ce6c0c438688e1adcf37242e2f0e1.jpg)
+
+![](images/2c8395177c695dd085d36e42b5856418d2e985a9761daf6162d4040bdd94d995.jpg)
+
+![](images/ad07d422a8159eb26d5d54266d112d32414f2bd332d94e9767163a9e3dc248b6.jpg)
+
+![](images/0e8847e8ce613ebf337399df08f281f5452a53b69c60edaf0d07a85bf2af47d3.jpg)
+
+Figure 1. Existing T2I methods that rely on text prompts (b) and those that combine text and layout conditioning (c-e) often struggle to produce images with accurate occlusion relationships. In this work, we propose a new framework called VODiff, which enhances control over object arrangement by introducing their visibility orders (indicated by numbers above their bounding box in the layout map in (a)) as auxiliary constraints. VODiff enables the generation of images with correct spatial arrangements and occlusion relationships.  
+![](images/e0cd9df4466b5e25d275006dd0334a0aaafe8427b37a82dd13563634a7e6fca4.jpg)  
+"A picture of a living room with a table and a plant in front of a sofa, and a dog on the sofa."
+
+![](images/cd961ee444dbf23977023005ab0f2dd6e8abeba45d987059b2b9ffabc9e27a6f.jpg)
+
+![](images/ba94a955f8739f650d20019a5b7ee45b70244a8f863f81f790cb4325b2158ee3.jpg)
+
+![](images/7c21e606ccfd361c39ecff52685c2c865629e60a794729a3d1840f630f373b41.jpg)
+
+![](images/3d4ca8009b08819f411e09f0503f73962549d923b4cc04d31f104db0bc7869c4.jpg)
+
+![](images/b363900ca272bb5e12e8a111fb2876bbf094d57e1f4087ab4f293a320cf2acb2.jpg)
+
+# Abstract
+
+Recent advancements in diffusion models have significantly enhanced the performance of text-to-image models in image synthesis. To enable control over the spatial locations of the generated objects, diffusion-based methods typically utilize object layout as an auxiliary input. However, we observe that this approach treats all objects as being on the same layer and neglect their visibility order, lead
+
+ing to the synthesis of overlapping objects with incorrect occlusions. To address this limitation, we introduce in this paper a new training-free framework that considers object visibility order explicitly and allows users to place overlapping objects in a stack of layers. Our framework consists of two visibility-based designs. First, we propose a novel Sequential Denoising Process (SDP) to divide the whole image generation into multiple stages for different objects, each stage primarily focuses on an object. Second, we propose a novel Visibility-Order-Aware (VOA) Loss to trans
+
+form the layout and occlusion constraints into an attention map optimization process to improve the accuracy of synthesizing object occlusions in complex scenes. By merging these two novel components, our framework, dubbed VODiff, enables the generation of photorealistic images that satisfy user-specified spatial constraints and object occlusion relationships. In addition, we introduce VOBench, a diverse benchmark dataset containing 200 curated samples, each with a reference image, text prompts, object visibility orders and layout maps. We conduct extensive evaluations on this dataset to demonstrate the superiority of our approach.
+
+# 1. Introduction
+
+With the advancements of diffusion models [20, 59], text-to-image (T2I) models [54, 55, 57] have revolutionized the field of image generation. These large-scale T2I models, e.g., Stable Diffusion [55], can generate diverse, high-quality images conforming to given prompts, facilitating more efficient image editing [4] and artistic creation [44].
+
+Despite the success, existing T2I models often struggle to produce satisfactory results in complex scenes involving multiple objects with occlusions. On the one hand, text descriptions alone lack spatial controllability, causing text-based methods to likely generate incorrect object positions and relationships, as shown in Fig. 1(b). On the other hand, although some condition-based methods [27, 30, 83] introduce layout conditions (e.g., bbox) into pre-trained T2I models to improve positional accuracy, they do not consider the occlusion relationship of objects and may produce unsatisfactory results, as shown in Fig. 1(c)-(e).
+
+In this work, we observe that when generating a scene with multiple objects, existing methods typically employ a one-layer-for-all strategy to synthesize all objects simultaneously on a single canvas. Such a design requires the model to associate the objects with the input text description and accurately obtain their spatial/occlusion relationships from it, if available. This is inherently challenging and often results in mixed or even incorrect object arrangements. Even when layout constraints [3, 27, 47, 68, 70] are provided, the model may still struggle with synthesizing the correct visibility order of overlapping objects, resulting in inaccurate occlusions. To address this problem, we propose a new training-free Visibility Order Diffusion (VODiff) framework to explicitly model the visibility order of objects by assigning each user-specified object to a separate layer and executing a novel multi-layer denoising process to produce the output image.
+
+We first propose a novel Sequential Denoising Process (SDP) to render the input objects one by one. SDP stratifies the objects into distinct layers, from bottom to top, according to the visibility order. For example, in the second example of Fig. 1, the layers are ordered with the "dog" at the
+
+bottom, the "suitcase" in the middle, and the "banana" at the top. Second, SDP divides the complete denoising process into several sub-stages, with each sub-stage primarily focusing on generating one object. Concurrently, during the denoising process, SDP assigns dynamically varying visual guidance to different objects/layers at different stages using a newly proposed smooth guidance mechanism. To prevent the misalignment of inter-object relations caused by deviations in visual guidance during denoising, we further propose a novel Visibility-Order-Aware (VOA) loss to transform layout and occlusion constraints into attention map optimization. When synthesizing an object, we first divide the corresponding layer into three regions, overlapping, visible, and background regions, based on its visibility w.r.t. those of the other objects, and then apply different constraints on these regions.
+
+To address the lack of a benchmark in visibility-order-aware T2I generation, we also curate a new benchmark called VOBench, which comprises tailored visibility orders, layouts, referenced images, and prompts. To validate the effectiveness of our VODiff, we conduct extensive evaluations on VOBench. Results show that VODiff generates prompt-consistent objects within the user-specified layout regions and precisely delineates the occlusion relationship, as shown in Fig. 1(f).
+
+Our key contributions can be summarized as:
+
+- To the best of our knowledge, we are the first to consider visibility order when generating images with T2I models. We present a new training-free framework named VODiff, which explicitly models object visibility order to control object occlusion relationships.  
+- We propose two novel designs to support the explicit modeling of object visibility order: a Sequential Denoising Process (SDP) to generate different objects sequentially at various sub-stages of the denoising process, and a Visibility-Order-Aware (VOA) Loss to incorporate occlusion relationships into the attention optimization process.  
+- We curate a visibility-order-aware benchmark called VOBench, and conduct extensive evaluations on it. Results show that compared to existing SOTA methods, VODiff can synthesize more photorealistic images, which satisfy user-specified spatial constraints as well as object visibility order.
+
+# 2. Related Work
+
+# 2.1. Text-to-Image (T2I) models
+
+Substantial progress has been made on T2I models through the use of diffusion models [20, 59, 60]. These models conceptualize image synthesis as an iterative denoising process guided by textual prompts. The denoising is conditioned on textual embeddings generated by language
+
+encoders [52, 53] and is performed either in pixel space [2, 43, 54, 57] or latent space [12, 16, 36, 48, 55, 74], followed by cascaded super-resolution [21] or latent-to-image decoding [11] for high-resolution image generation. By training on extensive datasets of image-text pairs [58], these models learn a unified generative space in which visual and linguistic modalities are closely associated.
+
+Nonetheless, current T2I diffusion models still struggle to accurately interpret spatial layout instructions, making it challenging to control the occlusion relationships among multiple objects using textual descriptions alone.
+
+# 2.2. Layout-guided Text-to-Image (T2I) Models
+
+Layout-guided T2I models aim to generate images that contain objects as specified by the input text description, with their locations aligned with the input layout.
+
+Training-based Models can be categorized into those that fine-tune pre-trained models [1, 7, 10, 23, 31, 51, 72, 76] using the specific input conditions, and those that train additional plug-ins [25, 30, 35, 37, 38, 40, 50, 67, 69, 79] to support various conditions. Some works achieve instance-level control [64, 65, 82, 83] by assigning individual prompts to each object to precisely control their attributes. However, these methods require additional training on large-scale datasets, resulting in significant computational and labor costs.
+
+Several methods [22, 26, 71, 81] employ multiple control conditions within a single generation. Sun et al. [63] address the complex relationships among different control conditions by training a multi-control encoder to maintain harmonious results when combining multiple controls. Although these approaches can generate images with harmonious occlusion relationships, they still cannot directly control the occlusion relationship of individual objects.
+
+Sampling-based Models often integrate multiple denoising processes, with each process targeting a specific component or region of the image. Liu et al. [33] interpret diffusion models as energy-based models and propose two compositional operators to combine the score functions of different conditions in a single sampling step. MultiDiff. [3] and LRDiff. [49] then propose to utilize shared parameters or constraints to ensure consistency/accuracy of each object in the generated image.
+
+Although several methods [39, 41] have proposed training-free multimodal control for the T2I generation, they primarily focus on enhancing control techniques without specifically addressing the object occlusion issues.
+
+Attention-based Models can be categorized into forward and backward approaches based on how they employ the attention maps. Forward methods [2, 13, 15, 27, 73] manipulate attention values directly on the feature map. The accuracy of these methods heavily depends on the precision of the layout condition (e.g., segmentation masks). Back-
+
+ward methods [5, 8, 47, 68, 70, 73] combine the layout condition with the attention map, and utilize an energy function to translate the spatial constraints into the noisy input for iterative updates. These methods also primarily emphasize on positional accuracy while neglecting the occlusion relationships among objects, leading to the unsatisfactory generation of overlapping regions.
+
+In addition, while several customized T2I models [9, 29, 32, 61, 62, 72, 80] attempt to integrate specific input objects into designated regions of the generated image by using additional layout input, they overlook the visibility order among the integrated objects and other objects within or near those regions. This limitation hinders their effectiveness in addressing occlusion issues.
+
+# 3. Method
+
+In this work, our goal is to enhance the ability of pre-trained text-to-image (T2I) diffusion models to correctly handle occlusion relationships in multi-object scenarios without training. We note that the overlapping regions of the objects are the main challenge, and they affect the occlusion relationship of the objects in the synthesized image. However, it is difficult for T2I models to understand occlusion relations among objects through text prompts only. In addition, current spatial layouts can only help constrain the spatial location of the objects. They do not indicate the order of objects in the overlapping regions.
+
+To achieve our goal, we propose a training-free method, VODiff, which introduces object visibility order into the layout-guided T2I generation to effectively address the incorrect object arrangement problem. Fig. 2 shows our framework. It consists of two novel visibility-aware designs: a Sequential Denoising Process (SDP) and a Visibility-Order-Aware (VOA) loss. When generating an image, our SDP strategy divides the whole denoising process into several sub-stages according to the number of user-specified objects and generates the specified objects in each stage sequentially in a bottom-up manner. Meanwhile, at each denoising step, our VOA loss applies dynamic losses to regions with different visibility states, ensuring accurate occlusion relationships among the generated objects.
+
+For the rest of this section, we first briefly introduce the preliminary information of diffusion models and attention layers in Sec. 3.1. We then present how we design the SDP in Sec. 3.2 and detail the VOA loss function in Sec. 3.3.
+
+# 3.1. Preliminaries
+
+Diffusion Models [20, 59, 60] learns to approximate a data distribution by progressively denoising a random variable sampled from a unit Gaussian distribution  $\mathcal{N}(0,I)$ . At a denoising step  $t\in [1,\dots ,T]$ , the forward-time Markov chain is formulated as  $x_{t} = \sqrt{1 - \beta_{t}} x_{t - 1} + \sqrt{\beta_{t}}\epsilon_{t - 1}$ , where  $\epsilon \sim \mathcal{N}(0,I)$ , and  $\beta_{t}$  is a pre-defined variance schedule.
+
+![](images/2e4359b52be4386cfa615893b5865b4fff7bc1f690515d4dacbb3bb7a3af8615.jpg)  
+Figure 2. Our VODiff incorporates two visibility-aware designs: a sequential denoising process (SDP) and a visibility-order-aware (VOA) loss function. SDP divides the denoising process into several stages according to the visibility order, and performs denoising on different parts sequentially. Meanwhile, the VOA loss is utilized to eliminate the local spatial shifts in the attention layers, further preventing position deviations and incorrect occlusion relationships. Given a text prompt, the spatial layout, and the visibility order of the objects to be synthesized as inputs, our VODiff can generate images that simultaneously adhere to all the constraints.
+
+The reverse-time denoising process can then be defined as:  $x_{t - 1} = \tilde{\alpha}_t x_t + \tilde{\beta}_t \hat{s}_t + \sigma_t \epsilon_t$ , where  $\tilde{\alpha}_t, \tilde{\beta}_t$ , and  $\sigma_t$  are calculated coefficients. To generate an image, given a noisy input  $x_T$  sampled from  $\mathcal{N}(0, I)$ , we run denoiser  $\epsilon_\theta$  iteratively to obtain intermediate representations  $x_{T - 1}, \dots, x_1, x_0$  where  $x_0 \approx x$  is the final result.
+
+Attention Layers are the cornerstone of Stable Diffusion [55]. They used to condition the generation based on the text prompt. A pretrained CLIP encoder [52] is usually utilized to encode the text prompt  $c = (c_{1}, c_{2}, \dots, c_{n})$ , where  $n$  is the number of tokens, into text embeddings. These tokens are then used as keys  $K \in \mathbb{R}^{n \times d}$  and values  $V \in \mathbb{R}^{n \times d}$  in the Cross-Attention layer. The symbol  $d$  denotes the feature dimension. For a set of queries  $Q \in \mathbb{R}^{hw \times d}$  computed from the feature map of size  $h \times w$ , the cross-attention map  $A_{t}$  at time step  $t$  is calculated as:  $A_{t} = \text{softmax}\left(\frac{QK^{\top}}{\sqrt{d}}\right) \in [0, 1]^{hw \times n}$ , where  $A_{t}$  contains  $n$  attention maps  $\{A_{t}^{1}, \dots, A_{t}^{n}\}$ , and  $A_{t}^{i} \in [0, 1]^{hw}$  indicates the strength of association between each word token  $c_{i}$  and each spatial location on the feature map. A higher attention value in  $A_{t}^{i}$  indicates a higher probability of generating the object  $c_{i}$  at that spatial location.
+
+# 3.2. Sequential Denoising Process (SDP)
+
+Instead of directly generating all objects on a single layer at once, like what existing works [27, 39, 47] do, we are inspired by the painter's algorithm [42], which sorts all polygons according to their z values and then render them starting from the farthest polygon, and the layer concept in image editing tools such as Photoshop. We propose to divide the generation of the target scene into layers, where each object is assigned a separate layer, and synthesize the ob
+
+jects sequentially, layer by layer from bottom to top, with the following steps.
+
+First, we stratify the target objects into distinct layers according to their visibility order. Specifically, given an input text prompt  $C$  containing  $N$  target objects; a layout condition represented by bounding boxes  $B = \{b_{1}, b_{2}, \ldots, b_{N}\}$ , where  $b_{n}$  is the bounding box of the  $n$ -th target object; and visibility orders  $V = \{v_{1}, v_{2}, \ldots, v_{N}\}$ , where  $v_{n}$  is the visibility order of the  $n$ -th target object, we convert the layout condition  $B$  into  $N$  masks,  $M = \{m_{1}, m_{2}, \ldots, m_{N}\}$ , where  $m_{n}$  is the mask corresponding to the bounding box  $b_{n}$  with a value of 1 inside the bounding box region and 0 elsewhere. Meanwhile, we extract  $N$  object entities from the input text prompt  $C$  and denote the set as  $\{c_{1}, c_{2}, \dots, c_{N}\}$ . Each bounding box  $b_{n}$  is associated with an object prompt  $c_{n}$ , which describes the content of the target object inside  $b_{n}$ . We order the  $N$  target objects in an increasing object visibility order specified by the user, even though there may be non-overlapping objects. The layers from bottom to top can thus be represented as  $L = \{(m_{1}, c_{1}, v_{1}), (m_{2}, c_{2}, v_{2}), \dots, (m_{N}, c_{N}, v_{N})\}$ .
+
+Second, we divide the reverse-time denoising process (with a total of  $T$  steps) of our SDP into two main parts,  $[T, T_g]$  and  $[T_g, 0]$ , where  $T_g$  denotes the start time of global denoising. The global denoising is to delineate the overall scene based on the input text prompt  $C$ . The first part of our SDP focuses on per-layer generation based on layers  $L$ . We divide it into  $N$  sub-stages,  $S = \{s_1, s_2, \dots, s_N\}$ , starting from the bottom layer (i.e., the object with the lowest visibility order). Each sub-stage  $s_n$  contains  $k$  denoising steps, ranging from  $[s_n, s_n + k]$ , i.e.,  $T - T_g = kN$ . Since the diffusion denoising process is reversed,  $s_n + k$  represents the start time for substage  $s_n$ . At each sub-stage, to
+
+force  $c_{n}$  to be generated within the object region, we explicitly incorporate visual guidance into this region. However, we observe that focusing exclusively on a single object  $n$  at each stage can lead to the failure of generating other objects, particularly those whose start time is near  $T_{g}$ , due to the missing of structure information [2] at later denoising steps. To address this problem, at each stage  $n$ , we apply visual guidance not only for the target object  $n$  but also for other objects  $j$  ( $j \neq n$ ), to ensure the generated layout adheres to the overall requirements. Our scheme for adding visual guidance can be formulated as:
+
+$$
+x _ {t, n} = x _ {t} + \lambda_ {t, n} \odot m _ {n}, \tag {1}
+$$
+
+where  $\lambda_{t,n}$  represents the visual guidance scale of the target object  $n$  at time step  $t$ . Note that we allocate a dynamically varying smooth  $\lambda_{t,n}$  to each object to ensure correct occlusion relationships. Specifically, when processing object  $n$  at sub-stage  $n$ , the SDP assigns a guidance scale  $a$  to object  $n$  and  $a/2$  to all other objects. As the process transitions from sub-stage  $n$  to sub-stage  $n + 1$ , the guidance scale for object  $n$  decreases linearly to  $a/2$ , while the guidance scale for object  $n + 1$  increases linearly to  $a$ , and guidance scale for the other objects remains unchanged, as:
+
+$$
+\lambda_ {t, n} = \left\{ \begin{array}{l l} a - \frac {a}{2} \frac {t - s _ {n - 1}}{k} & \text {i f} s _ {n - 1} <   t <   = s _ {n - 1} + k, \\ \frac {a}{2} + \frac {a}{2} \frac {t - s _ {n}}{k} & \text {i f} s _ {n} <   t <   = s _ {n} + k, \\ \frac {a}{2} & \text {o t h e r w i s e}, \end{array} \right. \tag {2}
+$$
+
+where  $a$  denotes the initial visual guidance scale (refer to Sec. D.4 of the Supp. for more discussion) and  $s_0 = T$ . By assigning a smoothly transitioning visual guidance, we ensure that each sub-stage focuses more on a specific object while maintaining correct occlusion relationships between it and other objects, such that in overlapping regions, objects with a higher visibility order occlude those with a lower one. Finally, we forward the updated noisy input  $x_{t,n}$  to the denoising model  $\epsilon_{\theta}$  to estimate noise for each object separately and then combine them together, as:
+
+$$
+\epsilon_ {\theta} \left(x _ {t}, t, C\right) = \sum_ {n = 1} ^ {N} \frac {m _ {n}}{\sum_ {n = 1} ^ {N} m _ {n}} \otimes \epsilon_ {\theta} \left(x _ {t, n}, t, c _ {n}\right), \tag {3}
+$$
+
+where  $\epsilon_{\theta}(x_t,t,C)$  is the estimated noise used to update  $x_{t - 1}$  via the standard diffusion scheduler.
+
+# 3.3. Visibility-Order-Aware (VOA) Loss
+
+Although SDP can help constrain target generation within the specified region using visual guidance, the consecutive conv. layers of the denoising model introduce local shifts, which are then amplified by the attention layers. As a result, adding visual guidance at the input level (i.e.,  $x_{t}$ ) as a constraint is insufficient and tends to cause object position deviation and false occlusion relationships (see Fig. 5 column 2 for a visual illustration).
+
+![](images/0726e66f0937cde8976af8f2f45d781c5d3838e665730c3ac944c38a82b01c79.jpg)  
+Layout + Order
+
+![](images/b7b3e19830a58b789a6962629bf870cdb82bf3b59d89d78141f629252401ffba.jpg)  
+moverl mvisible mbackgrd  
+Figure 3. Region separation illustration in the VOA loss.  $m_{n}^{\mathrm{overlap}}$  represents the regions where the current object  $n$  occludes other objects,  $m_{n}^{\mathrm{visible}}$  and  $m_{n}^{\mathrm{background}}$  denote where the current object should and should not be generated in the final image.
+
+![](images/76a7c050cf5dd8b17b460fa1a69427b6309273834ae26436aa2950603ffac531.jpg)  
+Output Image
+
+To address the above problem, we propose constraining the spatial location and occlusion relationship at feature-level. Specifically, when generating an image with multiple target objects, different object prompts usually have different activations in different regions of the cross-attention map. At a given spatial location, the higher the attention value of a prompt, the more likely the corresponding object will be generated in that region [17]. In addition, when two objects overlap each other, their occlusion relationship is directly influenced by their visibility order. Once the order changes, the content of the overlapping region changes accordingly. Based on this, we transform the spatial layout and occlusion relationship among the objects to the cross-attention map and optimize it for correct image generation.
+
+As shown in Fig. 3, when processing target object  $n$ , we first divide the mask  $m_{n}$  into three parts, overlapping regions  $m_{n}^{\mathrm{overlap}}$ , visible regions  $m_{n}^{\mathrm{visible}}$ , and background regions  $m_{n}^{\mathrm{background}}$ , based on the visibility order and masks of all objects in the layer set  $L$ . Specifically, for  $m_{n}^{\mathrm{overlap}}$ , we compute the overlapping regions (i.e., dark orange region) of the current object mask  $m_{n}$  and the masks of all objects with lower visibility order than the current object (i.e.,  $\{m_{j}\}$ , with  $j \in [1, n-1]$ ) as  $m_{n}^{\mathrm{overlap}} = \bigcup_{j=1}^{n-1} m_{n} \odot m_{j}$ . This is because an object located at an upper layer should obtain a higher attention score in the overlapping regions. For  $m_{n}^{\mathrm{visible}}$ , we first compute the overlapping regions of the current mask  $m_{n}$  with the masks of all objects that have a higher visibility order (i.e.,  $\{m_{j}\}$ , with  $j \in [n+1, N]$ ). We then subtract these overlapping regions from the current object mask, which can be expressed as  $m_{n}^{\mathrm{visible}} = m_{n} - \bigcup_{j=n+1}^{N} m_{n} \odot m_{j}$ .  $m_{n}^{\mathrm{background}}$  is then obtained by inverting  $m_{n}^{\mathrm{visible}}$  as  $m_{n}^{\mathrm{background}} = 1 - m_{n}^{\mathrm{visible}}$ .
+
+During the denoising phase for the  $n$ -th target object at time step  $t$ , as indicated in Eq. 3, we extract the cross-attention maps referred to by the object prompt  $c_{n}$  from each layer and average them to produce  $\mathbf{A}_{t,n}$ . We then compute the loss between the attention map  $\mathbf{A}_{t,n}$  and each of the regions. For the overlapping regions, which are affected by the visibility order, we ensure that objects in the upper layers obtain higher attention scores, as:
+
+$$
+\mathcal {L} _ {t, n} ^ {\mathrm {o}} = \alpha_ {o} \cdot \left[ 1 - \min  \left(\mathbf {A} _ {t, n} \odot m _ {n} ^ {\text {o v e r l a p}}\right) \right]. \tag {4}
+$$
+
+![](images/f286221b7c77d1d015e6b94307e065093d084e920aad7a540bd50ee82e31d13a.jpg)  
+"A picture of a cat, in front of a CD."
+
+![](images/ce17ba2ab414c47029b14a0a20f03afb96287a2ec61845bd18ae535b319a9614.jpg)
+
+![](images/9d734215b5e77186a88ff5776f6b8b030ff6c2c7c48e1b2eb44940b5c2ef9eb3.jpg)
+
+![](images/ae0960715533bf88598d93aa1a4b70a2e58a5b9d23565941ed7ba18916b0665f.jpg)
+
+![](images/9df1fea1ee6af4ad2d64997fcbd1a400e5f7a3ca3fb6bd32d23023877c492a3a.jpg)
+
+![](images/da0711edb9b22e5fdf281266ad5ec27723e5eceeb5cec78dccb2430f09990a8f.jpg)
+
+![](images/b31f5795205c9628c116d0e5ec8c6ccd4c8ae76d331d965d80864c8ea4538fb7.jpg)
+
+![](images/ba6d297d1203906f1275aa759b9b5eb5e6079fedfe8cbecfe9d52c6cdab61c4e.jpg)
+
+![](images/74a4e7daffbb1f5605e43931c895c8613910e365c9a98c008e380c9e4a5ec4f0.jpg)  
+"A laptop on the table with a plant behind it, and a cup near it."
+
+![](images/063d3786b086993b97a559f2741de26895aec842663364a2495a8ef5dbe6da10.jpg)
+
+![](images/563b95716ece1982d690cc18978f54f7343da30aa3588490567b348b7d60b8ed.jpg)
+
+![](images/0ee7477788993e3024b0a376df272a5d9cef0127487f94255cadb7092d74d8ba.jpg)
+
+![](images/be4407ad3e94c30cddaea4eda069f97ac82efc83503b3d7b739568a267444da5.jpg)
+
+![](images/2edd5e00bffc71990355a2151cd9a70192563c824b967ef62ed8d8e5878a8756.jpg)
+
+![](images/21dd74696ef2ff62afd698794830001a938fa8943442d8ec0f388e11b3a7fcf0.jpg)
+
+![](images/7e3083d800d017630758cd36e8de6ab22b08836b478e4a84ee7f9aff1ab6c428.jpg)
+
+![](images/b280df6c33bf337eb85bf78871cc45a18bbc7e12810b012de9d1b3af6746ca54.jpg)  
+"Left chair and right chair, a lamp and a plant in front of a bookshelf."
+
+![](images/582497669611b385566a15c4e9b3e0e11549314bd54777fb6575cb48db381b10.jpg)  
+Figure 4. Qualitative comparison with six SOTA methods in Tab. 1. Training-based methods are labeled with  $\dagger$ .
+
+![](images/657650ba2ba1848ac104b557bea57d507980cc91ad17e2655f20a7914cc106f4.jpg)
+
+![](images/3d2e2486f491fba8ad002fc48000a214a3012963043f68fa76f9ec785fba5d19.jpg)
+
+![](images/afeaeed0f2717ad1f2d0133eb34d2c20d8ed2957758e4817b3180e49a6e86e4b.jpg)
+
+![](images/2613240aa8c6356dd8c6dc2ff6cc69d524b1762a3f90e9205cd64cddc8d06d93.jpg)
+
+![](images/d91a5301d1caf3d62be6c3decee53a5b22e71be91798fc9d6818783dc72f24ba.jpg)
+
+![](images/e1c8306413edda32a103e8f9d238359270439bbeb8079cf3253c0cb0a13dbe7e.jpg)
+
+We then encourage target object  $n$  to be generated within the visible regions by increasing the attention score through  $\mathcal{L}_{t,n}^{\mathrm{v}}$ , and prevent the target deviation by weakening the attention score of the background regions via  $\mathcal{L}_{t,n}^{\mathrm{b}}$ :
+
+$$
+\mathcal {L} _ {t, n} ^ {\mathrm {v}} = \alpha_ {v i s} \cdot \left[ 1 - \max  \left(\mathbf {A} _ {t, n} \odot m _ {n} ^ {\text {v i s i b l e}}\right) \right], \tag {5}
+$$
+
+$$
+\mathcal {L} _ {t, n} ^ {\mathrm {b}} = \alpha_ {b a c} \cdot \max  \left(\mathbf {A} _ {t, n} \odot m _ {n} ^ {\text {b a c k g r o u n d}}\right), \tag {6}
+$$
+
+where  $\alpha_{o}$ ,  $\alpha_{vis}$  and  $\alpha_{bac}$  are coefficients for the three loss functions. Finally, the total loss is:
+
+$$
+\mathcal {L} _ {t} = \sum_ {n} ^ {N} \left(\mathcal {L} _ {t, n} ^ {\mathrm {o}} + \mathcal {L} _ {t, n} ^ {\mathrm {v}} + \mathcal {L} _ {t, n} ^ {\mathrm {b}}\right). \tag {7}
+$$
+
+With the visibility-order-aware loss, we modify the noise sample  $x_{t}$  at each denoising step to minimize the loss using gradient descent  $\hat{x}_{t} \gets x_{t} - \alpha \nabla_{x_{t}}\mathcal{L}_{t}$ , where  $\alpha$  controls the influence of the optimization in the denoising process. In each step  $t$ , we iterate the gradient descent process  $\tau$  times. By applying the VOA loss on all objects at each sub-stage, VODiff alleviates the guidance deviation and achieves more accurate object spatial position and occlusions.
+
+# 4. Experiment
+
+# 4.1. Metrics
+
+We adopt a few metrics to evaluate our model:
+
+- FID: We use Fréchet Inception Distance [18] to evaluate the image quality. It compares the distribution of generated images to that of real images. A lower FID score indicates a higher image quality, reflecting closer alignment with real images on visual features and diversity.  
+- CLIP-Score: We use the CLIP-Score to compute the distance between input textual features and generated image features, evaluating the fidelity of the generated image to the textual input.
+
+- **Layout Alignment (LA):** We use Grounding-DINO [34] to assess Layout alignment accuracy. It detects bounding boxes for each instance and then compute the mean IoU between the detected and Ground Truth boxes. The box threshold is set to 0.5 to avoid incorrect box detection.  
+- Occlusion Alignment (OA): We employ GPT-4V to determine the occlusion relationships between objects in the generated images by prompting it with the image and the object names contained in the image. We then compare the visual order returned by GPT-4V with the ground truth visibility order. If the two match with each other, we count it as accurate. Finally, we compute the percentage of accurate occlusion relationships from all the samples.
+
+# 4.2.VObench
+
+Since there are no visibility-order-aware datasets available for evaluation, we curate a VOBench for this purpose. All the raw image data in VOBench are manually collected from the Internet. The data annotation process is as follows. We first employ GroundingDINO [34] and GPT-4V to generate the bounding boxes and text prompts for each object in the image, and manually assign the visibility order to each object box. Based on the number of boxes detected by GroundingDINO, we then divide the images into four categories containing 2-5 objects. For each category, we manually select 50 images based on the degree of overlap between boxes to make sure that various occlusion relations are in our VOBench.
+
+# 4.3. Comparison with State-of-the-Art Methods
+
+We compare VODiff with 12 layout-guided text-to-image generation methods, including seven training-free methods: BoxDiff [70], R&B [68], Attention-Refocusing [47],
+
+Table 1. Quantitative comparison of our VODiff with training-based and training-free methods. LA and OA represent the layout and occlusion alignment accuracy, respectively. AR denotes the average ranking from the user study, while Q, L, and O indicate that the ranking is based on Quality, Layout, and Occlusion. The best and second-best performances are marked in bold and underlined, respectively.  
+
+<table><tr><td rowspan="2">Metrics</td><td colspan="5">Training-based</td><td colspan="8">Training-free</td></tr><tr><td>SmartCtrl [35]</td><td>ControlNet [79]</td><td>MIGC[83]</td><td>InstanceDiff. [65]</td><td>AnyControl [63]</td><td>BoxDiff [70]</td><td>R&amp;B [68]</td><td>AR [47]</td><td>MultiDiff. [3]</td><td>DenseDiff. [27]</td><td>FreeCtrl [39]</td><td>RPG [73]</td><td>Ours</td></tr><tr><td>FID ↓</td><td>10.58</td><td>18.91</td><td>9.82</td><td>10.21</td><td>10.35</td><td>21.99</td><td>12.37</td><td>14.87</td><td>16.34</td><td>21.68</td><td>17.35</td><td>11.66</td><td>10.03</td></tr><tr><td>CLIP-Score ↑</td><td>29.45</td><td>29.34</td><td>29.61</td><td>29.53</td><td>29.63</td><td>29.46</td><td>28.72</td><td>29.59</td><td>27.42</td><td>29.33</td><td>29.12</td><td>29.57</td><td>29.73</td></tr><tr><td>LA ↑</td><td>54.81</td><td>52.73</td><td>54.62</td><td>53.58</td><td>54.77</td><td>13.50</td><td>30.88</td><td>29.79</td><td>16.60</td><td>24.11</td><td>36.14</td><td>51.15</td><td>55.11</td></tr><tr><td>OA ↑</td><td>55.00</td><td>51.50</td><td>65.00</td><td>63.50</td><td>69.50</td><td>18.50</td><td>34.50</td><td>32.00</td><td>21.50</td><td>24.00</td><td>28.50</td><td>51.50</td><td>82.50</td></tr><tr><td>AR-Q ↓</td><td>5.33</td><td>12.13</td><td>2.47</td><td>6.20</td><td>3.05</td><td>8.15</td><td>10.05</td><td>4.13</td><td>12.68</td><td>7.10</td><td>9.00</td><td>5.12</td><td>1.50</td></tr><tr><td>AR-L ↓</td><td>2.16</td><td>6.21</td><td>4.04</td><td>5.11</td><td>3.12</td><td>12.98</td><td>9.02</td><td>10.07</td><td>12.03</td><td>11.14</td><td>8.08</td><td>7.17</td><td>1.73</td></tr><tr><td>AR-O ↓</td><td>6.25</td><td>5.12</td><td>3.18</td><td>4.02</td><td>2.05</td><td>12.74</td><td>9.01</td><td>10.11</td><td>12.04</td><td>11.09</td><td>8.03</td><td>7.10</td><td>1.26</td></tr></table>
+
+Table 2. Ablation study on components. Constant and Dynamic represent the state of change of guidance scale in the SDP.  
+
+<table><tr><td>SDP 
+Constant Dynamic</td><td>VOA</td><td>FID ↓</td><td>Clip-Score ↑</td><td>LA ↑</td><td>OA ↑</td></tr><tr><td>✓</td><td></td><td>9.98</td><td>27.85</td><td>32.33</td><td>54.50</td></tr><tr><td></td><td>✓</td><td>13.34</td><td>29.01</td><td>43.25</td><td>73.50</td></tr><tr><td>✓</td><td>✓</td><td>11.68</td><td>29.33</td><td>43.36</td><td>69.50</td></tr><tr><td>✓</td><td>✓</td><td>10.03</td><td>29.73</td><td>55.11</td><td>82.50</td></tr></table>
+
+MultiDiffusion [3], DenseDiffusion [27], FreeControl [39] and RPG-Diffusion [73], and five training-based methods: SmartControl [35], ControlNet [79], AnyControl [63], MIGC [83] and InstanceDiffusion [65]. Since these methods do not support visibility order as an input, we include descriptions of object occlusion relationships in the text prompts to ensure fairness.
+
+Qualitative Comparison. As shown in Fig. 4, we visually compare VODiff with six top-performing methods. The images generated by VODiff exhibit superior spatial consistency and occlusion handling. Specifically, our method handles varying numbers of objects and degrees of occlusion remarkably well. In contrast, other methods exhibit issues such as incorrect spatial positioning (e.g., the 2nd row, (b-f)) or missing objects (e.g., the 3rd row, (d-f)). In addition, although some methods manage to generate correct spatial relationships, they fail in occlusion handling (e.g., the 1st row, (b,e,f,g)).
+
+Quantitative Comparison. As shown in Tab. 1, we evaluate our method against both training-free and training-based approaches. Our method significantly outperforms the training-free methods across all metrics. We achieve higher CLIP-Score (29.73), LA (55.11), and OA (82.50) scores, indicating superior image quality, layout alignment, and occlusion handling. Despite being training-free, our method outperforms training-based methods in six out of seven metrics. We achieve higher CLIP-Score (29.73 vs 29.63 by AnyControl), LA (55.11 vs 54.81 by SmartControl), and OA (82.50 vs 69.50 by AnyControl) scores. This demonstrates the effectiveness of our method in generating text-/layout-aligned and occlusion handled images without
+
+![](images/cecbb91620f2c77915d98270bf5fde2d7c24e901ab3f4010a80789567c45d16f.jpg)  
+"A bowl of strawberries and a cup on a blanket"
+
+![](images/41c8934317f3065450f9d280f8e83029236964029378cae0582e953881982acd.jpg)  
+Layout + Order  
+SDP
+
+![](images/f8d3e1200333f75876254ebc53169bda58386ae479f89ca832f206c1de7f6068.jpg)  
+VOA
+
+![](images/3bac4bcb48d2485fd1ec6048dd06ef7ea760667f5fd8498758ef20b9962b4110.jpg)  
+VOA+SDPc
+
+![](images/2e8af02634b88617c21adff2ce9baaeb55674dacddc9f187f1cd5a5d25c3caab.jpg)  
+Ours
+
+![](images/03374d504f92570bcf9c0141629f3e33f3305ce8cf55a682808345a9935959b1.jpg)  
+"A man and a dog in front of a sofa"  
+Figure 5. Visual comparison of the ablation study in Tab. 2.  $\mathrm{SDP}_c$  denotes that the constant guidance scale is used. By default, the dynamic guidance scale is used in SDP.
+
+![](images/299d32385b9527f212ba2dea35a6140e9074ea35062a227b4059664b609b1b09.jpg)
+
+![](images/997686e00c4a7784e1cdeb12ce40b00a953f11cf87dcc97cb21ebaf6cf1824d0.jpg)
+
+![](images/31d3aaf85111125ebeaf052529692cc6c988e95a07f05d51e6f1c22d159fafbc.jpg)
+
+![](images/8fa3e6cd82c0e860c844c50297115a37476eb73992c9f004202559cbe3ccce5d.jpg)
+
+additional training. Although the training-based method, MIGC, achieves the best FID score of 9.82, our method achieves a similarly FID of 10.03. The slight reduction can be attributed to the optimization on attention maps [47].
+
+User Study. We also conduct a user study with 63 participants to assess various methods through subjective evaluation. We present users with ten sets of 8 images along the same input conditions (i.e., prompt and layout map), and ask them to rank the given images based on image quality (AR-Q), layout (AR-L), and occlusion accuracy (AR-O). Each pair of images are shown randomly, and we collect ranking scores from different participants. We then compute the average ranking (AR) for three aspects: quality, layout, and occlusion as shown in Tab. 1. The results show a clear preference for our method among the users.
+
+# 4.4. Ablation Study
+
+Component Analysis. We ablate our VODiff by removing SDP and VOA separately, and replacing the smooth guidance with constant guidance (denoted as  $\mathrm{VOA + SDP}_c$ ). Tab. 2 and Fig. 5 show the quantitative and qualitative results. We have three conclusions. Using only SDP (Dynamic) results in the worst LA, OA accuracy and Clip-Score but produces better FID. This is because SDP only adds visual guidance to guide the denoising direction, which does not impose stronger spatial constraints, and thus does not significantly affect the denoising process. VOA significantly improves the layout and occlusion accuracy by optimizing the cross-attention map, but at the same time, causes the model to produce less diversified results, thus impacting
+
+Table 3. Ablation study on the VOA loss.  ${\mathcal{L}}_{\mathrm{v}},{\mathcal{L}}_{\mathrm{o}}$  and  ${\mathcal{L}}_{\mathrm{b}}$  represent the loss on visible, occluded and background regions, respectively.  
+
+<table><tr><td>Lv</td><td>Lo</td><td>Lb</td><td>FID ↓</td><td>Clip-Score ↑</td><td>LA ↑</td><td>OA ↑</td></tr><tr><td>✓</td><td></td><td></td><td>11.41</td><td>28.05</td><td>42.13</td><td>60.50</td></tr><tr><td>✓</td><td></td><td>✓</td><td>10.23</td><td>29.25</td><td>46.42</td><td>61.50</td></tr><tr><td>✓</td><td>✓</td><td></td><td>14.65</td><td>28.42</td><td>44.47</td><td>77.50</td></tr><tr><td></td><td>✓</td><td>✓</td><td>11.68</td><td>28.10</td><td>29.14</td><td>58.00</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>10.03</td><td>29.73</td><td>55.11</td><td>82.50</td></tr></table>
+
+![](images/8536f1f2a099ef466ac33c162f3050e267cd2dec4c9d3031f865bd85476b664e.jpg)  
+Layout+Order  
+"A man sitting in front of a table, a laptop and a cup on the table"
+
+![](images/04ae46556d5dac17117f5acbc6f8fe93b109eacf9f266bca9425045c7ded64d0.jpg)  
+GLIGEN[30] ControlNet[79] MIGC[83] InstanceDiff.[65]  
+Figure 6. Enhancement of existing methods using VODiff. Row 1 and 2 show the vanilla and the enhanced results, respectively.
+
+the FID metrics. ⑧ Although constant visual guidance can improve FID, CLIP-Score, and LA, when combined with VOA, it confuses the order of objects in overlapping regions during the denoising process, interfering with attention optimization and worsening the OA. Finally, combining all the proposed components helps achieve the best results.
+
+Visibility-Order-Aware (VOA) loss. We verify the efficacy of different loss items in the VOA by removing them separately. Tab. 3 shows the quantitative results, and Supp. Fig. 3 shows the visual results. We can see that  $\mathbf{1} \mathcal{L}_v$  is indispensable for layout and occlusion constraints (4th row), but with  $\mathcal{L}_v$  alone is also insufficient (1st row);  $\mathbf{2} \mathcal{L}_v + \mathcal{L}_b$  produces inferior OA results as the overlapping regions lack constraints;  $\mathbf{3} \mathcal{L}_v + \mathcal{L}_o$  has the remarkable OA results, but the worst FID because this combination only considers each object by itself but ignores background constraints (e.g., misaligned "man" and "dog" in the 1st row, 5-th column of Fig. 3 in Supp.). Finally, combining all items helps achieve the best performance on all metrics.
+
+# 4.5. Diverse Applications
+
+Our VODiff is training-free and can be easily integrated into many existing layout-to-image methods [30, 65, 79, 83], significantly enhancing the accuracy of occlusion relationship generation in current pretrained models. As shown in Fig. 6, the result from each method without (or with) using our approach is presented in the first (or second) row. Our VODiff effectively assists these training-based methods in addressing issues such as object disappearance (columns 2-4), as well as incorrect positional (column 5) and occlusion (columns 2 and 5) relationships.
+
+Our VODiff supports customized generation by integrat-
+
+![](images/9de675883694e7e7c6e6fd7d0049d94fa2cf0b4b7498f0e0a221bb5a6fae2665.jpg)  
+(a) Object insertion with accurate visibility orders
+
+![](images/15a480cd80dce19f2db67f37c02eb28b9b5290e460103822fc6e5957453cf8e3.jpg)  
+(b) Exemplar based image editing with visibility orders  
+Figure 7. Customized generation with VODiff in object insertion (row 1) and exemplar-based editing (row 2).
+
+ing it with existing techniques, e.g., DreamBooth [56]. In Fig. 7(a), we first demonstrate the insertion of newly generated objects into an existing image via Inversion [59], where the inserted objects adhere to the preset visibility order relative to the original objects in the image, while preserving the background unchanged. Fig. 7(b) shows the integration of our method with DreamBooth [56], enabling the generated objects to not only maintain the identity of the input objects but also ensure that the positions and visibility order among multiple objects adhere to the preset arrangement.
+
+![](images/92bcd44445f22b7da1ba9d6868606ec75fa9a0ddec913bdea472536451d1b417.jpg)  
+"A bedroom with a cabinet behind a plant, and there is also a bed, a flower on the right and a boy in front of a lamp who is feeding a dog, and a glass bottle is in front of him."  
+Figure 8. A failure case of VODiff. From left to right are: inputs, results of InstanceDiff. [65] and Our method, respectively.
+
+# 5. Conclusion
+
+In this paper, we have introduced VODiff to address the challenges of generating complex scenes with object occlusions. We first propose a Sequential Denoising Process that performs object generation sequentially from the bottom to the top layer based on the visibility order of objects. We then propose a Visibility-Order-Aware loss to constrain object positions and occlusion relationships by transforming the layout and visibility order into cross-attention map optimization. The results on our curated VOBench show superior performances of our method in controlling object spatial position and occlusion relationships.
+
+VODiff does has limitations. For example, as shown in Fig. 8, like existing method [65], our method may struggle to handle scenarios where the number of objects is exceptionally high due to the challenge of balancing the image quality with precise spatial and occlusion relationships.
+
+# Acknowledgements
+
+This work is supported in part by Ant Group and General Program of National Natural Science Foundation of China (NSFC, No. 6207071897) and a GRF from the Research Grants Council of Hong Kong (Ref.: 11205620).
+
+# References
+
+[1] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In CVPR, pages 18370-18380, 2023. 3  
+[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv:2211.01324, 2022. 3.5  
+[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: fusing diffusion paths for controlled image generation. In ICML, pages 1737-1752, 2023. 2, 3, 7, 5  
+[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 2  
+[5] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM TOG, 42(4):1-10, 2023. 3, 1  
+[6] Hong Chen, Yipeng Zhang, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. Disenbooth: Disentangled parameter-efficient tuning for subject-driven text-to-image generation. arXiv:2305.03374, 3, 2023. 1  
+[7] Kai Chen, Enze Xie, Zhe Chen, Yibo Wang, HONG Lanqing, Zhenguo Li, and Dit-Yan Yeung. Geodiffusion: Text-prompted geometric control for object detection data generation. In ICLR, 2024. 3  
+[8] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In WACV, pages 5343-5353, 2024. 3, 2  
+[9] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In CVPR, pages 6593-6602, 2024. 3, 1, 6  
+[10] Jiaxin Cheng, Xiao Liang, Xingjian Shi, Tong He, Tianjun Xiao, and Mu Li. Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv preprint arXiv:2302.08908, 2023. 3  
+[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 3  
+[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 3  
+[13] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric
+
+Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv:2212.05032, 2022. 3  
+[14] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv:2208.01618, 2022. 1  
+[15] Biao Gong, Siteng Huang, Yutong Feng, Shiwei Zhang, Yuyuan Li, and Yu Liu. Check locate rectify: A training-free layout calibration system for text-to-image generation. In CVPR, pages 6624-6634, 2024. 3  
+[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, pages 10696-10706, 2022. 3  
+[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 5  
+[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017. 6  
+[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NIPS Workshop, 2021. 5  
+[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NIPS, 2020. 2, 3  
+[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 23(47):1-33, 2022. 3  
+[22] Hexiang Hu, Kelvin CK Chan, Yu-Chuan Su, Wenhu Chen, Yandong Li, Kihyuk Sohn, Yang Zhao, Xue Ben, Boqing Gong, William Cohen, et al. Instruct-imagen: Image generation with multi-modal instruction. In CVPR, pages 4754-4763, 2024. 3  
+[23] Chengyou Jia, Minnan Luo, Zhuohang Dang, Guang Dai, Xiaojun Chang, Mengmeng Wang, and Jingdong Wang. Ssmg: Spatial-semantic map guided diffusion model for free-form layout-to-image generation. In AAAI, pages 2480–2488, 2024. 3  
+[24] Yuming Jiang, Nanxuan Zhao, Qing Liu, Krishna Kumar Singh, Shuai Yang, Chen Change Loy, and Ziwei Liu. Groupdiff: Diffusion-based group portrait editing. arXiv:2409.14379, 2024. 1  
+[25] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. Scedit: Efficient and controllable image diffusion generation via skip connection editing. In CVPR, pages 8995–9004, 2024. 3  
+[26] Sungnyun Kim, Junsoo Lee, Kibeom Hong, Daesik Kim, and Namhyuk Ahn. Diffblender: Scalable and composable multimodal text-to-image diffusion models. arXiv:2305.15194, 2023. 3  
+[27] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In ICCV, pages 7701-7711, 2023. 1, 2, 3, 4, 7, 5
+
+[28] Leheng Li, Weichao Qiu, Xu Yan, Jing He, Kaiqiang Zhou, Yingjie Cai, Qing Lian, Bingbing Liu, and Ying-Cong Chen. Omnibooth: Learning latent control for image synthesis with multi-modal instruction. arXiv:2410.04932, 2024. 1  
+[29] Tianle Li, Max Ku, Cong Wei, and Wenhu Chen. Dreamedit: Subject-driven image editing. arXiv:2306.12624, 2023. 3, 1  
+[30] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, pages 22511-22521, 2023. 1, 2, 3, 8, 5, 6, 9  
+[31] Yaqi Li, Han Fang, Zerun Feng, Kaijing Ma, Chao Ban, Xianghao Zang, LanXiang Zhou, Zhongjiang He, Jingyan Chen, Jiani Hu, et al. Goal: Grounded text-to-image synthesis with joint layout alignment tuning. In ACM MM, pages 7055-7064, 2024. 3  
+[32] Chang Liu, Xiangtai Li, and Henghui Ding. Referring image editing: Object-level image editing via referring expressions. In CVPR, pages 13128-13138, 2024. 3, 1  
+[33] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In ECCV, pages 423-439, 2022. 3  
+[34] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv:2303.05499, 2023.6.1  
+[35] Xiaoyu Liu, Yuxiang Wei, Ming Liu, Xianhui Lin, Peiran Ren, Xuansong Xie, and Wangmeng Zuo. Smartcontrol: Enhancing controlnet for handling rough visual conditions. arXiv:2404.06451, 2024. 3, 7  
+[36] Yuhao Liu, Zhanghan Ke, Fang Liu, Nanxuan Zhao, and Rynson WH Lau. Diff-plugin: Revitalizing details for diffusion-based low-level tasks. In CVPR, pages 4197-4208, 2024. 3  
+[37] Grace Luo, Trevor Darrell, Oliver Wang, Dan B Goldman, and Aleksander Holynski. Readout guidance: Learning control from diffusion features. In CVPR, pages 8217-8227, 2024. 3  
+[38] Kangfu Mei, Mauricio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. Codi: Conditional diffusion distillation for higher-fidelity and faster image generation. In CVPR, pages 9048-9058, 2024. 3  
+[39] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In CVPR, pages 7465-7475, 2024. 3, 4, 6, 7, 5  
+[40] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, pages 4296-4304, 2024. 3  
+[41] Nithin Gopalakrishnan Nair, Jeya Maria Jose Valanarasu, and Vishal M Patel. Maxfusion: Plug&play multi-modal generation in text-to-image diffusion models. In ECCV, pages 93-110, 2025. 3
+
+[42] Martin E Newell, RG Newell, and Tom L Sancha. A solution to the hidden surface problem. In Seminal graphics: pioneering efforts that shaped the field, pages 27-32. 1998. 4  
+[43] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv:2112.10741, 2021. 3  
+[44] NovelAI. Novelai: Ai generated stories and art. 2021. 2  
+[45] Lingzhi Pan, Tong Zhang, Bingyuan Chen, Qi Zhou, Wei Ke, Sabine Süssstrunk, and Mathieu Salzmann. Coherent and multi-modality image inpainting via latent space optimization. arXiv:2407.08019, 2024. 1  
+[46] Yulin Pan, Chaojie Mao, Zeyinzi Jiang, Zhen Han, and Jingfeng Zhang. Locate, assign, refine: Taming customized image inpainting with text-subject guidance. arXiv:2403.19534, 2024. 1  
+[47] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. In CVPR, pages 7932-7942, 2024. 2, 3, 4, 6, 7, 5  
+[48] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv:2307.01952, 2023. 3, 5  
+[49] Zipeng Qi, Guoxi Huang, Chenyang Liu, and Fei Ye. Layered rendering diffusion model for zero-shot guided image synthesis. In ECCV, 2024. 3, 1  
+[50] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, et al. Uniconcontrol: A unified diffusion model for controllable visual generation in the wild. In NIPS, 2023. 3  
+[51] Leigang Qu, Shengqiong Wu, Hao Fei, Liqiang Nie, and TatSeng Chua. Layoutllm-t2i: Eliciting layout guidance from llm for text-to-image generation. In ACM MM, pages 643-654, 2023. 3  
+[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 3, 4  
+[53] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 3  
+[54] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv:2204.06125, 2022. 2, 3  
+[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 3, 4, 5  
+[56] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine
+
+tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 8, 1, 6  
+[57] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NIPS, pages 36479-36494, 2022. 2, 3  
+[58] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NIPS, pages 25278–25294, 2022. 3, 1  
+[59] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 8, 1, 6  
+[60] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2, 3  
+[61] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In CVPR, pages 18310-18319, 2023. 3, 1  
+[62] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, He Zhang, Wei Xiong, and Daniel Aliaga. Imprint: Generative object compositing by learning identity-preserving representation. In CVPR, pages 8048-8058, 2024. 3, 1, 6  
+[63] Yanan Sun, Yanchen Liu, Yinhao Tang, Wenjie Pei, and Kai Chen. Anycontrol: Create your artwork with versatile control on text-to-image generation, 2024. 3, 6, 7  
+[64] Zhenhong Sun, Junyan Wang, Zhiyu Tan, Daoyi Dong, Hailan Ma, Hao Li, and Dong Gong. Eggen: Image generation with multi-entity prior learning through entity guidance. In ACM MM, pages 6637-6645, 2024. 3  
+[65] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation. In CVPR, pages 6232-6242, 2024. 3, 6, 7, 8, 5, 9  
+[66] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, pages 15943-15953, 2023. 1  
+[67] Yinwei Wu, Xianpan Zhou, Bing Ma, Xuefeng Su, Kai Ma, and Xinchao Wang. Ifadapter: Instance feature control for grounded text-to-image generation. arXiv:2409.08240, 2024.3  
+[68] Jiayu Xiao, Henglei Lv, Liang Li, Shuhui Wang, and Qingming Huang. R&b: Region and boundary aware zero-shot grounded text-to-image generation. In ICLR, 2023. 2, 3, 6, 7, 5  
+[69] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv:2409.11340, 2024. 3
+
+[70] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In ICCV, pages 7452-7461, 2023. 2, 3, 6, 7, 5  
+[71] Xingqian Xu, Jiayi Guo, Zhangyang Wang, Gao Huang, Irfan Essa, and Humphrey Shi. Prompt-free diffusion: Taking" text" out of text-to-image diffusion models. In CVPR, pages 8682-8692, 2024. 3  
+[72] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In CVPR, pages 18381-18391, 2023. 3, 1, 6  
+[73] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In ICML, 2024. 3, 6, 7, 5  
+[74] Ling Yang, Zhilong Zhang, Zhaochen Yu, Jingwei Liu, Minkai Xu, Stefano Ermon, and Bin Cui. Cross-modal contextualized diffusion models for text-guided visual generation and editing. arXiv:2402.16627, 2024. 3  
+[75] Yang Yang, Wen Wang, Liang Peng, Chaotian Song, Yao Chen, Hengjia Li, Xiaolong Yang, Qinglin Lu, Deng Cai, Boxi Wu, et al. Lora-composer: Leveraging low-rank adaptation for multi-concept customization in training-free diffusion models. arXiv:2403.11627, 2024. 1  
+[76] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In CVPR, pages 14246-14255, 2023. 3  
+[77] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv:2308.06721, 2023. 1  
+[78] Ziyang Yuan, Mingdeng Cao, Xintao Wang, Zhonggang Qi, Chun Yuan, and Ying Shan. Customnet: Zero-shot object customization with variable-viewpoints in text-to-image diffusion models. arXiv:2310.19784, 2023. 1  
+[79] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 3, 6, 7, 8, 5, 9  
+[80] Lirui Zhao, Tianshuo Yang, Wenqi Shao, Yuxin Zhang, Yu Qiao, Ping Luo, Kaipeng Zhang, and Rongrong Ji. Diffree: Text-guided shape free object inpainting with diffusion model. arXiv:2407.16982, 2024. 3, 1  
+[81] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. In NIPS, 2023. 3  
+[82] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Mige  $^{+ + }$  : Advanced multi-instance generation controller for image synthesis. arXiv:2407.02329, 2024. 3  
+[83] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. In CVPR, pages 6818-6828, 2024. 1, 2, 3, 6, 7, 8, 5, 9

2025/VODiff_ Controlling Object Visibility Order in Text-to-Image Generation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb0d976c972f2796dac6214a643fe2176446d694a1c63f6eb8848cdde834337f
+size 692842

2025/VSNet_ Focusing on the Linguistic Characteristics of Sign Language/1135c1a4-3af0-45ac-a497-2c9557104d8c_content_list.json

@@ -0,0 +1,1666 @@
+[
+    {
+        "type": "text",
+        "text": "VSNet: Focusing on the Linguistic Characteristics of Sign Language",
+        "text_level": 1,
+        "bbox": [
+            153,
+            130,
+            843,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yuhao Li $^{1}$ , Xinyue Chen $^{1}$ , Hongkai Li $^{1}$ , Xiaorong Pu $^{1,3}$ , Peng Jin $^{2}$ , Yazhou Ren $^{1,3,*}$ $^{1}$ School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, China;  $^{2}$ Sichuan Provincial Key Lab of Philosophy and Social Science for Language Intelligence in Special Education, Leshan Normal University, Leshan, China;  $^{3}$ Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China, Shenzhen, China",
+        "bbox": [
+            156,
+            179,
+            864,
+            271
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            304,
+            326,
+            319
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sign language is a visual language expressed through complex movements of the upper body. The human skeleton plays a critical role in sign language recognition due to its good separation from the video background. However, mainstream skeleton-based sign language recognition models often overly focus on the natural connections between joints, treating sign language as ordinary human movements, which neglects its linguistic characteristics. We believe that just as letters form words, each sign language gloss can also be decomposed into smaller visual symbols. To fully harness the potential of skeleton data, this paper proposes a novel joint fusion strategy and a visual symbol attention model. Specifically, we first input the complete set of skeletal joints, and after dynamically exchanging joint information, we discard the parts with the weakest connections to other joints, resulting in a fused, simplified skeleton. Then, we group the joints most likely to express the same visual symbol and discuss the joint movements within each group separately. To validate the superiority of our method, we conduct extensive experiments on multiple public benchmark datasets. The results show that, without complex pretraining, we still achieve new state-of-the-art performance. The code is available at https://github.com/atinyboy/VSNet.",
+        "bbox": [
+            89,
+            335,
+            483,
+            684
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            91,
+            729,
+            220,
+            744
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sign language is a visual language expressed through body movements and serves as the primary communication tool for the deaf community [13]. Although sign language and spoken language have distinctly different grammatical structures [1, 7], they share common linguistic features [39, 44, 59]. Accurate sign language translation can facilitate communication between deaf and hearing people, fostering a more inclusive and accessible society. Among",
+        "bbox": [
+            89,
+            753,
+            482,
+            876
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/0ef7164ed9cc0e33de67ef10e8993b7a4a21ae4ff844e5a166b44fb3d111af05.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            303,
+            614,
+            375
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/44fe15157fcbe074d6c1d7cd80d06ce1f7f18531872187c473e9164ff0c00f43.jpg",
+        "image_caption": [
+            "gloss: 'accent'"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            617,
+            303,
+            709,
+            375
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/f9bf732e363bf30348c2d295f318f5d05c852e71aa7b797239453d99c7a09b96.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            710,
+            303,
+            803,
+            375
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/ca5a47bc0d09f7265a256a59d8607956570ca3711d2ec5e38228d7f2796734ad.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            805,
+            303,
+            897,
+            375
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/fdeac63056da7ab56e05b49bb7c6ccfa708b60992d811714809658431cc4ca36.jpg",
+        "image_caption": [
+            "Figure 1. In the sign for \"accent\", the index and middle fingers (or only index finger) are extended toward the neck while other fingers are curled, forming the \"pointing\" VS. In the sign for \"algebra\", movements can be decomposed into multiple VS such as \"crossed arms,\" \"fist,\" and \"thumb-up.\""
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            396,
+            614,
+            467
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/18fa9e3dbae1a5067929bcacce7062c5be4971a7cddd1ad711e479f11e05de4b.jpg",
+        "image_caption": [
+            "gloss:'algebra'"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            617,
+            396,
+            707,
+            467
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/ad3e18e00f9ec6d67c41252c0742477cc98782276705a52e360e368fa82d5f19.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            710,
+            396,
+            803,
+            467
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/10785b8de7ae97caba6d78a7d9b114606070865a8f7c70ff8eea76801732b85a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            805,
+            396,
+            895,
+            467
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "the key subtasks in sign language translation is isolated sign language recognition (ISLR), which focuses on translating individual glosses. ISLR has garnered significant attention as it aids people in accurately understanding numerous gestures and quickly learning sign language.",
+        "bbox": [
+            511,
+            594,
+            906,
+            671
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In recent years, human skeletons have been increasingly applied in the ISLR field due to their strong operability and clear separation from video backgrounds. Although most skeleton-based recognition methods have achieved promising results, they typically treat sign language as typical human motion [12, 21, 23, 30, 31, 41, 58], overlooking the intrinsic relationships among joints and the unique linguistic characteristics [25, 52] of sign language itself. We believe that: (1) The importance of joints in sign language is not equivalent; joints with stronger connections to others are more significant than isolated joints, and connections among joints with similar movement tendencies hold more meaning than those among dispersed joints. (2) Similar to how letters form words, each sign language gloss can be decomposed into smaller visual symbols (VS). While VS",
+        "bbox": [
+            511,
+            674,
+            908,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author (yazhou.ren@uestc.edu.cn).",
+        "bbox": [
+            107,
+            886,
+            382,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "24320",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "lack explicit meaning on their own, from a skeletal perspective, they capture distinct movement patterns of joints, effectively decomposing sign language actions (see Fig. 1). Therefore, our work focuses on these two core aspects:",
+        "bbox": [
+            89,
+            90,
+            482,
+            151
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Core 1: Focus on identifying essential joints. Including too many joints can impede training, while too few can lead to information loss. On the one hand, many prior models have used manually trimmed, simplified skeletons, but manually selecting the optimal joints is not trivial. On the other hand, enabling a model to automatically choose key joints can be computationally intensive. Fortunately, the dynamically learned skeletons from GCNs approximate real structures more closely, enabling us to reduce the dependency on an initial skeleton. By computing the weight of each joint, we can identify and retain the most important joints while discarding less critical local joints. Specifically, for the complete hand skeleton, we discard a small number of joints based on computed weights after each GCN layer, relearning new skeleton connections in the subsequent GCN layer. We control the number of joints discarded within each region to ensure that the final GCN output retains a recognizable structure. Notably, even joints with the lowest weights undergo at least one convolution, so their information is not entirely discarded.",
+        "bbox": [
+            91,
+            154,
+            482,
+            455
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Core 2: Focus on emphasizing visual symbols (VS) in sign language. If each joint is considered the smallest element of a sign language action, then the combination of joints across all frames becomes too vast to compute self-attention. In practice, sign language is not random motion; it is designed for human recognition. When a signer interprets sign language, they do not scrutinize the position of every joint but instead form a general impression of each hand. Thus, in this work, we concentrate on recognizing VS. Unfortunately, isolating each VS from movements is as challenging as the entire SLR task itself. Therefore, we adopt a simplified approach, segmenting the skeleton into a generalized VS in a fixed manner, as shown in Fig. 2. We then design two self-attention modules to capture spatial and temporal relationships between these VS. Empirically, we find that this approach significantly enhances the model's discriminative power.",
+        "bbox": [
+            91,
+            458,
+            482,
+            715
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our main contributions are summarized as follows:",
+        "bbox": [
+            109,
+            717,
+            449,
+            731
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose a novel module with an automated skeleton simplification function for sign language recognition. This module is designed to dynamically learn global skeleton connections while progressively discarding isolated local joints.",
+            "- Considering the linguistic characteristics of sign language, we introduce a new self-attention model called VSformer, which captures temporal and spatial patterns in sign language motions and extracts knowledge from the human skeleton based on the concept of VS.",
+            "- We conduct extensive experiments on ISLR to validate"
+        ],
+        "bbox": [
+            89,
+            734,
+            482,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/643f5b9a8f8824d4be18635ddea1b3f7e73bb10c6cc91a7c3c93649d5a9787aa.jpg",
+        "image_caption": [
+            "Figure 2. Decomposing generalized visual symbols from videos"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            90,
+            893,
+            256
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "the effectiveness of our proposed approach. Compared to previous methods, our approach achieves SoTA performance on four public benchmarks, namely WLASL [32], MSASL [26], NMFs-CSR [16], and Slovo [27].",
+        "bbox": [
+            524,
+            310,
+            903,
+            371
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Works",
+        "text_level": 1,
+        "bbox": [
+            513,
+            383,
+            663,
+            398
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Sign Language Recognition",
+        "text_level": 1,
+        "bbox": [
+            513,
+            409,
+            759,
+            425
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Sign language is the primary means of communication within the deaf community. With significant advances in machine learning and computer vision over the past decade, exploring sign language recognition (SLR) has become increasingly important [28, 43, 45], as it can automatically interpret sign language and help deaf and mute individuals communicate more easily in daily life. Unlike directly translating sign language sentences, isolated sign language recognition (ISLR) is a subtask in the field of sign language translation [45], aimed at translating each individual gloss [40, 53]. ISLR is not only useful for guiding beginners in learning sign language and assisting in the creation of new sign language gestures, but it also lays a solid foundation for SLR [1, 4, 53]. Research in this area can be divided into two categories based on the input modality: RGB-based methods and skeleton-based methods.",
+        "bbox": [
+            511,
+            431,
+            906,
+            672
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "RGB-based Recognition. RGB-based methods, similar to other video recognition tasks, typically use CNNs to extract visual features from RGB videos [5, 15, 19, 20, 29, 55-57], and then aggregate temporal and spatial information to predict sign language. However, RGB-based models face two significant challenges: first, they are heavily affected by changes in the video background, and second, video data occupies a large amount of memory resources.",
+        "bbox": [
+            511,
+            674,
+            906,
+            794
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Skeleton-based Recognition. Compared to RGB images, human skeletons [8, 47, 50, 54] describe the motion information conveyed in sign language videos using simplified points and edges, reducing interference from complex backgrounds and the appearance of the signer. Moreover, operations on three-dimensional spatial points with actual physical meanings offer stronger interpretability and",
+        "bbox": [
+            511,
+            795,
+            906,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "24321",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/a0fe65ed0df5a97890f903093b93c51daf44c0b86f7de44230e2f011d4ddd927.jpg",
+        "image_caption": [
+            "Figure 3. Overview of our VSNet. Here, Down sample consists of a convolution layer with a kernel size of 7 and a stride of 2 along the temporal dimension, followed by a batch norm layer, Average Pooling represent transforming the data into a single-dimensional feature by taking the mean across both temporal and spatial dimensions. For details on the Skeleton- $(a,b,c)$ , please refer to Fig. 4."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            99,
+            88,
+            890,
+            345
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "lower memory consumption. In earlier works, identifying the correct human keypoints in a video was a major challenge. However, with the development of deep learning, human pose estimation models have become increasingly reliable [10, 24, 51].",
+        "bbox": [
+            89,
+            422,
+            483,
+            500
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2. Human Skeleton",
+        "text_level": 1,
+        "bbox": [
+            89,
+            511,
+            256,
+            526
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Due to the natural physical connections between skeleton joints, Graph Convolutional Network (GCN) has been widely applied to model spatio-temporal relationships [12, 23, 31, 34, 41, 46]. GCN require predefined connectivity of the key points to construct the graph network. However, the physically natural skeleton connections may not always be the most suitable edges for recognition. Relying solely on the skeleton itself to construct the graph can limit the information flow between nodes. To address this issue, SL-GCN [21] employs a dynamic graph with decoupling between different frames, while TD-GCN [37] adds supernodes to control the connections between nodes. A common drawback of these approaches is that they require complex networks to model the key points. Furthermore, while dynamic graphs alleviate the limitations of the original skeleton, blindly allowing all key points to connect introduces redundant information, leading to suboptimal models.",
+        "bbox": [
+            88,
+            535,
+            482,
+            792
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Therefore, we argue that relying solely on GCN does not fully exploit the true potential of skeleton data.",
+        "bbox": [
+            89,
+            792,
+            483,
+            824
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In recent years, attention-based methods [3, 6, 9, 38, 42, 48] have also made progress in human skeleton recognition. Unlike GCN, attention mechanisms do not require the researcher to provide information on joint connections, instead equally attending to all nodes. However, captur",
+        "bbox": [
+            89,
+            825,
+            483,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ing the correlations between all joints across all frames requires substantial memory resources. To address this, SLGTformer [49] decomposes the spatio-temporal pose sequence into spatial graphs and temporal windows, while SkateFormer [11] partitions joints and frames and performs self-attention on each partition. Based on these theories, we design a model where GCNs and self-attention networks complement each other. GCNs are responsible for exploring the spatial correlations between human joints, while self-attention networks focus on capturing the motion of the skeleton across multiple time sequences and spatial combinations.",
+        "bbox": [
+            511,
+            422,
+            906,
+            603
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Methods",
+        "text_level": 1,
+        "bbox": [
+            513,
+            622,
+            612,
+            636
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Data Pre-Processing",
+        "text_level": 1,
+        "bbox": [
+            511,
+            647,
+            705,
+            665
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Key Point Extraction. Sign language is a visual language primarily conveyed through hand and body movements, with facial expressions also playing a crucial role. However, the human skeleton mainly describes bones and joints, excluding muscles, making it unable to accurately capture facial expressions in sign language videos. Our investigation revealed that a significant number of signers, especially beginners, do not exhibit extensive facial expressions during sign language communication. Therefore, our main experiments exclude facial expressions, while the appendix provides additional details on how the concept of VS can be extended to the face.",
+        "bbox": [
+            511,
+            672,
+            906,
+            852
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In previous GCN models, finger joints were often simplified to avoid making the graph too complex, leading to a loss of information. In this study, we selected a complete",
+        "bbox": [
+            511,
+            854,
+            908,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "24322",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/3b302e9367d27a69b446d6fd883fe21af9ef8d9b70365c49823d0a11074b393e.jpg",
+        "image_caption": [
+            "Figure 4. Illustration of the skeleton and grouping type-1.  $V_{0}$  represents a localization node, the dashed box indicates nodes that are grouped together and hollow dots indicate dropped nodes. Please note that (a) is the input real skeleton diagram, while (b) and (c) are simplified representations drawn after dropping for ease of analysis."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            90,
+            584,
+            215
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/b715aea289b6efb25c9f8e0f180402f2a4f181c1c5aefe14eb1470b89cdc6b34.jpg",
+        "image_caption": [
+            "Figure 5. Skeleton groupings type-2 and type-3. The other hand is grouped in a mirrored fashion."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            620,
+            90,
+            898,
+            234
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "set of 44 key points from both hands out of the 133 human body joints extracted using HRNet [24, 51], as shown in Fig. 4(a).",
+        "bbox": [
+            88,
+            311,
+            480,
+            357
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Data Enhancing. Given a video  $X \\in \\mathbb{R}^{T \\times V \\times C}$ , where T, V, and C represent the number of frames, joints, and coordinates respectively. We replace abnormal points—such as those out of range (less than 0, more than 512) or clearly distorted—caused by model extraction errors, with the average value from adjacent frames. Additionally, to standardize body positions and movement amplitudes across different signers in the sign language videos, we first use the hollow node  $V_{0}$  in Fig. 4 as the reference node for the entire video. Then, all key points in each frame are shifted based on this reference coordinate, as follows:",
+        "bbox": [
+            88,
+            357,
+            482,
+            523
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nX (t, v, c) = X (t, v, c) - X (t, v _ {0}, c) + [ x, y, z ] \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            130,
+            539,
+            482,
+            556
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $[x,y,z]$  is the average value of all the coordinates of  $V_{0}$  in the dataset.",
+        "bbox": [
+            89,
+            569,
+            483,
+            599
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In the time dimension, we randomly select 64 frames to standardize the video length and perform random mirroring and shifting of the video to generate augmented data.",
+        "bbox": [
+            89,
+            599,
+            483,
+            646
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Self-Pacing Dropping Block (SPD)",
+        "text_level": 1,
+        "bbox": [
+            89,
+            657,
+            390,
+            675
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As shown in Fig. 3, our SPD consists of two components: \"Keypoint Activation\" and \"Weak Joints Dropping\". The input is a complete hand skeleton where the node information is independent, and the output is a high-dimensional feature skeleton after activation. The entire block is stacked three times, where the third block only outputs features and does not build a new skeleton.",
+        "bbox": [
+            89,
+            683,
+            482,
+            787
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Keypoint Activation Layer. In this part, we define the interaction of each joint with others as an activation. We employ a decoupled graph convolution [21] to perform this activation, represented as:",
+        "bbox": [
+            89,
+            789,
+            483,
+            849
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {o u t} ^ {t} = \\frac {1}{p} \\sum_ {k = 0} ^ {p} D ^ {- \\frac {1}{2}} A _ {k} D ^ {- \\frac {1}{2}} \\phi \\left(F _ {i n} ^ {t}\\right) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            166,
+            864,
+            482,
+            904
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $A$  is the learnable adjacency matrix initialized as  $I + In + Out$ ,  $D$  presents the diagonal degree of  $A$ . Here,  $I$  denotes self-connections, while  $In$  and  $Out$  represent inward and outward connections, respectively. The symbol  $\\phi$  represents a linear layer, and  $p$  represents the decoupling quantity, with a different  $A$  applied each time, followed by averaging the results. Following the graph convolution, we apply the simplest Temporal Convolutional Network (TCN) to capture features along the temporal dimension.",
+        "bbox": [
+            511,
+            311,
+            906,
+            449
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Weak Joints Dropping Layer. Compared to the simplified skeletons used in [21, 49], our approach utilizes a complete set of skeleton keypoints that contains richer information. However, this comes at the cost of significantly increasing the model's burden and raising the risk of introducing outlier points. To leverage these additional points with minimal overhead, we have designed an adaptive skeleton simplification strategy.",
+        "bbox": [
+            511,
+            452,
+            908,
+            575
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Before formally introducing this layer, it is necessary to present our approach to skeleton decomposition. Based on the definition of VS described in Introduction, the decomposition concept involves grouping joints that exhibit strong overall correlation or frequently form VS together in sign language. These groupings can be viewed as a generalized or macro-level VS. A simple approach is grouping each finger and arm separately, but this creates too many groups, potentially affecting later interactions. Based on this, we conducted various experiments on VS, including treating individual fingers and distant fingers as VS. However, these experiments did not yield optimal results (the average performance in WLASL dataset decreased by  $1.20\\%$  and  $1.64\\%$ , respectively). We believe this is because the information from a single finger is insufficient to form a meaningful VS, while the information disparity between distant fingers is even greater. This also serves as the skeleton grouping basis for Figs. 4 and 5.",
+        "bbox": [
+            511,
+            578,
+            908,
+            851
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "After graph convolution in Eq. 2,  $A(i,j)$  has learned the weighted edge between node  $i$  and node  $j$ . These edges reflect the connection strength between  $i$  and  $j$ . Since these",
+        "bbox": [
+            511,
+            854,
+            908,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "24323",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "connections are not normalized via softmax, we can use  $A$  as a guide to calculate the total correlation of all keypoints, Cor, as follows:",
+        "bbox": [
+            89,
+            90,
+            483,
+            135
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nC o r _ {i} = \\sum_ {j} \\max  \\left(A _ {0} (i, j), A _ {1}, \\dots , A _ {p}\\right) \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            153,
+            148,
+            482,
+            180
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "A lower  $Cor$  to other keypoints indicates that the point is more independent. We refer to the two most independent points in each group as weak joints, which are dropped from the skeleton. To ensure that the number of keypoints in each group remains consistent, the groups involving both arms will have two fewer joints dropped.",
+        "bbox": [
+            89,
+            184,
+            483,
+            275
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3.VSformer",
+        "text_level": 1,
+        "bbox": [
+            89,
+            282,
+            205,
+            297
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After the joint fusion block, we obtained six different joint groupings, as shown in Fig. 4(c). Based on the concept of skeletal disassembly, we treat each grouping as a generalized VS. We now need to focus on two main questions: Q1, what are the relationships between different VS? And Q2, do these VS remain consistent throughout the entire video, or do they change after a brief period?",
+        "bbox": [
+            89,
+            306,
+            483,
+            411
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Inspired by [11], which proposes an efficient model for partitioning human limbs in action recognition, we employ two types of self-attention to constrain the model's focus on the aforementioned questions. Specifically, we calculate spatial pairing attention among different combinations within a frame to address Q1, and we compute temporal window attention for both global and local motion of the combinations across different frames to address Q2.",
+        "bbox": [
+            89,
+            412,
+            483,
+            531
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Spatial Pairing Attention (SPA). Given a frame  $F_{t} \\in \\mathbb{R}^{G \\times C}$ , where  $G = 6$  represents VS groupings, the ideal approach would involve calculating attention among all VS pairwise. However, this is inefficient and often purposeless, as VS that are far apart in the skeletal structure tend to have weak correlations. For instance, in the sign language gesture for \"accent\", it is unnecessary to compare the fingers of the left hand with those of the unused right hand. Based on this observation, we ultimately partition  $F_{t}$  into pairs as follows:",
+        "bbox": [
+            89,
+            532,
+            483,
+            681
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {t} = \\left[ P _ {\\text {a r m}}, P _ {\\text {l e f t}}, P _ {\\text {r i g h t}} \\right] \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            209,
+            683,
+            480,
+            700
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $P_{\\mathrm{arm}} \\in \\mathbb{R}^{2 \\times C}$  represents the grouping of both arms,  $P_{\\mathrm{left}}$  denotes the four fingers of the left hand, and  $P_{\\mathrm{right}}$  represents the remaining four fingers of the right hand. Represented as:",
+        "bbox": [
+            89,
+            704,
+            483,
+            763
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS P A _ {\\text {o u t}} = \\operatorname {c o n c a t} \\left(\\operatorname {s o f t m a x} \\left(\\frac {Q _ {i} K _ {i} ^ {T}}{\\sqrt {h}}\\right) V _ {i}\\right) \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            140,
+            768,
+            482,
+            804
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $Q_{i}, K_{i}, V_{i}$  are obtained from  $P_{i}$  after mapping through linear layers. The attention is actually calculated among the 8 joints within each pairing and is ultimately concatenated to form the output.",
+        "bbox": [
+            89,
+            809,
+            482,
+            869
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Temporal Window Attention (TWA). Depending on different sign language videos, the VS can be classified into",
+        "bbox": [
+            89,
+            869,
+            483,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/621111fb7906cca097896dddac1e940b4f52be4a0b9723b1c4add3959b904881.jpg",
+        "image_caption": [
+            "Figure 6. TWA windows. The small ticks represent frames, and the large ticks represent windows."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            95,
+            903,
+            135
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "two types: short duration and long duration, representing their respective duration periods. Accordingly, our TWA is also divided into two types: TWA-s and TWA-l. Given a video  $X \\in \\mathbb{R}^{T \\times G \\times C}$ , where  $T$  represents the total number of frames, the timeline is divided into  $n$  windows, each containing  $k$  frames, such that  $T = n \\times k$ , as shown in Fig. 6. For TWA-s, attention is computed within each window. For TWA-l, frames with the same index are selected sequentially from each window. Finally, the results from both types of attention are concatenated to produce the output, represented as:",
+        "bbox": [
+            511,
+            208,
+            906,
+            375
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nT W A _ {o u t} = \\operatorname {s o f t m a x} \\left(\\phi_ {1} (W) \\phi_ {2} ^ {T} (W) + b\\right) \\phi_ {3} (W) \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            535,
+            385,
+            906,
+            404
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $W$  represents the selected windows,  $\\phi_{n}$  denotes dividing the  $C$  dimension into three equal parts and selecting the  $n$ -th part, and  $b$  represents the added bias.",
+        "bbox": [
+            511,
+            414,
+            905,
+            458
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "It is worth noting that the discussions on Q1 and Q2 are not fully addressed by these two attention layers. The SPA only addresses the relationships between paired VS, while the windows defined in TWA do not entirely correspond to the duration of the VS. Therefore, we incorporate a simple GCN and TCN to enhance the model's ability to address the remaining scenarios. The connection method of SPA and TWA is shown in the right side of Fig. 3. After passing the input data  $X \\in \\mathbb{R}^{T \\times G \\times C}$  through a linear layer, we divide its channel dimension  $C$  into four equal parts, sending each part into one of the four modules and later concatenated. Additionally, unlike the Decoupled GCN in Sect. 3.2, the GCN here utilizes an adjacency matrix shared across certain frames, which significantly reduces the computational load.",
+        "bbox": [
+            511,
+            460,
+            906,
+            672
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4. Multi-Grouping Ensemble",
+        "text_level": 1,
+        "bbox": [
+            511,
+            680,
+            754,
+            696
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the previous sections, we introduced the concept of skeletal disassembly based on VS. However, since we cannot precisely determine the true VS composition for each sign language, using a fixed disassembly method appears to be a compromise, potentially weakening interpretability. Therefore, we adopt an ensemble strategy to reduce such discrepancies. Specifically, we designed two additional skeleton grouping schemes, as shown in the Fig. 5, and trained the model separately to obtain the output  $q$  from the final fully connected layer for each scheme. Following [21], we aggregate these results using a weighted sum, represented as:  $q_{\\mathrm{total}} = \\alpha_1q_{\\mathrm{type - 1}} + \\alpha_2q_{\\mathrm{type - 2}} + \\alpha_3q_{\\mathrm{type - 3}} + \\alpha_4q_{\\mathrm{bone}}$ , where  $\\alpha$  represents the weights. For a detailed discussion on  $\\alpha$ ,",
+        "bbox": [
+            511,
+            704,
+            906,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "24324",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/70acf9951ff5aaf6d8c7cd56e115a61e787330cd2da58e5baa16e5c647a0f337.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"3\">Methods</td><td colspan=\"4\">MSASL1000</td><td colspan=\"4\">MSASL200</td><td colspan=\"4\">MSASL100</td></tr><tr><td colspan=\"2\">P-I</td><td colspan=\"2\">P-C</td><td colspan=\"2\">P-I</td><td colspan=\"2\">P-C</td><td colspan=\"2\">P-I</td><td colspan=\"2\">P-C</td></tr><tr><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td></tr><tr><td>RGB-based</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>HMA [15]</td><td>49.16</td><td>69.75</td><td>46.27</td><td>68.60</td><td>66.30</td><td>84.03</td><td>67.47</td><td>84.03</td><td>73.45</td><td>89.70</td><td>74.59</td><td>89.70</td></tr><tr><td>I3D [5]*</td><td>-</td><td>-</td><td>57.69</td><td>81.05</td><td>-</td><td>-</td><td>81.97</td><td>93.79</td><td>-</td><td>-</td><td>81.76</td><td>95.16</td></tr><tr><td>BSL [2]</td><td>64.71</td><td>85.59</td><td>61.55</td><td>84.43</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>TCK [33]†</td><td>-</td><td>-</td><td>-</td><td>-</td><td>80.31</td><td>91.82</td><td>81.14</td><td>92.24</td><td>83.04</td><td>93.46</td><td>83.91</td><td>93.52</td></tr><tr><td>skeleton-based</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>ST-GCN [58]</td><td>36.03</td><td>59.92</td><td>32.32</td><td>57.15</td><td>52.91</td><td>76.67</td><td>54.20</td><td>77.62</td><td>59.84</td><td>82.03</td><td>60.79</td><td>82.96</td></tr><tr><td>BEST [60]</td><td>58.82</td><td>81.18</td><td>54.87</td><td>80.05</td><td>76.60</td><td>91.54</td><td>76.75</td><td>91.95</td><td>80.98</td><td>95.11</td><td>81.24</td><td>95.44</td></tr><tr><td>SignBERT [14]†</td><td>59.80</td><td>81.86</td><td>57.06</td><td>80.94</td><td>77.34</td><td>91.10</td><td>78.02</td><td>91.48</td><td>81.37</td><td>93.66</td><td>82.31</td><td>93.76</td></tr><tr><td>SignBERT+ [17]†</td><td>62.42</td><td>83.49</td><td>60.15</td><td>82.44</td><td>78.51</td><td>-</td><td>79.35</td><td>-</td><td>84.94</td><td>-</td><td>85.23</td><td>-</td></tr><tr><td>MASA [61]†</td><td>63.47</td><td>83.89</td><td>60.79</td><td>83.29</td><td>79.25</td><td>92.86</td><td>79.70</td><td>93.33</td><td>83.22</td><td>95.24</td><td>83.19</td><td>95.46</td></tr><tr><td>SSRL [62]†</td><td>65.22</td><td>85.09</td><td>62.68</td><td>84.38</td><td>82.93</td><td>94.78</td><td>83.43</td><td>95.03</td><td>86.26</td><td>96.96</td><td>86.63</td><td>96.79</td></tr><tr><td>Ours</td><td>70.71</td><td>86.36</td><td>68.38</td><td>85.71</td><td>84.62</td><td>94.04</td><td>85.19</td><td>94.08</td><td>86.26</td><td>95.24</td><td>86.68</td><td>95.39</td></tr><tr><td>Ours(4-crops)</td><td>73.18</td><td>88.64</td><td>70.88</td><td>87.59</td><td>86.17</td><td>95.51</td><td>86.64</td><td>95.61</td><td>88.11</td><td>97.09</td><td>88.32</td><td>97.18</td></tr></table>",
+        "bbox": [
+            91,
+            88,
+            906,
+            342
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/948ee5102452f7798c718f2b06fb8bcf7dfdc78e35f33a87fe17a339fff5a94e.jpg",
+        "table_caption": [
+            "Table 1. Comparison with previous work on MSASL. “†” indicates that the model was pre-trained, and “*” indicates methods that applied result fusion with other modalities in post-processing. Bold indicates the best performance, while underlined denotes the second-best."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"3\">Methods</td><td colspan=\"4\">WLASL2000</td><td colspan=\"4\">WLASL300</td><td colspan=\"4\">WLASL100</td></tr><tr><td colspan=\"2\">P-I</td><td colspan=\"2\">P-C</td><td colspan=\"2\">P-I</td><td colspan=\"2\">P-C</td><td colspan=\"2\">P-I</td><td colspan=\"2\">P-C</td></tr><tr><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td></tr><tr><td>RGB-based</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>I3D [5]*</td><td>32.48</td><td>57.31</td><td>-</td><td>-</td><td>56.14</td><td>79.94</td><td>56.24</td><td>78.38</td><td>65.89</td><td>84.11</td><td>67.01</td><td>84.58</td></tr><tr><td>HMA [15]</td><td>37.91</td><td>71.26</td><td>35.90</td><td>70.00</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>BSL [2]</td><td>46.82</td><td>79.36</td><td>44.72</td><td>78.47</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>TCK [33]†</td><td>-</td><td>-</td><td>-</td><td>-</td><td>68.56</td><td>89.52</td><td>68.75</td><td>89.41</td><td>77.52</td><td>91.08</td><td>77.55</td><td>91.42</td></tr><tr><td>NLA-SLR V64 [63]*</td><td>51.15</td><td>83.43</td><td>48.14</td><td>82.20</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>skeleton-based</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>P3D [30]</td><td>44.47</td><td>79.69</td><td>42.18</td><td>78.52</td><td>67.18</td><td>89.01</td><td>67.62</td><td>89.24</td><td>76.71</td><td>91.97</td><td>78.27</td><td>92.97</td></tr><tr><td>BEST [60]</td><td>46.25</td><td>79.33</td><td>43.52</td><td>77.65</td><td>67.66</td><td>89.22</td><td>68.31</td><td>89.57</td><td>77.91</td><td>91.47</td><td>77.83</td><td>92.50</td></tr><tr><td>SignBERT [14]†</td><td>47.46</td><td>83.32</td><td>45.17</td><td>82.32</td><td>70.36</td><td>88.92</td><td>71.17</td><td>89.36</td><td>79.07</td><td>93.80</td><td>80.05</td><td>94.17</td></tr><tr><td>MASA [61]†</td><td>49.06</td><td>82.90</td><td>46.91</td><td>81.80</td><td>73.65</td><td>91.77</td><td>74.33</td><td>92.13</td><td>83.72</td><td>93.80</td><td>84.47</td><td>94.30</td></tr><tr><td>NLA-SLR K64 [63]*</td><td>49.10</td><td>82.00</td><td>46.18</td><td>80.71</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SL-GCN [22]*</td><td>51.50</td><td>84.94</td><td>48.87</td><td>84.02</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DSTA [18]</td><td>53.68</td><td>-</td><td>51.17</td><td>-</td><td>79.97</td><td>-</td><td>80.56</td><td>-</td><td>82.38</td><td>-</td><td>83.09</td><td>-</td></tr><tr><td>Ours</td><td>53.54</td><td>83.18</td><td>51.18</td><td>82.03</td><td>79.04</td><td>94.01</td><td>79.74</td><td>94.25</td><td>84.50</td><td>94.57</td><td>85.25</td><td>94.33</td></tr><tr><td>Ours(4-crops)</td><td>55.98</td><td>87.07</td><td>53.54</td><td>86.04</td><td>80.09</td><td>94.76</td><td>80.85</td><td>94.92</td><td>85.66</td><td>94.96</td><td>86.25</td><td>95.08</td></tr></table>",
+        "bbox": [
+            91,
+            393,
+            906,
+            670
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Comparison with previous work on WLASL. “†” indicates that the model was pre-trained, and “*” indicates methods that applied result fusion with other modalities in post-processing. Bold indicates the best performance, while underlined denotes the second-best.",
+        "bbox": [
+            89,
+            680,
+            906,
+            709
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "please refer to Sect. 4.4. The term bone refers to the joint data of type-1 represented in vector form, generated by directing from the source joint to its target joint according to the natural connections of the human body.",
+        "bbox": [
+            89,
+            734,
+            482,
+            796
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            806,
+            223,
+            824
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Parameter Settings",
+        "text_level": 1,
+        "bbox": [
+            89,
+            832,
+            274,
+            848
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We employ the AdamW optimizer with a plateau learning rate schedule. The initial learning rate is set to  $1 \\times 10^{-3}$ , with a minimum learning rate of  $1 \\times 10^{-5}$  and a decay rate",
+        "bbox": [
+            89,
+            854,
+            483,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "of 0.5. The batch size is 32 and the loss function used is cross-entropy with label smoothing. Training spans a maximum of 200 epochs, with the first 15 epochs serving as a warm-up period starting from a learning rate of  $1 \\times 10^{-7}$ . A random sample of 64 frames is selected as input, and with a  $50\\%$  probability, image mirroring is applied. A random jitter in the range  $(-10, 10)$  is added to each joint point. In the last two layers of SPD, joints are randomly masked with a probability of  $(1 - p_{kr}) / 100 \\times \\text{epoch}$ , where  $p_{kr}$  is set to 0.9. All experiments were performed with NVIDIA A100.",
+        "bbox": [
+            511,
+            734,
+            906,
+            887
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "24325",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/4b0b34d9395710f0472923dc42a5987e8754994ed5e4212f2acabb0bcd27597a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">NMFs-CSR</td></tr><tr><td>Top-1</td><td>Top-5</td></tr><tr><td>RGB-based</td><td></td><td></td></tr><tr><td>I3D [5]*</td><td>64.4</td><td>88.0</td></tr><tr><td>HMA [15]</td><td>64.7</td><td>91.0</td></tr><tr><td>GLE-Net [36]</td><td>69.0</td><td>88.1</td></tr><tr><td>skeleton-based</td><td></td><td></td></tr><tr><td>ST-GCN [58]</td><td>59.9</td><td>86.8</td></tr><tr><td>BEST [60]</td><td>68.5</td><td>94.4</td></tr><tr><td>MASA [61]†</td><td>71.7</td><td>97.0</td></tr><tr><td>SignBERT [14]†</td><td>74.9</td><td>93.2</td></tr><tr><td>Ours</td><td>75.3</td><td>95.4</td></tr><tr><td>Ours(4-crops)</td><td>76.7</td><td>97.2</td></tr></table>",
+        "bbox": [
+            161,
+            88,
+            415,
+            289
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/6e0863849cf0b9d8091a93d7d3cd76aaf00e4eafdbdb07ae6c335cc5f9120a23.jpg",
+        "table_caption": [
+            "Table 3. Comparison on NMFs-CSR. Bold indicates the best performance, while underlined denotes the second-best."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">Slovo</td></tr><tr><td>Top-1</td><td>Top-5</td></tr><tr><td>Swin-large-48 [38]</td><td>55.66</td><td>-</td></tr><tr><td>MViTv2-small-48 [35]</td><td>62.18</td><td>-</td></tr><tr><td>MViTv2-small-32 [35]</td><td>64.09</td><td>-</td></tr><tr><td>Ours</td><td>76.53</td><td>93.08</td></tr><tr><td>Ours(4-crops)</td><td>77.83</td><td>95.17</td></tr></table>",
+        "bbox": [
+            142,
+            342,
+            434,
+            454
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Comparison with the baselines on Slovo. The final number indicates the number of frames input to the model.",
+        "bbox": [
+            89,
+            462,
+            482,
+            493
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Metrics. Following [22, 61, 63], we report top-1 and top-5 classification accuracy, as well as per-instance (P-I) and per-class (P-C) precision metrics. Each table includes results for single-crop and 4-crop inference, with all ablation studies conducted in the single-crop setting. For 4-crops inference, we sum the predictions from the three types of skeleton groupings along with the bone data to obtain the final prediction.",
+        "bbox": [
+            89,
+            518,
+            483,
+            640
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2. Dataset",
+        "text_level": 1,
+        "bbox": [
+            89,
+            650,
+            187,
+            664
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We conduct experiments on four publicly available benchmarks: WLASL, MSASL, NMFs-CSL, and Slovo. All training occurs on the training set, with hyperparameter tuning performed on the validation set.",
+        "bbox": [
+            89,
+            672,
+            482,
+            733
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "MSASL [26] is a widely used dataset for American Sign Language (ASL). It contains a total of 16,054, 5,287, and 4,172 samples in the training, validation, and test sets, respectively, with a vocabulary of 1,000 signs. Two subsets, consisting of the top 100 and 300 words, are defined as MSASL100 and MSASL300, respectively.",
+        "bbox": [
+            89,
+            734,
+            482,
+            824
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "WLASL [32] has a larger vocabulary but fewer total samples than MSASL. It contains 2,000 signs performed by over 100 signers, with 14,289, 3,916, and 2,878 samples in the training, validation, and test sets. Similarly, two subsets, WLASL100 and WLASL300, are created. Both",
+        "bbox": [
+            89,
+            824,
+            483,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "ASL datasets are collected from real-world scenarios, offering greater authenticity.",
+        "bbox": [
+            511,
+            90,
+            903,
+            121
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "NMFs-CSR [16] is a challenging Chinese Sign Language (CSL) dataset with a vocabulary of 1,067 and consists of 25,608 and 6,402 samples in the training and test sets, respectively.",
+        "bbox": [
+            511,
+            122,
+            903,
+            181
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Slovo [27] is the latest dataset for Russian Sign Language (RSL). It contains 20,000 videos across 1,000 isolated RSL gestures from 194 signers. Notably, the authors provide 42 hand landmarks based on MediaPipe, so we do not re-extract joint points.",
+        "bbox": [
+            511,
+            183,
+            906,
+            258
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3. Comparison with State-of-the-art Methods",
+        "text_level": 1,
+        "bbox": [
+            511,
+            270,
+            877,
+            286
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this section, we compare our model with several state-of-the-art approaches. For single-modality methods, we categorize them as either RGB-based or skeleton-based. For multi-modality methods, we report the performance of their single-modality models before ensemble.",
+        "bbox": [
+            511,
+            294,
+            903,
+            369
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "MSASL [26]. As shown in the Tab. 1, RGB-based methods have developed slowly in recent years, and our approach shows a significant improvement when compared to them. Among skeleton-based methods, SignBERT [14] and BEST [60] designed self-supervised learning strategies to pre-train the model in order to alleviate issues related to limited interpretability and overfitting due to the small size of the SLR datasets. NASA [61] and SSRL [62] address the problem of neglecting explicit motion information and lexical signs during pre-training. However, none of these methods take into account the linguistic characteristics of sign language itself. Compared to them, our method achieves a substantial top-1 improvement of  $7.96\\%$  in P-I and  $8.2\\%$  in P-C on the full dataset. However, SSRL performs slightly better than our single-crop method on the 100-class subset, which we speculate is due to SPD requiring a certain amount of data to effectively learn the droppable nodes.",
+        "bbox": [
+            511,
+            369,
+            906,
+            627
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "WLASL [32]. SAM-SLR [22] and NLA-SLR [63] are currently the two best-performing heavy multi-modality ensemble models. Among them, SL-GCN is the skeleton model used in SAM-SLR for later ensemble, employing a multi-stream approach to further improve performance. K64 and V64 are skeleton and RGB models in NLASLR that use 64-frame input, where the skeleton is represented as a heatmap sequence, allowing their keypoint encoder to share the same architecture with the video encoder. DSTA [18] was the previously most accurate model. Compared to their results, our top-1 accuracy achieved a  $2.30\\%$  improvement on WLASL2000.",
+        "bbox": [
+            511,
+            628,
+            905,
+            808
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "NMFs-CSL [16]. As shown in Tab. 3, NMFs-CSL is a balanced dataset in terms of class distribution, so we only report the P-I accuracy. The pretraining strategy of SignBERT [14] is highly effective on this dataset; however, its performance on top-5 is suboptimal. Our model does not use any additional training data, yet it shows a significant",
+        "bbox": [
+            511,
+            809,
+            905,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "24326",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/b906ec28865617bbba0152fb78ccecece5a76f63ed3253cd0e0fe93ef268ec63.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Skeleton</td><td colspan=\"2\">MSASL1000</td><td colspan=\"2\">MSASL200</td><td colspan=\"2\">MSASL100</td><td colspan=\"2\">WLASL2000</td><td colspan=\"2\">WLASL300</td><td colspan=\"2\">WLASL100</td><td colspan=\"2\">NMFs-CLR</td></tr><tr><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td></tr><tr><td>type-1</td><td>70.71</td><td>86.36</td><td>84.62</td><td>94.04</td><td>86.26</td><td>95.24</td><td>53.54</td><td>83.18</td><td>79.04</td><td>94.01</td><td>84.50</td><td>94.96</td><td>75.26</td><td>95.47</td></tr><tr><td>bone</td><td>68.41</td><td>85.88</td><td>82.41</td><td>93.38</td><td>84.68</td><td>94.06</td><td>50.76</td><td>82.45</td><td>74.70</td><td>91.17</td><td>84.11</td><td>93.80</td><td>73.63</td><td>94.31</td></tr><tr><td>type-2</td><td>70.35</td><td>86.10</td><td>85.28</td><td>94.11</td><td>87.45</td><td>95.64</td><td>53.02</td><td>83.32</td><td>78.44</td><td>93.41</td><td>83.72</td><td>93.80</td><td>74.70</td><td>94.86</td></tr><tr><td>type-3</td><td>70.35</td><td>86.10</td><td>84.92</td><td>94.11</td><td>85.60</td><td>95.11</td><td>52.26</td><td>82.49</td><td>78.14</td><td>93.41</td><td>84.88</td><td>93.41</td><td>74.87</td><td>95.55</td></tr></table>",
+        "bbox": [
+            102,
+            88,
+            893,
+            172
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/31e604748199a8431fe8b374c23dd2cd44acb63279e6389e154223e2aed8487b.jpg",
+        "table_caption": [
+            "Table 5. Performance of multi-grouping VSformer on three datasets."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">weighting</td><td colspan=\"2\">WLASL2000</td><td colspan=\"2\">MSASL1000</td><td colspan=\"2\">NMFs-CSL</td></tr><tr><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td></tr><tr><td>sum</td><td>55.98</td><td>87.07</td><td>73.18</td><td>88.64</td><td>76.65</td><td>97.17</td></tr><tr><td>acc</td><td>56.01</td><td>87.18</td><td>73.13</td><td>88.64</td><td>76.65</td><td>97.17</td></tr></table>",
+        "bbox": [
+            94,
+            220,
+            477,
+            277
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/a7028c8dd42754b2060fb0519da73a576752b18f71243ebf83ae6ff3357910fb.jpg",
+        "table_caption": [
+            "Table 6. Performance of the two ensemble methods"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">WLASL300</td><td colspan=\"2\">WLASL2000</td></tr><tr><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td></tr><tr><td>Linear</td><td>74.70</td><td>92.81</td><td>46.32</td><td>77.24</td></tr><tr><td>GCN</td><td>76.35</td><td>91.77</td><td>51.18</td><td>80.82</td></tr><tr><td>SPD(ours)</td><td>79.04</td><td>94.01</td><td>53.54</td><td>83.18</td></tr></table>",
+        "bbox": [
+            124,
+            316,
+            449,
+            396
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/a41391df11b191da55a27851e8377608ec051734a9a098a54cd740bb7ff4d18e.jpg",
+        "table_caption": [
+            "Table 7. Comparing our SPD with GCN and linear layers."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">SPA</td><td rowspan=\"2\">TMA</td><td colspan=\"2\">WLASL300</td><td colspan=\"2\">WLASL2000</td></tr><tr><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td></tr><tr><td>✓</td><td></td><td>77.69</td><td>93.41</td><td>52.43</td><td>82.24</td></tr><tr><td></td><td>✓</td><td>77.99</td><td>93.26</td><td>52.36</td><td>81.83</td></tr><tr><td>✓</td><td>✓</td><td>79.04</td><td>94.01</td><td>53.54</td><td>83.18</td></tr></table>",
+        "bbox": [
+            117,
+            435,
+            457,
+            515
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 8. Ablation study on the impact of our two types of attention mechanisms.",
+        "bbox": [
+            89,
+            525,
+            482,
+            554
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "improvement over SignBERT in overall performance, with a  $1.8\\%$  improvement in top-1 accuracy and a  $4\\%$  improvement in top-5 accuracy.",
+        "bbox": [
+            89,
+            580,
+            482,
+            625
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Slovo [27]. Slovo is a recently released RSL dataset, with only limited experimental data available. We provide an up-to-date benchmark for better evaluating this dataset. As shown in Tab. 4, the results for Swin [38] and MViTv2 [35] were provided as baselines by the dataset authors. Notably, the skeleton landmarks provided by the authors includes only 42 hand joints, without arm data, which partially limits our model's performance. Despite this, we still achieve results significantly above the baselines.",
+        "bbox": [
+            89,
+            626,
+            482,
+            762
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.4. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            89,
+            771,
+            243,
+            787
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this section, we conduct several key ablation experiments on the WLASL dataset [32] and report P-I accuracy to validate the effectiveness of our approach.",
+        "bbox": [
+            89,
+            794,
+            482,
+            840
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "SPD. We first replace SPD with two alternative methods. As shown in Tab. 7, \"Linear\" denotes a three-layer Conv-2D module with the same output dimension as SPD, while \"GCN\" refers to three decoupled graph convolution",
+        "bbox": [
+            89,
+            839,
+            483,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "layer without the weak joints dropping strategy. To ensure consistent output data shape, these two methods directly input the simplified 24 joint points. The complete SFB implementation achieves the highest accuracy, indicating that our dropping strategy better facilitates the extraction of sign language action features.",
+        "bbox": [
+            511,
+            224,
+            903,
+            314
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Multi-Grouping. As shown in Tab. 5, we present the final performance of four types of crops. The overall performance of each grouping type varies only slightly, mainly depending on the sample distribution of the dataset. However, integrating all four crops significantly enhances action classification performance. The performance of the bone-based grouping is slightly lower than that of joint-based results, primarily because our model is designed for points and is less sensitive to vector-based data.",
+        "bbox": [
+            511,
+            316,
+            905,
+            452
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "SAM-SLR [21] ultimately using manually set weights to control the ensemble. However, we argue that such weights lack interpretability. As shown in Tab. 6, we employed two methods to integrate the results: \"acc\" refers to using the individual top-1 accuracy of each crop as the weight, while \"sum\" represents directly summing the outputs of the final layer of the model. Both methods resulted in an about  $2\\%$  improvement in top-1 accuracy, demonstrating that our ensemble approach is both general and robust.",
+        "bbox": [
+            511,
+            455,
+            905,
+            592
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "VSformer. For our two types of self-attention, we replace one with a linear layer to test the performance of the other individually. As shown in Tab. 8, the full VSformer achieves significantly better performance, indicating that features in both temporal and spatial dimensions are crucial for distinguishing sign language actions.",
+        "bbox": [
+            511,
+            594,
+            905,
+            686
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            511,
+            705,
+            633,
+            722
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this work, we design a skeleton recognition model tailored specifically for ISLR, termed VSNet. The core idea of our approach is to focus on the truly essential elements of sign language. To this end, we first introduce a weak joints dropping strategy to prioritize more critical points in the skeleton. We then propose the VSformer to capture the linguistic characteristics of sign language, where Visual Symbols (VS) represent smaller components of sign language actions. We conduct extensive experiments to validate the effectiveness of the proposed VSNet, achieving state-of-the-art performance across four benchmarks.",
+        "bbox": [
+            511,
+            734,
+            905,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "24327",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            91,
+            90,
+            258,
+            107
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work was supported in part by National Natural Science Foundation of China (No. 62476052), Sichuan Science and Technology Program (No. 2024NSFSC1473), the open project of Sichuan Provincial Key Laboratory of Philosophy and Social Science for Language Intelligence in Special Education (No. YYZN-2023-3), the Humanities and Social Sciences Project of the Ministry of Education (No. 23YJA740013), and Central Guidance for Local Science and Technology Development Fund Projects (No. 2024ZYD0268).",
+        "bbox": [
+            89,
+            116,
+            485,
+            267
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            91,
+            282,
+            187,
+            299
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Nikolas Adaloglou, Theocharis Chatzis, Ilias Papastratis, Andreas Stergioulas, Georgios Th Papadopoulos, Vassia Zacharopoulou, George J Xydopoulos, Klimnis Atzakas, Dimitris Papazachariou, and Petros Daras. A comprehensive study on deep learning-based methods for sign language recognition. IEEE TMM, 24:1750-1762, 2021. 1, 2",
+            "[2] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. Bsl-1k: Scaling up co-articulated sign language recognition using mouthing cues. In ECCV. Springer, 2020. 6",
+            "[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 3",
+            "[4] Amanda Cardoso Duarte, Samuel Albanie, Xavier Giro Nieto, and Gül Varol. Sign language video retrieval with freeform textual queries. In CVPR, pages 14074-14084, 2022. 2",
+            "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 2, 6, 7",
+            "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 3",
+            "[7] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In CVPR, pages 5120-5130, 2022. 1",
+            "[8] Ke Cheng, Yifan Zhang, Congqi Cao, Lei Shi, Jian Cheng, and Hanqing Lu. Decoupling gcn with dropgraph module for skeleton-based action recognition. In ECCV, pages 536-553. Springer, 2020. 2",
+            "[9] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. NeurIPS, 34:9355-9366, 2021. 3",
+            "[10] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.3",
+            "[11] Jeonghyeok Do and Munchurl Kim. Skateformer: Skeletal-temporal transformer for human action recognition. arXiv preprint arXiv:2403.09508, 2024. 3, 5"
+        ],
+        "bbox": [
+            93,
+            309,
+            480,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[12] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In CVPR, pages 1110-1118, 2015. 1, 3",
+            "[13] Karen Emmorey. Language, cognition, and the brain: Insights from sign language research. Psychology Press, 2001. 1",
+            "[14] Hezhen Hu, Weichao Zhao, Wengang Zhou, Yuechen Wang, and Houqiang Li. Signbert: Pre-training of hand-model-aware representation for sign language recognition. In ICCV, pages 11087-11096, 2021. 6, 7",
+            "[15] Hezhen Hu, Wengang Zhou, and Houqiang Li. Hand-model-aware sign language recognition. In AAAI, pages 1558-1566, 2021. 2, 6, 7",
+            "[16] Hezhen Hu, Wengang Zhou, Junfu Pu, and Houqiang Li. Global-local enhancement network for nmf-aware sign language recognition. ACM transactions on multimedia computing, communications, and applications (TOMM), 17(3): 1-19, 2021. 2, 7",
+            "[17] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE TPAMI, 45 (9):11221-11239, 2023. 6",
+            "[18] Lianyu Hu, Liqing Gao, Zekang Liu, and Wei Feng. Dynamic spatial-temporal aggregation for skeleton-aware sign language recognition. In LREC/COLING, 2024. 6, 7",
+            "[19] Jie Huang, Wengang Zhou, Houqiang Li, and Weiping Li. Attention-based 3d-cnns for large-vocabulary sign language recognition. IEEE TCSVT, 29(9):2822-2832, 2018. 2",
+            "[20] Jie Huang, Wengang Zhou, Qilin Zhang, Houqiang Li, and Weiping Li. Video-based sign language recognition without temporal segmentation. In AAAI, 2018. 2",
+            "[21] Songyao Jiang, Bin Sun, Lichen Wang, Yue Bai, Kunpeng Li, and Yun Fu. Skeleton aware multi-modal sign language recognition. In CVPR, pages 3413-3423, 2021. 1, 3, 4, 5, 8",
+            "[22] Songyao Jiang, Bin Sun, Lichen Wang, Yue Bai, Kunpeng Li, and Yun Fu. Sign language recognition via skeleton-aware multi-model ensemble. arXiv preprint arXiv:2110.06161, 2021. 6, 7",
+            "[23] Tao Jiang, Necati Cihan Camgoz, and Richard Bowden. Skeleton: Skeletal transformers for robust body-posed estimation. In CVPR, pages 3394–3402, 2021. 1, 3",
+            "[24] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In ECCV, pages 196-214. Springer, 2020. 3, 4",
+            "[25] Trevor Johnston and Adam Schembri. Australian Sign Language (Auslan): An introduction to sign language linguistics. Cambridge University Press, 2007. 1",
+            "[26] Hamid Reza Vaezi Joze and Oscar Koller. Ms-asl: A large-scale data set and benchmark for understanding american sign language. arXiv preprint arXiv:1812.01053, 2018. 2, 7",
+            "[27] Alexander Kapitanov, Kvanchiani Karina, Alexander Nagaev, and Petrova Elizaveta. Slovo: Russian sign language dataset. In International Conference on Computer Vision Systems, pages 63-73. Springer, 2023. 2, 7, 8"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "24328",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[28] Oscar Koller. Quantitative survey of the state of the art in sign language recognition. arXiv preprint arXiv:2008.09918, 2020. 2",
+            "[29] Oscar Koller, Sepehr Zargaran, Hermann Ney, and Richard Bowden. Deep sign: Enabling robust statistical continuous sign language recognition via hybrid cnn-hmms. *IJCV*, 126: 1311–1325, 2018. 2",
+            "[30] Taeryung Lee, Yeonguk Oh, and Kyoung Mu Lee. Human part-wise 3d motion context learning for sign language recognition. In ICCV, pages 20740-20750, 2023. 1, 6",
+            "[31] Chao Li, Qiaoyong Zhong, Di Xie, and Shiliang Pu. Cooccurrence feature learning from skeleton data for action recognition and detection with hierarchical aggregation. arXiv preprint arXiv:1804.06055, 2018. 1, 3",
+            "[32] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1459-1469, 2020. 2, 7, 8",
+            "[33] Dongxu Li, Xin Yu, Chenchen Xu, Lars Petersson, and Hongdong Li. Transferring cross-domain knowledge for video sign language recognition. In CVPR, pages 6205-6214, 2020. 6",
+            "[34] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In CVPR, pages 3595–3603, 2019. 3",
+            "[35] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mviv2: Improved multiscale vision transformers for classification and detection. In CVPR, pages 4804-4814, 2022. 7, 8",
+            "[36] Jiajia Liao, Yujun Liu, Yingchao Piao, Jinhe Su, Guorong Cai, and Yundong Wu. Gle-net: A global and local ensemble network for aerial object detection. International Journal of Computational Intelligence Systems, 15(1):2, 2022. 7",
+            "[37] Jinfu Liu, Xinshun Wang, Can Wang, Yuan Gao, and Mengyuan Liu. Temporal decoupling graph convolutional network for skeleton-based gesture recognition. IEEE TMM, 26:811-823, 2023. 3",
+            "[38] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, pages 3202-3211, 2022. 3, 7, 8",
+            "[39] Anna Mindess. Reading between the signs: Intercultural communication for sign language interpreters. Nicholas Brealey, 2014. 1",
+            "[40] Liliane Momeni, Hannah Bull, KR Prajwal, Samuel Albanie, Gül Varol, and Andrew Zisserman. Automatic dense annotation of large-vocabulary sign language videos. In ECCV, pages 671-690. Springer, 2022. 2",
+            "[41] Evonne Ng, Shiry Ginosar, Trevor Darrell, and Hanbyul Joo. Body2hands: Learning to infer 3d hands from conversational gesture body dynamics. In CVPR, pages 11865-11874, 2021. 1, 3",
+            "[42] Ladislav Rampášek, Michael Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. Recipe for a general, powerful, scalable graph transformer. NeurIPS, 35:14501-14515, 2022. 3"
+        ],
+        "bbox": [
+            91,
+            90,
+            480,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[43] Razieh Rastgoo, Kourosh Kiani, and Sergio Escalera. Sign language recognition: A deep survey. Expert Systems with Applications, 164:113794, 2021. 2",
+            "[44] Wendy Sandler. Sign language and linguistic universals. Cambridge University, 2006. 1",
+            "[45] Noha Sarhan and Simone Frintrop. Unraveling a decade: a comprehensive survey on isolated sign language recognition. In ICCV, pages 3210-3219, 2023. 2",
+            "[46] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In CVPR, pages 12026-12035, 2019. 3",
+            "[47] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with multi-stream adaptive graph convolutional networks. IEEE TIP, 29:9532-9545, 2020. 2",
+            "[48] Chenyang Si, Wentao Chen, Wei Wang, Liang Wang, and Tieniu Tan. An attention enhanced graph convolutional LSTM network for skeleton-based action recognition. In CVPR, pages 1227-1236, 2019. 3",
+            "[49] Neil Song and Yu Xiang. Slgtformer: An attention-based approach to sign language recognition. arXiv preprint arXiv:2212.10746, 2022. 3, 4",
+            "[50] Yi-Fan Song, Zhang Zhang, Caifeng Shan, and Liang Wang. Stronger, faster and more explainable: A graph convolutional baseline for skeleton-based action recognition. In proceedings of the 28th ACM international conference on multimedia, pages 1625-1633, 2020. 2",
+            "[51] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, pages 5693-5703, 2019. 3, 4",
+            "[52] Clayton Valli. Linguistics of American Sign Language: An Introduction. Gallaudet University Press, 2000. 1",
+            "[53] Gül Varol, Liliane Momeni, Samuel Albanie, Triantafyllos Afouras, and Andrew Zisserman. Scaling up sign spotting through sign language dictionaries. *IJCV*, 130(6):1416–1439, 2022. 2",
+            "[54] Hongsong Wang and Liang Wang. Modeling temporal dynamics and spatial configurations of actions using two-stream recurrent neural networks. In CVPR, pages 499-508, 2017. 2",
+            "[55] Hanjie Wang, Xiujuan Chai, and Xin Chen. A novel sign language recognition framework using hierarchical grassmann covariance matrix. IEEE TMM, 21(11):2806-2814, 2019. 2",
+            "[56] Chengcheng Wei, Jian Zhao, Wengang Zhou, and Houqiang Li. Semantic boundary detection with reinforcement learning for continuous sign language recognition. IEEE TCSVT, 31(3):1138-1149, 2020.",
+            "[57] Pan Xie, Mengyi Zhao, and Xiaohui Hu. Pisltrc: Position-informed sign language transformer with content-aware convolution. IEEE TMM, 24:3908-3919, 2021. 2",
+            "[58] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 1, 6, 7",
+            "[59] Aoxiong Yin, Zhou Zhao, Weike Jin, Meng Zhang, Xingshan Zeng, and Xiaofei He. Mlslt: Towards multilingual sign language translation. In CVPR, pages 5109-5119, 2022. 1"
+        ],
+        "bbox": [
+            514,
+            90,
+            903,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "24329",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[60] Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. Best: Bert pre-training for sign language recognition with coupling tokenization. In AAAI, pages 3597-3605, 2023. 6, 7",
+            "[61] Weichao Zhao, Hezhen Hu, Wengang Zhou, Yunyao Mao, Min Wang, and Houjiang Li. Masa: Motion-aware masked autoencoder with semantic alignment for sign language recognition. IEEE TCSVT, 2024. 6, 7",
+            "[62] Weichao Zhao, Wengang Zhou, Hezhen Hu, Min Wang, and Houqiang Li. Self-supervised representation learning with spatial-temporal consistency for sign language recognition. arXiv preprint arXiv:2406.10501, 2024. 6, 7",
+            "[63] Ronglai Zuo, Fangyun Wei, and Brian Mak. Natural language-assisted sign language recognition. In CVPR, pages 14890-14900, 2023. 6, 7"
+        ],
+        "bbox": [
+            91,
+            90,
+            480,
+            303
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "24330",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2025/VSNet_ Focusing on the Linguistic Characteristics of Sign Language/1135c1a4-3af0-45ac-a497-2c9557104d8c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25ea6623303544c6406f405e9b1bdbebd55a0a5bf1bd89647bb6cfea9ddc822c
+size 3520878

2025/VSNet_ Focusing on the Linguistic Characteristics of Sign Language/full.md

@@ -0,0 +1,331 @@
+# VSNet: Focusing on the Linguistic Characteristics of Sign Language
+
+Yuhao Li $^{1}$ , Xinyue Chen $^{1}$ , Hongkai Li $^{1}$ , Xiaorong Pu $^{1,3}$ , Peng Jin $^{2}$ , Yazhou Ren $^{1,3,*}$ $^{1}$ School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, China;  $^{2}$ Sichuan Provincial Key Lab of Philosophy and Social Science for Language Intelligence in Special Education, Leshan Normal University, Leshan, China;  $^{3}$ Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China, Shenzhen, China
+
+# Abstract
+
+Sign language is a visual language expressed through complex movements of the upper body. The human skeleton plays a critical role in sign language recognition due to its good separation from the video background. However, mainstream skeleton-based sign language recognition models often overly focus on the natural connections between joints, treating sign language as ordinary human movements, which neglects its linguistic characteristics. We believe that just as letters form words, each sign language gloss can also be decomposed into smaller visual symbols. To fully harness the potential of skeleton data, this paper proposes a novel joint fusion strategy and a visual symbol attention model. Specifically, we first input the complete set of skeletal joints, and after dynamically exchanging joint information, we discard the parts with the weakest connections to other joints, resulting in a fused, simplified skeleton. Then, we group the joints most likely to express the same visual symbol and discuss the joint movements within each group separately. To validate the superiority of our method, we conduct extensive experiments on multiple public benchmark datasets. The results show that, without complex pretraining, we still achieve new state-of-the-art performance. The code is available at https://github.com/atinyboy/VSNet.
+
+# 1. Introduction
+
+Sign language is a visual language expressed through body movements and serves as the primary communication tool for the deaf community [13]. Although sign language and spoken language have distinctly different grammatical structures [1, 7], they share common linguistic features [39, 44, 59]. Accurate sign language translation can facilitate communication between deaf and hearing people, fostering a more inclusive and accessible society. Among
+
+![](images/0ef7164ed9cc0e33de67ef10e8993b7a4a21ae4ff844e5a166b44fb3d111af05.jpg)
+
+![](images/44fe15157fcbe074d6c1d7cd80d06ce1f7f18531872187c473e9164ff0c00f43.jpg)  
+gloss: 'accent'
+
+![](images/f9bf732e363bf30348c2d295f318f5d05c852e71aa7b797239453d99c7a09b96.jpg)
+
+![](images/ca5a47bc0d09f7265a256a59d8607956570ca3711d2ec5e38228d7f2796734ad.jpg)
+
+![](images/fdeac63056da7ab56e05b49bb7c6ccfa708b60992d811714809658431cc4ca36.jpg)  
+Figure 1. In the sign for "accent", the index and middle fingers (or only index finger) are extended toward the neck while other fingers are curled, forming the "pointing" VS. In the sign for "algebra", movements can be decomposed into multiple VS such as "crossed arms," "fist," and "thumb-up."
+
+![](images/18fa9e3dbae1a5067929bcacce7062c5be4971a7cddd1ad711e479f11e05de4b.jpg)  
+gloss:'algebra'
+
+![](images/ad3e18e00f9ec6d67c41252c0742477cc98782276705a52e360e368fa82d5f19.jpg)
+
+![](images/10785b8de7ae97caba6d78a7d9b114606070865a8f7c70ff8eea76801732b85a.jpg)
+
+the key subtasks in sign language translation is isolated sign language recognition (ISLR), which focuses on translating individual glosses. ISLR has garnered significant attention as it aids people in accurately understanding numerous gestures and quickly learning sign language.
+
+In recent years, human skeletons have been increasingly applied in the ISLR field due to their strong operability and clear separation from video backgrounds. Although most skeleton-based recognition methods have achieved promising results, they typically treat sign language as typical human motion [12, 21, 23, 30, 31, 41, 58], overlooking the intrinsic relationships among joints and the unique linguistic characteristics [25, 52] of sign language itself. We believe that: (1) The importance of joints in sign language is not equivalent; joints with stronger connections to others are more significant than isolated joints, and connections among joints with similar movement tendencies hold more meaning than those among dispersed joints. (2) Similar to how letters form words, each sign language gloss can be decomposed into smaller visual symbols (VS). While VS
+
+lack explicit meaning on their own, from a skeletal perspective, they capture distinct movement patterns of joints, effectively decomposing sign language actions (see Fig. 1). Therefore, our work focuses on these two core aspects:
+
+Core 1: Focus on identifying essential joints. Including too many joints can impede training, while too few can lead to information loss. On the one hand, many prior models have used manually trimmed, simplified skeletons, but manually selecting the optimal joints is not trivial. On the other hand, enabling a model to automatically choose key joints can be computationally intensive. Fortunately, the dynamically learned skeletons from GCNs approximate real structures more closely, enabling us to reduce the dependency on an initial skeleton. By computing the weight of each joint, we can identify and retain the most important joints while discarding less critical local joints. Specifically, for the complete hand skeleton, we discard a small number of joints based on computed weights after each GCN layer, relearning new skeleton connections in the subsequent GCN layer. We control the number of joints discarded within each region to ensure that the final GCN output retains a recognizable structure. Notably, even joints with the lowest weights undergo at least one convolution, so their information is not entirely discarded.
+
+Core 2: Focus on emphasizing visual symbols (VS) in sign language. If each joint is considered the smallest element of a sign language action, then the combination of joints across all frames becomes too vast to compute self-attention. In practice, sign language is not random motion; it is designed for human recognition. When a signer interprets sign language, they do not scrutinize the position of every joint but instead form a general impression of each hand. Thus, in this work, we concentrate on recognizing VS. Unfortunately, isolating each VS from movements is as challenging as the entire SLR task itself. Therefore, we adopt a simplified approach, segmenting the skeleton into a generalized VS in a fixed manner, as shown in Fig. 2. We then design two self-attention modules to capture spatial and temporal relationships between these VS. Empirically, we find that this approach significantly enhances the model's discriminative power.
+
+Our main contributions are summarized as follows:
+
+- We propose a novel module with an automated skeleton simplification function for sign language recognition. This module is designed to dynamically learn global skeleton connections while progressively discarding isolated local joints.  
+- Considering the linguistic characteristics of sign language, we introduce a new self-attention model called VSformer, which captures temporal and spatial patterns in sign language motions and extracts knowledge from the human skeleton based on the concept of VS.  
+- We conduct extensive experiments on ISLR to validate
+
+![](images/643f5b9a8f8824d4be18635ddea1b3f7e73bb10c6cc91a7c3c93649d5a9787aa.jpg)  
+Figure 2. Decomposing generalized visual symbols from videos
+
+the effectiveness of our proposed approach. Compared to previous methods, our approach achieves SoTA performance on four public benchmarks, namely WLASL [32], MSASL [26], NMFs-CSR [16], and Slovo [27].
+
+# 2. Related Works
+
+# 2.1. Sign Language Recognition
+
+Sign language is the primary means of communication within the deaf community. With significant advances in machine learning and computer vision over the past decade, exploring sign language recognition (SLR) has become increasingly important [28, 43, 45], as it can automatically interpret sign language and help deaf and mute individuals communicate more easily in daily life. Unlike directly translating sign language sentences, isolated sign language recognition (ISLR) is a subtask in the field of sign language translation [45], aimed at translating each individual gloss [40, 53]. ISLR is not only useful for guiding beginners in learning sign language and assisting in the creation of new sign language gestures, but it also lays a solid foundation for SLR [1, 4, 53]. Research in this area can be divided into two categories based on the input modality: RGB-based methods and skeleton-based methods.
+
+RGB-based Recognition. RGB-based methods, similar to other video recognition tasks, typically use CNNs to extract visual features from RGB videos [5, 15, 19, 20, 29, 55-57], and then aggregate temporal and spatial information to predict sign language. However, RGB-based models face two significant challenges: first, they are heavily affected by changes in the video background, and second, video data occupies a large amount of memory resources.
+
+Skeleton-based Recognition. Compared to RGB images, human skeletons [8, 47, 50, 54] describe the motion information conveyed in sign language videos using simplified points and edges, reducing interference from complex backgrounds and the appearance of the signer. Moreover, operations on three-dimensional spatial points with actual physical meanings offer stronger interpretability and
+
+![](images/a0fe65ed0df5a97890f903093b93c51daf44c0b86f7de44230e2f011d4ddd927.jpg)  
+Figure 3. Overview of our VSNet. Here, Down sample consists of a convolution layer with a kernel size of 7 and a stride of 2 along the temporal dimension, followed by a batch norm layer, Average Pooling represent transforming the data into a single-dimensional feature by taking the mean across both temporal and spatial dimensions. For details on the Skeleton- $(a,b,c)$ , please refer to Fig. 4.
+
+lower memory consumption. In earlier works, identifying the correct human keypoints in a video was a major challenge. However, with the development of deep learning, human pose estimation models have become increasingly reliable [10, 24, 51].
+
+# 2.2. Human Skeleton
+
+Due to the natural physical connections between skeleton joints, Graph Convolutional Network (GCN) has been widely applied to model spatio-temporal relationships [12, 23, 31, 34, 41, 46]. GCN require predefined connectivity of the key points to construct the graph network. However, the physically natural skeleton connections may not always be the most suitable edges for recognition. Relying solely on the skeleton itself to construct the graph can limit the information flow between nodes. To address this issue, SL-GCN [21] employs a dynamic graph with decoupling between different frames, while TD-GCN [37] adds supernodes to control the connections between nodes. A common drawback of these approaches is that they require complex networks to model the key points. Furthermore, while dynamic graphs alleviate the limitations of the original skeleton, blindly allowing all key points to connect introduces redundant information, leading to suboptimal models.
+
+Therefore, we argue that relying solely on GCN does not fully exploit the true potential of skeleton data.
+
+In recent years, attention-based methods [3, 6, 9, 38, 42, 48] have also made progress in human skeleton recognition. Unlike GCN, attention mechanisms do not require the researcher to provide information on joint connections, instead equally attending to all nodes. However, captur
+
+ing the correlations between all joints across all frames requires substantial memory resources. To address this, SLGTformer [49] decomposes the spatio-temporal pose sequence into spatial graphs and temporal windows, while SkateFormer [11] partitions joints and frames and performs self-attention on each partition. Based on these theories, we design a model where GCNs and self-attention networks complement each other. GCNs are responsible for exploring the spatial correlations between human joints, while self-attention networks focus on capturing the motion of the skeleton across multiple time sequences and spatial combinations.
+
+# 3. Methods
+
+# 3.1. Data Pre-Processing
+
+Key Point Extraction. Sign language is a visual language primarily conveyed through hand and body movements, with facial expressions also playing a crucial role. However, the human skeleton mainly describes bones and joints, excluding muscles, making it unable to accurately capture facial expressions in sign language videos. Our investigation revealed that a significant number of signers, especially beginners, do not exhibit extensive facial expressions during sign language communication. Therefore, our main experiments exclude facial expressions, while the appendix provides additional details on how the concept of VS can be extended to the face.
+
+In previous GCN models, finger joints were often simplified to avoid making the graph too complex, leading to a loss of information. In this study, we selected a complete
+
+![](images/3b302e9367d27a69b446d6fd883fe21af9ef8d9b70365c49823d0a11074b393e.jpg)  
+Figure 4. Illustration of the skeleton and grouping type-1.  $V_{0}$  represents a localization node, the dashed box indicates nodes that are grouped together and hollow dots indicate dropped nodes. Please note that (a) is the input real skeleton diagram, while (b) and (c) are simplified representations drawn after dropping for ease of analysis.
+
+![](images/b715aea289b6efb25c9f8e0f180402f2a4f181c1c5aefe14eb1470b89cdc6b34.jpg)  
+Figure 5. Skeleton groupings type-2 and type-3. The other hand is grouped in a mirrored fashion.
+
+set of 44 key points from both hands out of the 133 human body joints extracted using HRNet [24, 51], as shown in Fig. 4(a).
+
+Data Enhancing. Given a video  $X \in \mathbb{R}^{T \times V \times C}$ , where T, V, and C represent the number of frames, joints, and coordinates respectively. We replace abnormal points—such as those out of range (less than 0, more than 512) or clearly distorted—caused by model extraction errors, with the average value from adjacent frames. Additionally, to standardize body positions and movement amplitudes across different signers in the sign language videos, we first use the hollow node  $V_{0}$  in Fig. 4 as the reference node for the entire video. Then, all key points in each frame are shifted based on this reference coordinate, as follows:
+
+$$
+X (t, v, c) = X (t, v, c) - X (t, v _ {0}, c) + [ x, y, z ] \tag {1}
+$$
+
+where  $[x,y,z]$  is the average value of all the coordinates of  $V_{0}$  in the dataset.
+
+In the time dimension, we randomly select 64 frames to standardize the video length and perform random mirroring and shifting of the video to generate augmented data.
+
+# 3.2. Self-Pacing Dropping Block (SPD)
+
+As shown in Fig. 3, our SPD consists of two components: "Keypoint Activation" and "Weak Joints Dropping". The input is a complete hand skeleton where the node information is independent, and the output is a high-dimensional feature skeleton after activation. The entire block is stacked three times, where the third block only outputs features and does not build a new skeleton.
+
+Keypoint Activation Layer. In this part, we define the interaction of each joint with others as an activation. We employ a decoupled graph convolution [21] to perform this activation, represented as:
+
+$$
+F _ {o u t} ^ {t} = \frac {1}{p} \sum_ {k = 0} ^ {p} D ^ {- \frac {1}{2}} A _ {k} D ^ {- \frac {1}{2}} \phi \left(F _ {i n} ^ {t}\right) \tag {2}
+$$
+
+where  $A$  is the learnable adjacency matrix initialized as  $I + In + Out$ ,  $D$  presents the diagonal degree of  $A$ . Here,  $I$  denotes self-connections, while  $In$  and  $Out$  represent inward and outward connections, respectively. The symbol  $\phi$  represents a linear layer, and  $p$  represents the decoupling quantity, with a different  $A$  applied each time, followed by averaging the results. Following the graph convolution, we apply the simplest Temporal Convolutional Network (TCN) to capture features along the temporal dimension.
+
+Weak Joints Dropping Layer. Compared to the simplified skeletons used in [21, 49], our approach utilizes a complete set of skeleton keypoints that contains richer information. However, this comes at the cost of significantly increasing the model's burden and raising the risk of introducing outlier points. To leverage these additional points with minimal overhead, we have designed an adaptive skeleton simplification strategy.
+
+Before formally introducing this layer, it is necessary to present our approach to skeleton decomposition. Based on the definition of VS described in Introduction, the decomposition concept involves grouping joints that exhibit strong overall correlation or frequently form VS together in sign language. These groupings can be viewed as a generalized or macro-level VS. A simple approach is grouping each finger and arm separately, but this creates too many groups, potentially affecting later interactions. Based on this, we conducted various experiments on VS, including treating individual fingers and distant fingers as VS. However, these experiments did not yield optimal results (the average performance in WLASL dataset decreased by  $1.20\%$  and  $1.64\%$ , respectively). We believe this is because the information from a single finger is insufficient to form a meaningful VS, while the information disparity between distant fingers is even greater. This also serves as the skeleton grouping basis for Figs. 4 and 5.
+
+After graph convolution in Eq. 2,  $A(i,j)$  has learned the weighted edge between node  $i$  and node  $j$ . These edges reflect the connection strength between  $i$  and  $j$ . Since these
+
+connections are not normalized via softmax, we can use  $A$  as a guide to calculate the total correlation of all keypoints, Cor, as follows:
+
+$$
+C o r _ {i} = \sum_ {j} \max  \left(A _ {0} (i, j), A _ {1}, \dots , A _ {p}\right) \tag {3}
+$$
+
+A lower  $Cor$  to other keypoints indicates that the point is more independent. We refer to the two most independent points in each group as weak joints, which are dropped from the skeleton. To ensure that the number of keypoints in each group remains consistent, the groups involving both arms will have two fewer joints dropped.
+
+# 3.3.VSformer
+
+After the joint fusion block, we obtained six different joint groupings, as shown in Fig. 4(c). Based on the concept of skeletal disassembly, we treat each grouping as a generalized VS. We now need to focus on two main questions: Q1, what are the relationships between different VS? And Q2, do these VS remain consistent throughout the entire video, or do they change after a brief period?
+
+Inspired by [11], which proposes an efficient model for partitioning human limbs in action recognition, we employ two types of self-attention to constrain the model's focus on the aforementioned questions. Specifically, we calculate spatial pairing attention among different combinations within a frame to address Q1, and we compute temporal window attention for both global and local motion of the combinations across different frames to address Q2.
+
+Spatial Pairing Attention (SPA). Given a frame  $F_{t} \in \mathbb{R}^{G \times C}$ , where  $G = 6$  represents VS groupings, the ideal approach would involve calculating attention among all VS pairwise. However, this is inefficient and often purposeless, as VS that are far apart in the skeletal structure tend to have weak correlations. For instance, in the sign language gesture for "accent", it is unnecessary to compare the fingers of the left hand with those of the unused right hand. Based on this observation, we ultimately partition  $F_{t}$  into pairs as follows:
+
+$$
+F _ {t} = \left[ P _ {\text {a r m}}, P _ {\text {l e f t}}, P _ {\text {r i g h t}} \right] \tag {4}
+$$
+
+where  $P_{\mathrm{arm}} \in \mathbb{R}^{2 \times C}$  represents the grouping of both arms,  $P_{\mathrm{left}}$  denotes the four fingers of the left hand, and  $P_{\mathrm{right}}$  represents the remaining four fingers of the right hand. Represented as:
+
+$$
+S P A _ {\text {o u t}} = \operatorname {c o n c a t} \left(\operatorname {s o f t m a x} \left(\frac {Q _ {i} K _ {i} ^ {T}}{\sqrt {h}}\right) V _ {i}\right) \tag {5}
+$$
+
+where  $Q_{i}, K_{i}, V_{i}$  are obtained from  $P_{i}$  after mapping through linear layers. The attention is actually calculated among the 8 joints within each pairing and is ultimately concatenated to form the output.
+
+Temporal Window Attention (TWA). Depending on different sign language videos, the VS can be classified into
+
+![](images/621111fb7906cca097896dddac1e940b4f52be4a0b9723b1c4add3959b904881.jpg)  
+Figure 6. TWA windows. The small ticks represent frames, and the large ticks represent windows.
+
+two types: short duration and long duration, representing their respective duration periods. Accordingly, our TWA is also divided into two types: TWA-s and TWA-l. Given a video  $X \in \mathbb{R}^{T \times G \times C}$ , where  $T$  represents the total number of frames, the timeline is divided into  $n$  windows, each containing  $k$  frames, such that  $T = n \times k$ , as shown in Fig. 6. For TWA-s, attention is computed within each window. For TWA-l, frames with the same index are selected sequentially from each window. Finally, the results from both types of attention are concatenated to produce the output, represented as:
+
+$$
+T W A _ {o u t} = \operatorname {s o f t m a x} \left(\phi_ {1} (W) \phi_ {2} ^ {T} (W) + b\right) \phi_ {3} (W) \tag {6}
+$$
+
+where  $W$  represents the selected windows,  $\phi_{n}$  denotes dividing the  $C$  dimension into three equal parts and selecting the  $n$ -th part, and  $b$  represents the added bias.
+
+It is worth noting that the discussions on Q1 and Q2 are not fully addressed by these two attention layers. The SPA only addresses the relationships between paired VS, while the windows defined in TWA do not entirely correspond to the duration of the VS. Therefore, we incorporate a simple GCN and TCN to enhance the model's ability to address the remaining scenarios. The connection method of SPA and TWA is shown in the right side of Fig. 3. After passing the input data  $X \in \mathbb{R}^{T \times G \times C}$  through a linear layer, we divide its channel dimension  $C$  into four equal parts, sending each part into one of the four modules and later concatenated. Additionally, unlike the Decoupled GCN in Sect. 3.2, the GCN here utilizes an adjacency matrix shared across certain frames, which significantly reduces the computational load.
+
+# 3.4. Multi-Grouping Ensemble
+
+In the previous sections, we introduced the concept of skeletal disassembly based on VS. However, since we cannot precisely determine the true VS composition for each sign language, using a fixed disassembly method appears to be a compromise, potentially weakening interpretability. Therefore, we adopt an ensemble strategy to reduce such discrepancies. Specifically, we designed two additional skeleton grouping schemes, as shown in the Fig. 5, and trained the model separately to obtain the output  $q$  from the final fully connected layer for each scheme. Following [21], we aggregate these results using a weighted sum, represented as:  $q_{\mathrm{total}} = \alpha_1q_{\mathrm{type - 1}} + \alpha_2q_{\mathrm{type - 2}} + \alpha_3q_{\mathrm{type - 3}} + \alpha_4q_{\mathrm{bone}}$ , where  $\alpha$  represents the weights. For a detailed discussion on  $\alpha$ ,
+
+<table><tr><td rowspan="3">Methods</td><td colspan="4">MSASL1000</td><td colspan="4">MSASL200</td><td colspan="4">MSASL100</td></tr><tr><td colspan="2">P-I</td><td colspan="2">P-C</td><td colspan="2">P-I</td><td colspan="2">P-C</td><td colspan="2">P-I</td><td colspan="2">P-C</td></tr><tr><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td></tr><tr><td>RGB-based</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>HMA [15]</td><td>49.16</td><td>69.75</td><td>46.27</td><td>68.60</td><td>66.30</td><td>84.03</td><td>67.47</td><td>84.03</td><td>73.45</td><td>89.70</td><td>74.59</td><td>89.70</td></tr><tr><td>I3D [5]*</td><td>-</td><td>-</td><td>57.69</td><td>81.05</td><td>-</td><td>-</td><td>81.97</td><td>93.79</td><td>-</td><td>-</td><td>81.76</td><td>95.16</td></tr><tr><td>BSL [2]</td><td>64.71</td><td>85.59</td><td>61.55</td><td>84.43</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>TCK [33]†</td><td>-</td><td>-</td><td>-</td><td>-</td><td>80.31</td><td>91.82</td><td>81.14</td><td>92.24</td><td>83.04</td><td>93.46</td><td>83.91</td><td>93.52</td></tr><tr><td>skeleton-based</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>ST-GCN [58]</td><td>36.03</td><td>59.92</td><td>32.32</td><td>57.15</td><td>52.91</td><td>76.67</td><td>54.20</td><td>77.62</td><td>59.84</td><td>82.03</td><td>60.79</td><td>82.96</td></tr><tr><td>BEST [60]</td><td>58.82</td><td>81.18</td><td>54.87</td><td>80.05</td><td>76.60</td><td>91.54</td><td>76.75</td><td>91.95</td><td>80.98</td><td>95.11</td><td>81.24</td><td>95.44</td></tr><tr><td>SignBERT [14]†</td><td>59.80</td><td>81.86</td><td>57.06</td><td>80.94</td><td>77.34</td><td>91.10</td><td>78.02</td><td>91.48</td><td>81.37</td><td>93.66</td><td>82.31</td><td>93.76</td></tr><tr><td>SignBERT+ [17]†</td><td>62.42</td><td>83.49</td><td>60.15</td><td>82.44</td><td>78.51</td><td>-</td><td>79.35</td><td>-</td><td>84.94</td><td>-</td><td>85.23</td><td>-</td></tr><tr><td>MASA [61]†</td><td>63.47</td><td>83.89</td><td>60.79</td><td>83.29</td><td>79.25</td><td>92.86</td><td>79.70</td><td>93.33</td><td>83.22</td><td>95.24</td><td>83.19</td><td>95.46</td></tr><tr><td>SSRL [62]†</td><td>65.22</td><td>85.09</td><td>62.68</td><td>84.38</td><td>82.93</td><td>94.78</td><td>83.43</td><td>95.03</td><td>86.26</td><td>96.96</td><td>86.63</td><td>96.79</td></tr><tr><td>Ours</td><td>70.71</td><td>86.36</td><td>68.38</td><td>85.71</td><td>84.62</td><td>94.04</td><td>85.19</td><td>94.08</td><td>86.26</td><td>95.24</td><td>86.68</td><td>95.39</td></tr><tr><td>Ours(4-crops)</td><td>73.18</td><td>88.64</td><td>70.88</td><td>87.59</td><td>86.17</td><td>95.51</td><td>86.64</td><td>95.61</td><td>88.11</td><td>97.09</td><td>88.32</td><td>97.18</td></tr></table>
+
+Table 1. Comparison with previous work on MSASL. “†” indicates that the model was pre-trained, and “*” indicates methods that applied result fusion with other modalities in post-processing. Bold indicates the best performance, while underlined denotes the second-best.  
+
+<table><tr><td rowspan="3">Methods</td><td colspan="4">WLASL2000</td><td colspan="4">WLASL300</td><td colspan="4">WLASL100</td></tr><tr><td colspan="2">P-I</td><td colspan="2">P-C</td><td colspan="2">P-I</td><td colspan="2">P-C</td><td colspan="2">P-I</td><td colspan="2">P-C</td></tr><tr><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td></tr><tr><td>RGB-based</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>I3D [5]*</td><td>32.48</td><td>57.31</td><td>-</td><td>-</td><td>56.14</td><td>79.94</td><td>56.24</td><td>78.38</td><td>65.89</td><td>84.11</td><td>67.01</td><td>84.58</td></tr><tr><td>HMA [15]</td><td>37.91</td><td>71.26</td><td>35.90</td><td>70.00</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>BSL [2]</td><td>46.82</td><td>79.36</td><td>44.72</td><td>78.47</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>TCK [33]†</td><td>-</td><td>-</td><td>-</td><td>-</td><td>68.56</td><td>89.52</td><td>68.75</td><td>89.41</td><td>77.52</td><td>91.08</td><td>77.55</td><td>91.42</td></tr><tr><td>NLA-SLR V64 [63]*</td><td>51.15</td><td>83.43</td><td>48.14</td><td>82.20</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>skeleton-based</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>P3D [30]</td><td>44.47</td><td>79.69</td><td>42.18</td><td>78.52</td><td>67.18</td><td>89.01</td><td>67.62</td><td>89.24</td><td>76.71</td><td>91.97</td><td>78.27</td><td>92.97</td></tr><tr><td>BEST [60]</td><td>46.25</td><td>79.33</td><td>43.52</td><td>77.65</td><td>67.66</td><td>89.22</td><td>68.31</td><td>89.57</td><td>77.91</td><td>91.47</td><td>77.83</td><td>92.50</td></tr><tr><td>SignBERT [14]†</td><td>47.46</td><td>83.32</td><td>45.17</td><td>82.32</td><td>70.36</td><td>88.92</td><td>71.17</td><td>89.36</td><td>79.07</td><td>93.80</td><td>80.05</td><td>94.17</td></tr><tr><td>MASA [61]†</td><td>49.06</td><td>82.90</td><td>46.91</td><td>81.80</td><td>73.65</td><td>91.77</td><td>74.33</td><td>92.13</td><td>83.72</td><td>93.80</td><td>84.47</td><td>94.30</td></tr><tr><td>NLA-SLR K64 [63]*</td><td>49.10</td><td>82.00</td><td>46.18</td><td>80.71</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SL-GCN [22]*</td><td>51.50</td><td>84.94</td><td>48.87</td><td>84.02</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DSTA [18]</td><td>53.68</td><td>-</td><td>51.17</td><td>-</td><td>79.97</td><td>-</td><td>80.56</td><td>-</td><td>82.38</td><td>-</td><td>83.09</td><td>-</td></tr><tr><td>Ours</td><td>53.54</td><td>83.18</td><td>51.18</td><td>82.03</td><td>79.04</td><td>94.01</td><td>79.74</td><td>94.25</td><td>84.50</td><td>94.57</td><td>85.25</td><td>94.33</td></tr><tr><td>Ours(4-crops)</td><td>55.98</td><td>87.07</td><td>53.54</td><td>86.04</td><td>80.09</td><td>94.76</td><td>80.85</td><td>94.92</td><td>85.66</td><td>94.96</td><td>86.25</td><td>95.08</td></tr></table>
+
+Table 2. Comparison with previous work on WLASL. “†” indicates that the model was pre-trained, and “*” indicates methods that applied result fusion with other modalities in post-processing. Bold indicates the best performance, while underlined denotes the second-best.
+
+please refer to Sect. 4.4. The term bone refers to the joint data of type-1 represented in vector form, generated by directing from the source joint to its target joint according to the natural connections of the human body.
+
+# 4. Experiments
+
+# 4.1. Parameter Settings
+
+We employ the AdamW optimizer with a plateau learning rate schedule. The initial learning rate is set to  $1 \times 10^{-3}$ , with a minimum learning rate of  $1 \times 10^{-5}$  and a decay rate
+
+of 0.5. The batch size is 32 and the loss function used is cross-entropy with label smoothing. Training spans a maximum of 200 epochs, with the first 15 epochs serving as a warm-up period starting from a learning rate of  $1 \times 10^{-7}$ . A random sample of 64 frames is selected as input, and with a  $50\%$  probability, image mirroring is applied. A random jitter in the range  $(-10, 10)$  is added to each joint point. In the last two layers of SPD, joints are randomly masked with a probability of  $(1 - p_{kr}) / 100 \times \text{epoch}$ , where  $p_{kr}$  is set to 0.9. All experiments were performed with NVIDIA A100.
+
+<table><tr><td rowspan="2">Methods</td><td colspan="2">NMFs-CSR</td></tr><tr><td>Top-1</td><td>Top-5</td></tr><tr><td>RGB-based</td><td></td><td></td></tr><tr><td>I3D [5]*</td><td>64.4</td><td>88.0</td></tr><tr><td>HMA [15]</td><td>64.7</td><td>91.0</td></tr><tr><td>GLE-Net [36]</td><td>69.0</td><td>88.1</td></tr><tr><td>skeleton-based</td><td></td><td></td></tr><tr><td>ST-GCN [58]</td><td>59.9</td><td>86.8</td></tr><tr><td>BEST [60]</td><td>68.5</td><td>94.4</td></tr><tr><td>MASA [61]†</td><td>71.7</td><td>97.0</td></tr><tr><td>SignBERT [14]†</td><td>74.9</td><td>93.2</td></tr><tr><td>Ours</td><td>75.3</td><td>95.4</td></tr><tr><td>Ours(4-crops)</td><td>76.7</td><td>97.2</td></tr></table>
+
+Table 3. Comparison on NMFs-CSR. Bold indicates the best performance, while underlined denotes the second-best.  
+
+<table><tr><td rowspan="2">Methods</td><td colspan="2">Slovo</td></tr><tr><td>Top-1</td><td>Top-5</td></tr><tr><td>Swin-large-48 [38]</td><td>55.66</td><td>-</td></tr><tr><td>MViTv2-small-48 [35]</td><td>62.18</td><td>-</td></tr><tr><td>MViTv2-small-32 [35]</td><td>64.09</td><td>-</td></tr><tr><td>Ours</td><td>76.53</td><td>93.08</td></tr><tr><td>Ours(4-crops)</td><td>77.83</td><td>95.17</td></tr></table>
+
+Table 4. Comparison with the baselines on Slovo. The final number indicates the number of frames input to the model.
+
+Metrics. Following [22, 61, 63], we report top-1 and top-5 classification accuracy, as well as per-instance (P-I) and per-class (P-C) precision metrics. Each table includes results for single-crop and 4-crop inference, with all ablation studies conducted in the single-crop setting. For 4-crops inference, we sum the predictions from the three types of skeleton groupings along with the bone data to obtain the final prediction.
+
+# 4.2. Dataset
+
+We conduct experiments on four publicly available benchmarks: WLASL, MSASL, NMFs-CSL, and Slovo. All training occurs on the training set, with hyperparameter tuning performed on the validation set.
+
+MSASL [26] is a widely used dataset for American Sign Language (ASL). It contains a total of 16,054, 5,287, and 4,172 samples in the training, validation, and test sets, respectively, with a vocabulary of 1,000 signs. Two subsets, consisting of the top 100 and 300 words, are defined as MSASL100 and MSASL300, respectively.
+
+WLASL [32] has a larger vocabulary but fewer total samples than MSASL. It contains 2,000 signs performed by over 100 signers, with 14,289, 3,916, and 2,878 samples in the training, validation, and test sets. Similarly, two subsets, WLASL100 and WLASL300, are created. Both
+
+ASL datasets are collected from real-world scenarios, offering greater authenticity.
+
+NMFs-CSR [16] is a challenging Chinese Sign Language (CSL) dataset with a vocabulary of 1,067 and consists of 25,608 and 6,402 samples in the training and test sets, respectively.
+
+Slovo [27] is the latest dataset for Russian Sign Language (RSL). It contains 20,000 videos across 1,000 isolated RSL gestures from 194 signers. Notably, the authors provide 42 hand landmarks based on MediaPipe, so we do not re-extract joint points.
+
+# 4.3. Comparison with State-of-the-art Methods
+
+In this section, we compare our model with several state-of-the-art approaches. For single-modality methods, we categorize them as either RGB-based or skeleton-based. For multi-modality methods, we report the performance of their single-modality models before ensemble.
+
+MSASL [26]. As shown in the Tab. 1, RGB-based methods have developed slowly in recent years, and our approach shows a significant improvement when compared to them. Among skeleton-based methods, SignBERT [14] and BEST [60] designed self-supervised learning strategies to pre-train the model in order to alleviate issues related to limited interpretability and overfitting due to the small size of the SLR datasets. NASA [61] and SSRL [62] address the problem of neglecting explicit motion information and lexical signs during pre-training. However, none of these methods take into account the linguistic characteristics of sign language itself. Compared to them, our method achieves a substantial top-1 improvement of  $7.96\%$  in P-I and  $8.2\%$  in P-C on the full dataset. However, SSRL performs slightly better than our single-crop method on the 100-class subset, which we speculate is due to SPD requiring a certain amount of data to effectively learn the droppable nodes.
+
+WLASL [32]. SAM-SLR [22] and NLA-SLR [63] are currently the two best-performing heavy multi-modality ensemble models. Among them, SL-GCN is the skeleton model used in SAM-SLR for later ensemble, employing a multi-stream approach to further improve performance. K64 and V64 are skeleton and RGB models in NLASLR that use 64-frame input, where the skeleton is represented as a heatmap sequence, allowing their keypoint encoder to share the same architecture with the video encoder. DSTA [18] was the previously most accurate model. Compared to their results, our top-1 accuracy achieved a  $2.30\%$  improvement on WLASL2000.
+
+NMFs-CSL [16]. As shown in Tab. 3, NMFs-CSL is a balanced dataset in terms of class distribution, so we only report the P-I accuracy. The pretraining strategy of SignBERT [14] is highly effective on this dataset; however, its performance on top-5 is suboptimal. Our model does not use any additional training data, yet it shows a significant
+
+<table><tr><td rowspan="2">Skeleton</td><td colspan="2">MSASL1000</td><td colspan="2">MSASL200</td><td colspan="2">MSASL100</td><td colspan="2">WLASL2000</td><td colspan="2">WLASL300</td><td colspan="2">WLASL100</td><td colspan="2">NMFs-CLR</td></tr><tr><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td></tr><tr><td>type-1</td><td>70.71</td><td>86.36</td><td>84.62</td><td>94.04</td><td>86.26</td><td>95.24</td><td>53.54</td><td>83.18</td><td>79.04</td><td>94.01</td><td>84.50</td><td>94.96</td><td>75.26</td><td>95.47</td></tr><tr><td>bone</td><td>68.41</td><td>85.88</td><td>82.41</td><td>93.38</td><td>84.68</td><td>94.06</td><td>50.76</td><td>82.45</td><td>74.70</td><td>91.17</td><td>84.11</td><td>93.80</td><td>73.63</td><td>94.31</td></tr><tr><td>type-2</td><td>70.35</td><td>86.10</td><td>85.28</td><td>94.11</td><td>87.45</td><td>95.64</td><td>53.02</td><td>83.32</td><td>78.44</td><td>93.41</td><td>83.72</td><td>93.80</td><td>74.70</td><td>94.86</td></tr><tr><td>type-3</td><td>70.35</td><td>86.10</td><td>84.92</td><td>94.11</td><td>85.60</td><td>95.11</td><td>52.26</td><td>82.49</td><td>78.14</td><td>93.41</td><td>84.88</td><td>93.41</td><td>74.87</td><td>95.55</td></tr></table>
+
+Table 5. Performance of multi-grouping VSformer on three datasets.  
+
+<table><tr><td rowspan="2">weighting</td><td colspan="2">WLASL2000</td><td colspan="2">MSASL1000</td><td colspan="2">NMFs-CSL</td></tr><tr><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td></tr><tr><td>sum</td><td>55.98</td><td>87.07</td><td>73.18</td><td>88.64</td><td>76.65</td><td>97.17</td></tr><tr><td>acc</td><td>56.01</td><td>87.18</td><td>73.13</td><td>88.64</td><td>76.65</td><td>97.17</td></tr></table>
+
+Table 6. Performance of the two ensemble methods  
+
+<table><tr><td rowspan="2">Methods</td><td colspan="2">WLASL300</td><td colspan="2">WLASL2000</td></tr><tr><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td></tr><tr><td>Linear</td><td>74.70</td><td>92.81</td><td>46.32</td><td>77.24</td></tr><tr><td>GCN</td><td>76.35</td><td>91.77</td><td>51.18</td><td>80.82</td></tr><tr><td>SPD(ours)</td><td>79.04</td><td>94.01</td><td>53.54</td><td>83.18</td></tr></table>
+
+Table 7. Comparing our SPD with GCN and linear layers.  
+
+<table><tr><td rowspan="2">SPA</td><td rowspan="2">TMA</td><td colspan="2">WLASL300</td><td colspan="2">WLASL2000</td></tr><tr><td>Top-1</td><td>Top-5</td><td>Top-1</td><td>Top-5</td></tr><tr><td>✓</td><td></td><td>77.69</td><td>93.41</td><td>52.43</td><td>82.24</td></tr><tr><td></td><td>✓</td><td>77.99</td><td>93.26</td><td>52.36</td><td>81.83</td></tr><tr><td>✓</td><td>✓</td><td>79.04</td><td>94.01</td><td>53.54</td><td>83.18</td></tr></table>
+
+Table 8. Ablation study on the impact of our two types of attention mechanisms.
+
+improvement over SignBERT in overall performance, with a  $1.8\%$  improvement in top-1 accuracy and a  $4\%$  improvement in top-5 accuracy.
+
+Slovo [27]. Slovo is a recently released RSL dataset, with only limited experimental data available. We provide an up-to-date benchmark for better evaluating this dataset. As shown in Tab. 4, the results for Swin [38] and MViTv2 [35] were provided as baselines by the dataset authors. Notably, the skeleton landmarks provided by the authors includes only 42 hand joints, without arm data, which partially limits our model's performance. Despite this, we still achieve results significantly above the baselines.
+
+# 4.4. Ablation Study
+
+In this section, we conduct several key ablation experiments on the WLASL dataset [32] and report P-I accuracy to validate the effectiveness of our approach.
+
+SPD. We first replace SPD with two alternative methods. As shown in Tab. 7, "Linear" denotes a three-layer Conv-2D module with the same output dimension as SPD, while "GCN" refers to three decoupled graph convolution
+
+layer without the weak joints dropping strategy. To ensure consistent output data shape, these two methods directly input the simplified 24 joint points. The complete SFB implementation achieves the highest accuracy, indicating that our dropping strategy better facilitates the extraction of sign language action features.
+
+Multi-Grouping. As shown in Tab. 5, we present the final performance of four types of crops. The overall performance of each grouping type varies only slightly, mainly depending on the sample distribution of the dataset. However, integrating all four crops significantly enhances action classification performance. The performance of the bone-based grouping is slightly lower than that of joint-based results, primarily because our model is designed for points and is less sensitive to vector-based data.
+
+SAM-SLR [21] ultimately using manually set weights to control the ensemble. However, we argue that such weights lack interpretability. As shown in Tab. 6, we employed two methods to integrate the results: "acc" refers to using the individual top-1 accuracy of each crop as the weight, while "sum" represents directly summing the outputs of the final layer of the model. Both methods resulted in an about  $2\%$  improvement in top-1 accuracy, demonstrating that our ensemble approach is both general and robust.
+
+VSformer. For our two types of self-attention, we replace one with a linear layer to test the performance of the other individually. As shown in Tab. 8, the full VSformer achieves significantly better performance, indicating that features in both temporal and spatial dimensions are crucial for distinguishing sign language actions.
+
+# 5. Conclusion
+
+In this work, we design a skeleton recognition model tailored specifically for ISLR, termed VSNet. The core idea of our approach is to focus on the truly essential elements of sign language. To this end, we first introduce a weak joints dropping strategy to prioritize more critical points in the skeleton. We then propose the VSformer to capture the linguistic characteristics of sign language, where Visual Symbols (VS) represent smaller components of sign language actions. We conduct extensive experiments to validate the effectiveness of the proposed VSNet, achieving state-of-the-art performance across four benchmarks.
+
+# Acknowledgements
+
+This work was supported in part by National Natural Science Foundation of China (No. 62476052), Sichuan Science and Technology Program (No. 2024NSFSC1473), the open project of Sichuan Provincial Key Laboratory of Philosophy and Social Science for Language Intelligence in Special Education (No. YYZN-2023-3), the Humanities and Social Sciences Project of the Ministry of Education (No. 23YJA740013), and Central Guidance for Local Science and Technology Development Fund Projects (No. 2024ZYD0268).
+
+# References
+
+[1] Nikolas Adaloglou, Theocharis Chatzis, Ilias Papastratis, Andreas Stergioulas, Georgios Th Papadopoulos, Vassia Zacharopoulou, George J Xydopoulos, Klimnis Atzakas, Dimitris Papazachariou, and Petros Daras. A comprehensive study on deep learning-based methods for sign language recognition. IEEE TMM, 24:1750-1762, 2021. 1, 2  
+[2] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. Bsl-1k: Scaling up co-articulated sign language recognition using mouthing cues. In ECCV. Springer, 2020. 6  
+[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 3  
+[4] Amanda Cardoso Duarte, Samuel Albanie, Xavier Giro Nieto, and Gül Varol. Sign language video retrieval with freeform textual queries. In CVPR, pages 14074-14084, 2022. 2  
+[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 2, 6, 7  
+[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 3  
+[7] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In CVPR, pages 5120-5130, 2022. 1  
+[8] Ke Cheng, Yifan Zhang, Congqi Cao, Lei Shi, Jian Cheng, and Hanqing Lu. Decoupling gcn with dropgraph module for skeleton-based action recognition. In ECCV, pages 536-553. Springer, 2020. 2  
+[9] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. NeurIPS, 34:9355-9366, 2021. 3  
+[10] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.3  
+[11] Jeonghyeok Do and Munchurl Kim. Skateformer: Skeletal-temporal transformer for human action recognition. arXiv preprint arXiv:2403.09508, 2024. 3, 5
+
+[12] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In CVPR, pages 1110-1118, 2015. 1, 3  
+[13] Karen Emmorey. Language, cognition, and the brain: Insights from sign language research. Psychology Press, 2001. 1  
+[14] Hezhen Hu, Weichao Zhao, Wengang Zhou, Yuechen Wang, and Houqiang Li. Signbert: Pre-training of hand-model-aware representation for sign language recognition. In ICCV, pages 11087-11096, 2021. 6, 7  
+[15] Hezhen Hu, Wengang Zhou, and Houqiang Li. Hand-model-aware sign language recognition. In AAAI, pages 1558-1566, 2021. 2, 6, 7  
+[16] Hezhen Hu, Wengang Zhou, Junfu Pu, and Houqiang Li. Global-local enhancement network for nmf-aware sign language recognition. ACM transactions on multimedia computing, communications, and applications (TOMM), 17(3): 1-19, 2021. 2, 7  
+[17] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE TPAMI, 45 (9):11221-11239, 2023. 6  
+[18] Lianyu Hu, Liqing Gao, Zekang Liu, and Wei Feng. Dynamic spatial-temporal aggregation for skeleton-aware sign language recognition. In LREC/COLING, 2024. 6, 7  
+[19] Jie Huang, Wengang Zhou, Houqiang Li, and Weiping Li. Attention-based 3d-cnns for large-vocabulary sign language recognition. IEEE TCSVT, 29(9):2822-2832, 2018. 2  
+[20] Jie Huang, Wengang Zhou, Qilin Zhang, Houqiang Li, and Weiping Li. Video-based sign language recognition without temporal segmentation. In AAAI, 2018. 2  
+[21] Songyao Jiang, Bin Sun, Lichen Wang, Yue Bai, Kunpeng Li, and Yun Fu. Skeleton aware multi-modal sign language recognition. In CVPR, pages 3413-3423, 2021. 1, 3, 4, 5, 8  
+[22] Songyao Jiang, Bin Sun, Lichen Wang, Yue Bai, Kunpeng Li, and Yun Fu. Sign language recognition via skeleton-aware multi-model ensemble. arXiv preprint arXiv:2110.06161, 2021. 6, 7  
+[23] Tao Jiang, Necati Cihan Camgoz, and Richard Bowden. Skeleton: Skeletal transformers for robust body-posed estimation. In CVPR, pages 3394–3402, 2021. 1, 3  
+[24] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In ECCV, pages 196-214. Springer, 2020. 3, 4  
+[25] Trevor Johnston and Adam Schembri. Australian Sign Language (Auslan): An introduction to sign language linguistics. Cambridge University Press, 2007. 1  
+[26] Hamid Reza Vaezi Joze and Oscar Koller. Ms-asl: A large-scale data set and benchmark for understanding american sign language. arXiv preprint arXiv:1812.01053, 2018. 2, 7  
+[27] Alexander Kapitanov, Kvanchiani Karina, Alexander Nagaev, and Petrova Elizaveta. Slovo: Russian sign language dataset. In International Conference on Computer Vision Systems, pages 63-73. Springer, 2023. 2, 7, 8
+
+[28] Oscar Koller. Quantitative survey of the state of the art in sign language recognition. arXiv preprint arXiv:2008.09918, 2020. 2  
+[29] Oscar Koller, Sepehr Zargaran, Hermann Ney, and Richard Bowden. Deep sign: Enabling robust statistical continuous sign language recognition via hybrid cnn-hmms. *IJCV*, 126: 1311–1325, 2018. 2  
+[30] Taeryung Lee, Yeonguk Oh, and Kyoung Mu Lee. Human part-wise 3d motion context learning for sign language recognition. In ICCV, pages 20740-20750, 2023. 1, 6  
+[31] Chao Li, Qiaoyong Zhong, Di Xie, and Shiliang Pu. Cooccurrence feature learning from skeleton data for action recognition and detection with hierarchical aggregation. arXiv preprint arXiv:1804.06055, 2018. 1, 3  
+[32] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1459-1469, 2020. 2, 7, 8  
+[33] Dongxu Li, Xin Yu, Chenchen Xu, Lars Petersson, and Hongdong Li. Transferring cross-domain knowledge for video sign language recognition. In CVPR, pages 6205-6214, 2020. 6  
+[34] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In CVPR, pages 3595–3603, 2019. 3  
+[35] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mviv2: Improved multiscale vision transformers for classification and detection. In CVPR, pages 4804-4814, 2022. 7, 8  
+[36] Jiajia Liao, Yujun Liu, Yingchao Piao, Jinhe Su, Guorong Cai, and Yundong Wu. Gle-net: A global and local ensemble network for aerial object detection. International Journal of Computational Intelligence Systems, 15(1):2, 2022. 7  
+[37] Jinfu Liu, Xinshun Wang, Can Wang, Yuan Gao, and Mengyuan Liu. Temporal decoupling graph convolutional network for skeleton-based gesture recognition. IEEE TMM, 26:811-823, 2023. 3  
+[38] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, pages 3202-3211, 2022. 3, 7, 8  
+[39] Anna Mindess. Reading between the signs: Intercultural communication for sign language interpreters. Nicholas Brealey, 2014. 1  
+[40] Liliane Momeni, Hannah Bull, KR Prajwal, Samuel Albanie, Gül Varol, and Andrew Zisserman. Automatic dense annotation of large-vocabulary sign language videos. In ECCV, pages 671-690. Springer, 2022. 2  
+[41] Evonne Ng, Shiry Ginosar, Trevor Darrell, and Hanbyul Joo. Body2hands: Learning to infer 3d hands from conversational gesture body dynamics. In CVPR, pages 11865-11874, 2021. 1, 3  
+[42] Ladislav Rampášek, Michael Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. Recipe for a general, powerful, scalable graph transformer. NeurIPS, 35:14501-14515, 2022. 3
+
+[43] Razieh Rastgoo, Kourosh Kiani, and Sergio Escalera. Sign language recognition: A deep survey. Expert Systems with Applications, 164:113794, 2021. 2  
+[44] Wendy Sandler. Sign language and linguistic universals. Cambridge University, 2006. 1  
+[45] Noha Sarhan and Simone Frintrop. Unraveling a decade: a comprehensive survey on isolated sign language recognition. In ICCV, pages 3210-3219, 2023. 2  
+[46] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In CVPR, pages 12026-12035, 2019. 3  
+[47] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with multi-stream adaptive graph convolutional networks. IEEE TIP, 29:9532-9545, 2020. 2  
+[48] Chenyang Si, Wentao Chen, Wei Wang, Liang Wang, and Tieniu Tan. An attention enhanced graph convolutional LSTM network for skeleton-based action recognition. In CVPR, pages 1227-1236, 2019. 3  
+[49] Neil Song and Yu Xiang. Slgtformer: An attention-based approach to sign language recognition. arXiv preprint arXiv:2212.10746, 2022. 3, 4  
+[50] Yi-Fan Song, Zhang Zhang, Caifeng Shan, and Liang Wang. Stronger, faster and more explainable: A graph convolutional baseline for skeleton-based action recognition. In proceedings of the 28th ACM international conference on multimedia, pages 1625-1633, 2020. 2  
+[51] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, pages 5693-5703, 2019. 3, 4  
+[52] Clayton Valli. Linguistics of American Sign Language: An Introduction. Gallaudet University Press, 2000. 1  
+[53] Gül Varol, Liliane Momeni, Samuel Albanie, Triantafyllos Afouras, and Andrew Zisserman. Scaling up sign spotting through sign language dictionaries. *IJCV*, 130(6):1416–1439, 2022. 2  
+[54] Hongsong Wang and Liang Wang. Modeling temporal dynamics and spatial configurations of actions using two-stream recurrent neural networks. In CVPR, pages 499-508, 2017. 2  
+[55] Hanjie Wang, Xiujuan Chai, and Xin Chen. A novel sign language recognition framework using hierarchical grassmann covariance matrix. IEEE TMM, 21(11):2806-2814, 2019. 2  
+[56] Chengcheng Wei, Jian Zhao, Wengang Zhou, and Houqiang Li. Semantic boundary detection with reinforcement learning for continuous sign language recognition. IEEE TCSVT, 31(3):1138-1149, 2020.  
+[57] Pan Xie, Mengyi Zhao, and Xiaohui Hu. Pisltrc: Position-informed sign language transformer with content-aware convolution. IEEE TMM, 24:3908-3919, 2021. 2  
+[58] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 1, 6, 7  
+[59] Aoxiong Yin, Zhou Zhao, Weike Jin, Meng Zhang, Xingshan Zeng, and Xiaofei He. Mlslt: Towards multilingual sign language translation. In CVPR, pages 5109-5119, 2022. 1
+
+[60] Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. Best: Bert pre-training for sign language recognition with coupling tokenization. In AAAI, pages 3597-3605, 2023. 6, 7  
+[61] Weichao Zhao, Hezhen Hu, Wengang Zhou, Yunyao Mao, Min Wang, and Houjiang Li. Masa: Motion-aware masked autoencoder with semantic alignment for sign language recognition. IEEE TCSVT, 2024. 6, 7  
+[62] Weichao Zhao, Wengang Zhou, Hezhen Hu, Min Wang, and Houqiang Li. Self-supervised representation learning with spatial-temporal consistency for sign language recognition. arXiv preprint arXiv:2406.10501, 2024. 6, 7  
+[63] Ronglai Zuo, Fangyun Wei, and Brian Mak. Natural language-assisted sign language recognition. In CVPR, pages 14890-14900, 2023. 6, 7

2025/VSNet_ Focusing on the Linguistic Characteristics of Sign Language/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9683809fddb618d4fe8a5df52fe9ed388309dc1ec835c8a11206d3894471c8e6
+size 680534

2025/VTON 360_ High-Fidelity Virtual Try-On from Any Viewing Direction/4af7cfba-2431-4854-85f1-f3f911dcc811_content_list.json

@@ -0,0 +1,1362 @@
+[
+    {
+        "type": "text",
+        "text": "VTON 360: High-Fidelity Virtual Try-On from Any Viewing Direction",
+        "text_level": 1,
+        "bbox": [
+            142,
+            130,
+            852,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zijian He<sup>1</sup> Yuwei Ning<sup>1</sup> Yipeng Qin<sup>2</sup> Guangrun Wang<sup>1</sup> Sibei Yang<sup>3</sup> Liang Lin<sup>1,4,5</sup> Guanbin Li<sup>1,4,5*</sup>  \n<sup>1</sup>Sun Yat-sen University  \n<sup>2</sup>Cardiff University  \n<sup>3</sup>ShanghaiTech University  \n<sup>4</sup>Guangdong Key Laboratory of Big Data Analysis and Processing  \n<sup>5</sup>Peng Cheng Laboratory  \nhezj39@mail2.sysu.edu.cn, yuwei.ning@hust.edu.cn, {qinyipengl991, wanggrun}@gmail.com, yangsb@shanghaiitech.edu.cn linliang@ieee.org, liguanbin@mail.sysu.edu.cn",
+        "bbox": [
+            99,
+            159,
+            898,
+            250
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/3e4ed0ca9fe9c0466bbb1e4f4019077b1eea386c4b063ea02f3897e6315ec9c9.jpg",
+        "image_caption": [
+            "Figure 1. Results of VTON 360. Our VTON 360 enables high-fidelity 3D Virtual Try-On (VTON) by seamlessly adapting E-commerce garments onto a clothed 3D human model, supporting full  $360^{\\circ}$  view rendering. The highlighted bounding boxes (dashed line) demonstrate our method's ability to preserve intricate clothing details and patterns (e.g., collar accessories, horizontal line patterns, logos, texts, numbers) across diverse garment types."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            252,
+            509,
+            474
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/18ca224d94844f9dde5541e3e34a54b63839bcf1d69569779a98633de60a2a1f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            252,
+            709,
+            474
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/a5889a307af41522f81b308115e028b9c3f02dce5c03771e67bcdd87075b042d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            714,
+            252,
+            906,
+            474
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            542,
+            326,
+            558
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Virtual Try-On (VTON) is a transformative technology in e-commerce and fashion design, enabling realistic digital visualization of clothing on individuals. In this work, we propose VTON 360, a novel 3D VTON method that addresses the open challenge of achieving high-fidelity VTON that supports any-view rendering. Specifically, we leverage the equivalence between a 3D model and its rendered multiview 2D images, and reformulate 3D VTON as an extension of 2D VTON that ensures 3D consistent results across multiple views. To achieve this, we extend 2D VTON models to include multi-view garments and clothing-agnostic human body images as input, and propose several novel techniques to enhance them, including: i) a pseudo-3D pose representation using normal maps derived from the SMPL-X 3D human model, ii) a multi-view spatial attention mechanism that models the correlations between features from different viewing angles, and iii) a multi-view CLIP embedding that enhances the garment CLIP features used in 2D VTON with camera information. Extensive experiments on large-scale real datasets and clothing images from e-commerce",
+        "bbox": [
+            88,
+            569,
+            485,
+            872
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "platforms demonstrate the effectiveness of our approach.  \nProject page: https://scnuhealthy.github.io/  \nVTON360.",
+        "bbox": [
+            511,
+            544,
+            906,
+            589
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            513,
+            601,
+            643,
+            616
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Virtual Try-On (VTON) enables realistic digital visualization of clothing on individuals and has emerged as a transformative technology in e-commerce and fashion design. While significant research efforts have been made on 2D VTON solutions [8, 12, 19, 33, 39], these approaches are inherently limited in their representation of view-related features. To overcome this limitation and enable high-fidelity any-view rendering, 3D VTON methods were introduced.",
+        "bbox": [
+            509,
+            627,
+            906,
+            748
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "3D VTON requires accurate garment transfer onto a 3D human body while ensuring realistic garment fitting, texture preservation, and 3D consistency. The two primary aims of 3D VTON are i) achieving high-fidelity and ii) supporting any-view rendering. Leveraging the inherent capability of 3D models for any-view rendering, early 3D VTON methods [13, 15, 28] make clothing simulation on synthetic human bodies. Specifically, these methods utilized 3D scanners to capture clothing meshes, followed by the development of specialized dressing algorithms. Although effec",
+        "bbox": [
+            509,
+            750,
+            908,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Corresponding author is Guanbin Li.",
+        "bbox": [
+            109,
+            886,
+            310,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "26388",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tive, these methods rely on costly 3D scanning equipment and the physical presence of the human body/clothing (i.e., not fully virtual), restricting their practicality in real-world applications. As a byproduct, most early methods focused on developing geometrically correct dressing algorithms using standard templates of human body and clothing models. Addressing this limitation, researchers extended 3D VTON by introducing algorithms that reconstruct 3D clothing models from input images, enabling the use of image-based clothing inputs [3, 32, 41, 42]. However, since input clothing images (usually frontal) are inherently 2D and lack multi-view information, this approach struggles to reconstruct high-fidelity clothing models that can be rendered well from all viewing directions.",
+        "bbox": [
+            89,
+            90,
+            480,
+            303
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To complement this missing information, DreamVTON [50] introduces a novel approach that leverages Text-to-Image (T2I) diffusion models to reconstruct both the human body and clothing from input images. Its key insight is that T2I models learned view-agnostic \"concepts\" of both bodies and garments during their training, and that the corresponding concepts for the input body and clothing images can be obtained using LoRA [21]. By utilizing Score Distillation Sampling (SDS) [37], DreamVTON can generate visual-pleasing 3D VTON results by ensuring consistency between renderings from arbitrary viewpoints and the concepts. Nonetheless, DreamVTON's high flexibility comes at the cost of low fidelity. This limitation stems from the fact that the concepts learned by T2I models are semantic in nature, thus lacking 3D geometric consistency and pixel-level accuracy with respect to the input body and clothing images. Recently, a concurrent work, namely GaussianVTON [6], partially addressed this limitation by formulating 3D VTON as a 3D scene editing task, where a given 3D human model is edited using multi-view images generated by 2D VTON methods. While it significantly enhances the fidelity of the human body, the fidelity and 3D consistency of clothing remain problematic, as there are no 2D VTON methods that can generate multi-view images with 3D consistency. Therefore, to the best of our knowledge, achieving high-fidelity 3D VTON that supports any-view rendering remains an open challenge.",
+        "bbox": [
+            91,
+            306,
+            482,
+            715
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this work, we address the above-mentioned challenge via proposing VTON 360, a novel 3D VTON method that achieves high-fidelity VTON from arbitrary viewing directions. Similar to GaussianVTON [6], our method edits a given 3D human model by inpainting the rendered images using a latent diffusion model. However, we set ourselves apart through our novel garment fidelity preservation strategy that can generate high-fidelity on-body garments in all viewing directions. Specifically, we first extend both the garment and clothing-agnostic human body inputs to typical 2D VTON models to leverage multi-view information, including paired front and back view garment images as well",
+        "bbox": [
+            89,
+            719,
+            482,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "as a set of multi-view clothing-agnostic human body images sampled from random azimuth angles. Then, we propose several novel enhancements to bridge the gap between typical 2D VTON methods and our multi-view 3D consistency requirements: i) We propose a pseudo-3D pose representation using normal maps derived from the SMPL-X 3D human model, which captures fine-grained surface orientation details and provides more consistent geometry across views compared to the 2D pose representations (semantic segmentation maps) used in 2D VTON models. ii) We design a Multi-view Spatial Attention mechanism that models the correlations between features from different viewing angles, featuring a novel \"correlation\" matrix modeling the relationships among different input views. iii) We propose a multi-view CLIP embedding that enhances the garment CLIP embedding used in 2D VTON methods with camera information, thereby facilitating network learning of features relevant to a particular view. Together, these innovations enable our 2D VTON model to generate high-quality, multi-view and 3D-consistent virtual try-on results. Extensive experiments on Thuman2.0 [55] and MVHumanNet [51] datasets demonstrate that our method achieves high fidelity 3D VTON which supports any-view rendering. In addition, we show the effectiveness and generalizability of our methodology by testing it using garments from e-commerce platforms. Our conclusions include:",
+        "bbox": [
+            511,
+            90,
+            903,
+            482
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose a novel 3D Virtual Try-On (VTON) method, namely VTON 360, which achieves high-fidelity VTON from arbitrary viewing directions.",
+            "- Leveraging the equivalence between a 3D model and its rendered multi-view 2D images, we reformulate 3D VTON as an extension of 2D VTON that ensures consistent results across multiple views. Specifically, we introduce several novel techniques, including: (i) pseudo-3D pose representation; (ii) multi-view spatial attention; and (iii) multi-view CLIP embedding. These innovations enhance traditional 2D VTON models to generate multiview and 3D-consistent results.",
+            "- Extensive experimental results on two large real datasets as well as real clothing images from e-commerce platforms demonstrate the effectiveness of our approach."
+        ],
+        "bbox": [
+            513,
+            483,
+            903,
+            710
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            513,
+            723,
+            653,
+            739
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2D Virtual Try-On. 2D Virtual Try-On (VTON) aims to transfer a desired garment to the corresponding region of a target human image while preserving the human pose and identity. Early methods [2, 8, 10, 11, 16, 18, 29, 31, 38, 54, 56] use Generative Adversarial Networks (GANs) to deform the garments to match the target body shape, which a critical step for achieving realistic VTON. However, accurately adapting to diverse real-world conditions remains a significant challenge. Addressing this issue, recent methods [12, 19, 33, 60] reframe 2D VTON as a conditioned in-",
+        "bbox": [
+            511,
+            750,
+            903,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "26389",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "painting task, leveraging the strong priors provided by diffusion models [20, 43, 45] to achieve promising results. This strategy is further improved by [9, 26, 53], which introduce a ReferenceNet to extract hierarchical garment features and apply attention mechanisms to condition the Main UNet.",
+        "bbox": [
+            89,
+            90,
+            480,
+            167
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3D Virutal Try-On. For 3D Virtual Try-On (VTON), traditional methods [4, 13, 15, 28, 36] rely on 3D scanning or cloth simulation to generate highly precise body and garment geometry. These methods were then extended by learning-based methods [3, 32] that employ differentiable rendering to dress the SMPL [30] model with a desired garment mesh. Despite their effectiveness, such methods rely on costly 3D scanning and the physical presence of human body/clothing, limiting their application in the real world. Addressing this limitation, M3D-VTON [59] proposes a depth-based 3D VTON framework to reconstruct 3D clothed human models from 2D human and garment images, but the results often suffer from explicit warping artifacts. To improve 3D VTON results, recent methods [23, 24, 50, 62] resort to text-to-image (T2I) diffusion models and employ the Score Distillation Sampling (SDS) loss [40] to ensure consistency among different viewing directions. Specifically, TeCH [24] adapts the generative priors of T2I diffusion model to the specific person and clothes by training descriptive text prompts with DreamBooth [40]. DreamWaltz [23] leverages Pose ControlNet [57] to attain clothed human body models. DreamVTON [50] introduces a multi-concept LoRA [21] to personalize the T2I diffusion model, and uses a template-based optimization mechanism that combines with SDS loss to better preserve patterns on the garment. Although effective, these methods often produce results lacking in fidelity, as the concepts learned by T2I models are semantic rather than at the pixel level. Concurrent to our work, GaussianVTON [6] proposes an alternative approach by combining Gaussian Splatting [25] with pre-trained 2D VTON models and formulate it as an editing task. However, since there are no 2D VTON methods that can generate multi-view images with 3D consistency, the fidelity and 3D consistency of the clothing generated remain problematic.",
+        "bbox": [
+            91,
+            178,
+            483,
+            708
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Radiance Field-based 3D Human or Scene Editing. Recently, radiance field-based editing has attracted significant interest due to its efficient differentiable rendering capabilities, sparking substantial advancements in text-driven 3D editing. For example, InstructN2N [17] employ an image-based diffusion model InstructP2P [5] to modify the rendered image by the user's text description with a variant of the score distillation sampling (SDS) [37] loss. GaussianEditor [7] applies Gaussian Splatting [25] as 3D representation instead of NeRF, adopting Gaussian semantic tracking to track target Gaussian values, significantly improving editing speed and controllability. To enable accu",
+        "bbox": [
+            89,
+            719,
+            482,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "rate location and appearance control, subsequent works [47, 61] specify the editing region using the attention score or with a segmentation model. TIP-Editor [63] proposes a content personalization step dedicated to the reference image based on LoRA, achieving the editing following a hybrid text-image prompt. GaussCtrl [48] leverage depth conditions and attention-based latent code alignment to achieve 3D-aware multi-view consistent editing instead of iteratively editing single views using SDS loss. However, these works primarily focus on global appearance modifications within a text-driven pipeline, while our approach emphasizes preserving fine textural details from different viewing directions throughout the editing process.",
+        "bbox": [
+            511,
+            90,
+            906,
+            287
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Preliminary",
+        "text_level": 1,
+        "bbox": [
+            511,
+            300,
+            638,
+            316
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Latent Diffusion Model. Latent Diffusion Model [39] is a variant of diffusion models that performs denoising within the latent space of a Variational Autoencoder (VAE) [27]. During training, given a fixed encoder  $\\mathcal{E}$ , an input image  $x$  is transformed into its latent representation  $z_0 = \\mathcal{E}(x)$ . A conditional diffusion model  $\\hat{\\epsilon}_{\\theta}$ , typically implemented with a UNet architecture, is then trained using a weighted denoising score matching objective:",
+        "bbox": [
+            511,
+            325,
+            906,
+            446
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {L D M} = \\mathbb {E} _ {\\mathbf {z}, \\mathbf {c}, \\boldsymbol {\\epsilon}, t} [ \\boldsymbol {\\epsilon} - \\| \\hat {\\epsilon} _ {\\theta} (\\mathbf {z} _ {\\mathbf {t}}; \\mathbf {c}, t) \\| _ {2} ^ {2} ] \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            583,
+            455,
+            906,
+            474
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{z}_t\\coloneqq \\alpha_t\\mathbf{x} + \\sigma_t\\boldsymbol {\\epsilon}$  denotes the forward diffusion process at timestep  $t$  .  $\\alpha_{t},\\sigma_{t}$  are time-dependent functions defined by the diffusion model formulation; c denotes the conditional input and  $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{1})$  is Gaussian noise. During inference, data samples are generated by initiating from Gaussian noise  $\\mathbf{z}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{1})$  and iteratively refining it using a DDIM [44] sampler.",
+        "bbox": [
+            511,
+            484,
+            908,
+            590
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4. Method",
+        "text_level": 1,
+        "bbox": [
+            511,
+            603,
+            604,
+            618
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our method leverages the equivalence between a 3D model and its rendered multi-view 2D images to achieve high-fidelity, any-view 3D VTON. Specifically, as Fig. 2 shows, given an input 3D human model and a garment image, our method 1) renders the 3D model into multi-view 2D images and 2) formulates 3D VTON as a consistent, unified 2D VTON process across these rendered views; 3) By reconstructing the edited images into a 3D model using existing 3D reconstruction methods, we ensure visual coherence and precise garment alignment from any viewing angle. Among them, the second step is crucial as existing 2D VTON methods [9, 26, 53] lack 3D knowledge, preventing them from generating multi-view images with 3D consistency.",
+        "bbox": [
+            511,
+            628,
+            908,
+            824
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To address this challenge, we propose several novel techniques (Sec. 4.2) that equip a typical 2D VTON network (Sec. 4.1), which is built on a latent diffusion model [39], with the capability to generate 3D-consistent results. We use Gaussian Splatting [25] as our 3D representation.",
+        "bbox": [
+            511,
+            825,
+            906,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "26390",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/50456d17e5ccbb76f3e583fe2b0ebac3fa363e2d03f09e2a13d08f82ca40471e.jpg",
+        "image_caption": [
+            "Figure 2. Overview of VTON 360. Given an input 3D human model  $\\mathbf{G}_{\\mathrm{src}}$  and a pair of garment images  $(g_{f}, g_{b})$ , our method 1) renders  $\\mathbf{G}_{\\mathrm{src}}$  into multi-view 2D images (left) and 2) edits the rendered multi-view 2D images (middle); 3) reconstructs the edited images into a 3D model  $\\mathbf{G}_{\\mathrm{VTON}}$  (right). In the crucial step 2), we propose three novel techniques to equip a typical 2D VTON network with the capability to generate 3D-consistent results: 1) Pseudo-3D Pose Input, 2) Multi-view Spatial Attention, and 3) Multi-view CLIP Embedding."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            89,
+            88,
+            903,
+            431
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/c3dd68bc18957e8ae4d2aa15c28ee2931189de321d4ce4c838cc8d0a92c0acbb.jpg",
+        "image_caption": [
+            "Figure 3. DensePose (2D) vs. SMPL-X normal map (pseudo-3D) representations. DensePose applies uniform labels per body part, lacking 3D consistency across views and causing artifacts and temporal inconsistencies (highlighted with red boxes). In contrast, SMPL-X normal maps capture fine surface details, ensuring geometric coherence and stable, realistic shading across views."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            503,
+            480,
+            679
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1. Recap of 2D VTON Framework",
+        "text_level": 1,
+        "bbox": [
+            89,
+            784,
+            372,
+            800
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Following previous works [12, 26, 53], we formulate 2D VTON as an exemplar-based inpainting problem, aiming to fill an input clothing-agnostic image  $\\mathbf{A}$  with a given garment image  $g$ , where  $\\mathbf{A}$  is obtained following the method used in [53]. As illustrated in Fig. 2 (middle), the network architecture is based on the latent diffusion model [39] with",
+        "bbox": [
+            89,
+            809,
+            483,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "an encoder  $\\mathcal{E}$  and comprises two components:",
+        "bbox": [
+            513,
+            506,
+            816,
+            520
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- A GarmentNet [9, 53] that extracts features from  $\\mathcal{E}(g)$ .",
+            "- A Main UNet that inpaints  $\\mathbf{A}$  according to i) detailed garment features extracted by the GarmentNet; ii) the 2D pose of  $\\mathbf{A}$  represented by semantic labels using Dense-Pose [14]; iii) CLIP embeddings of input garment  $g$ . Among them, i) and ii) together with noise are input to the self-attention layers, while iii) is input to the cross-attention layers of the Main UNet."
+        ],
+        "bbox": [
+            511,
+            520,
+            903,
+            640
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Both the GarmentNet and the Main UNet share the same network architecture.",
+        "bbox": [
+            511,
+            641,
+            903,
+            670
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2. Multi-view 2D VTON with 3D Consistency",
+        "text_level": 1,
+        "bbox": [
+            511,
+            681,
+            877,
+            698
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To enable the aforementioned 2D VTON model to generate multi-view and 3D-consistent results, we propose the following novel enhancements to its design:",
+        "bbox": [
+            511,
+            704,
+            903,
+            750
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Multi-view Inputs. We extend both inputs to the model:",
+        "bbox": [
+            513,
+            750,
+            887,
+            763
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Multi-view Garment Inputs: We extend the input garment representation from a single image  $g$  to paired front and back view images  $g_{f}$ ,  $g_{b}$ , providing comprehensive garment information across all viewing angles. Accordingly, we use the encoder  $\\mathcal{E}$  to encode  $g_{f}, g_{b}$  into their latent representations  $\\mathcal{E}(g_{f}), \\mathcal{E}(g_{b})$  and feed them into GarmentNet to obtain their multi-layer features  $F_{f}^{l}$  and  $F_{b}^{l}$  at layer  $l$ , respectively.",
+            "- Multi-view Clothing-agnostic Image Inputs: Based on"
+        ],
+        "bbox": [
+            513,
+            765,
+            905,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "26391",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "the equivalence between a 3D human model and its rendered multi-view 2D images, we extend the input human body representation from a single, clothing-agnostic image,  $\\mathbf{A}$ , to a set of  $m$  multi-view images, denoted as  $\\mathbf{A}_1, \\mathbf{A}_2, \\dots, \\mathbf{A}_m$ . These images are sampled from random azimuth angles, allowing the 2D VTON model to access comprehensive, multi-view information from the input 3D human model.",
+        "bbox": [
+            102,
+            90,
+            482,
+            210
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Pseudo-3D Pose Input. As shown in Fig. 3, the 2D Dense-Pose representations [14] commonly used in state-of-the-art 2D VTON methods [9, 26] assign a uniform semantic label to all pixels within each body part (e.g., thigh), inherently lack 3D geometric consistency across multiple views, and often introduce artifacts and temporal inconsistencies. To address these limitations, we propose a novel pseudo-3D pose representation: the normal maps  $\\mathbf{N}$  derived from the SMPL-X [35] model of the input body. These normal maps capture fine-grained surface orientation details, providing a more consistent representation across views by preserving geometric structure in the 3D space. Furthermore, they enable smoother, temporally stable transitions and realistic shading effects across multi-view scenarios. In practice, we employ a lightweight PoseEncoder  $\\mathcal{E}'$  [22] and feed  $\\mathcal{E}'(\\mathbf{N})$  into the Main UNet. We obtain the SMPL-X model from the multi-view images of the input body using EasyMoCap [1].",
+        "bbox": [
+            89,
+            212,
+            482,
+            469
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Accordingly, we concatenate three components as the enhanced input to the Main UNet: i) a noise latent  $z_{t}$ ; ii) the encoded pseudo-3D poses  $\\mathcal{E}'(\\mathbf{N_1}),\\mathcal{E}'(\\mathbf{N_2}),\\dots,\\mathcal{E}'(\\mathbf{N_m})$ ; and iii) the encoded multi-view clothing-agnostic images  $\\mathcal{E}(\\mathbf{A_1}),\\mathcal{E}(\\mathbf{A_2}),\\dots,\\mathcal{E}(\\mathbf{A_m})$ . Let  $F_{1}^{l},F_{2}^{l},\\dots,F_{m}^{l}$  be the feature representations at layer  $l$  of the Main UNet, and recall the garment features  $F_{f}^{l}$  and  $F_{b}^{l}$  defined above, we enhance the self-attention layers of the Main UNet as:",
+        "bbox": [
+            89,
+            470,
+            482,
+            589
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Multi-view Spatial Attention. To cope with the aforementioned multi-view input features and ensure their consistency, we draw insights from the temporal attention layer commonly used in video generation and editing [49, 58] and extend it to our multi-view spatial attention layer, denoted as MVAttention. The key distinction of our MVAttention is that its input multi-view features  $F_{1}^{l}, F_{2}^{l}, \\ldots, F_{m}^{l}$  are from images captured from non-uniform spatial intervals, with the viewing angles varying randomly. Consequently, features from similar views exhibit a higher correlation, while those from distinct views are largely independent. To model this relationship, we construct a \"correlation\" matrix  $C$  based on the angular disparity obtained from camera rotation matrices of the input multi-view images, and define our MVAttention as follows:",
+        "bbox": [
+            89,
+            590,
+            482,
+            816
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathbf {F} ^ {\\mathbf {l}} = \\left[ F _ {1} ^ {l} \\oplus F _ {2} ^ {l} \\dots \\oplus F _ {m} ^ {l} \\right], \\hat {\\mathbf {F}} ^ {\\mathbf {l}} = \\left[ \\mathbf {F} ^ {\\mathbf {l}} \\oplus F _ {f} ^ {l} \\oplus F _ {b} ^ {l} \\right] \\\\ Q = W ^ {Q} \\mathbf {F} ^ {1}, K = W ^ {K} \\hat {\\mathbf {F}} ^ {1}, V = W ^ {V} \\hat {\\mathbf {F}} ^ {1} \\tag {2} \\\\ A _ {i} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {i} \\times \\left(C _ {i} \\cdot K ^ {T}\\right)}{\\sqrt {d}}\\right), \\mathbf {H} _ {\\mathbf {i}} ^ {\\mathbf {l}} = A _ {i} \\times V _ {i} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            112,
+            825,
+            480,
+            905
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $i\\in \\{1,2,\\dots,m\\}$  denotes  $i$  -th view; the Query  $Q$  comes directly from  $\\mathbf{F}^1$  and the concatenation of  $[\\mathbf{F}^1,F_f^l,F_b^l ]$  serves as the key  $K$  and the value  $V$  .  $\\oplus$  indicates matrix concatenation along the token axis;  $d$  denotes the dimension;  $W^{Q},W^{K},W^{V}$  represent the linear transformation matrices; we omitted the  $l$  of the attention matrices and parameters for simplicity;  $C\\in \\mathbb{R}^{m\\times m}$ $C_i$  represents  $i$  -th row in  $C$  , and its \"correlation\" value between  $i$  -th and  $j$  -th features is  $C_{ij}$  ..",
+        "bbox": [
+            511,
+            90,
+            903,
+            227
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nC _ {i j} = \\left(\\left(\\operatorname {t r a c e} \\left(R _ {i} ^ {T} R _ {j}\\right) - 1\\right) / 2 + 1\\right) / 2 \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            583,
+            239,
+            903,
+            258
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $R_{i}$  and  $R_{j}$  are the extrinsic rotation matrices of the corresponding camera views,  $(\\mathrm{trace}(R_i^T R_j) - 1) / 2$  is the cosine value of the angle between these camera views.",
+        "bbox": [
+            511,
+            268,
+            903,
+            314
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Multi-view CLIP Embedding. Camera viewpoints can serve as an effective condition signal to enhance 3D consistency in video content generation [52]. Building on this insight, we incorporate camera condition within our try-on network by encoding camera parameters as an additional token, enabling the generation of more consistent multi-view images. Specifically, we define a world coordinate system in which the camera faces the subject directly. For each input image (view)  $\\mathbf{A}_{\\mathrm{i}}$ ,  $1 \\leq i \\leq m$ , we extract the rotation matrix from the camera's corresponding extrinsic matrix. This rotation matrix is then reshaped into a 9-dimensional tensor  $\\mathbf{r}_{\\mathrm{i}}$ , which undergoes positional encoding to effectively integrate the camera parameters into the feature representation  $F_{i}^{c}$ .",
+        "bbox": [
+            511,
+            315,
+            905,
+            527
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} F _ {i} ^ {c} = \\left(\\sin \\left(2 ^ {0} \\pi \\mathbf {r} _ {\\mathbf {i}}\\right), \\cos \\left(2 ^ {0} \\pi \\mathbf {r} _ {\\mathbf {i}}\\right), \\dots , \\right. \\tag {4} \\\\ \\sin (2 ^ {L - 1} \\pi \\mathbf {r _ {i}}), \\cos (2 ^ {L - 1} \\pi \\mathbf {r _ {i}})) \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            594,
+            539,
+            903,
+            578
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $L$  is the length of positional embedding. We then project  $F_{i}^{c}$  to match the dimensionality of the garment CLIP image embedding  $F^{g}$  via an MLP and concatenate them along the token axis to form  $Y_{i}$ . This combined representation,  $Y_{i}$ , is subsequently used in the key  $K_{\\mathrm{x}}$  and value  $V_{\\mathrm{x}}$  of the cross-attention layers of the Main UNet:",
+        "bbox": [
+            511,
+            590,
+            905,
+            679
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nY _ {i} = F ^ {g} \\oplus \\operatorname {M L P} \\left(F _ {i} ^ {c}\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            635,
+            693,
+            777,
+            709
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nQ _ {\\mathrm {x}} = W _ {\\mathrm {x}} ^ {Q} \\mathbf {H} _ {\\mathrm {i}} ^ {1}, K _ {\\mathrm {x}} = W _ {\\mathrm {x}} ^ {K} Y _ {i}, V _ {\\mathrm {x}} = W _ {\\mathrm {x}} ^ {V} Y _ {i} \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            560,
+            712,
+            903,
+            736
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {i} ^ {(l + 1)} = \\mathrm {s o f t m a x} (\\frac {Q _ {\\mathrm {x}} K _ {\\mathrm {x}} ^ {T}}{\\sqrt {d _ {\\mathrm {x}}}}) V _ {\\mathrm {x}},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            586,
+            734,
+            794,
+            768
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{H}^1$  is the output of the MVAttention of the  $l$ -th layer; we omitted the  $l$  of the cross attention matrices and parameters for simplicity.",
+        "bbox": [
+            511,
+            779,
+            903,
+            825
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training. Our enhanced multi-view 2D VTON network can be trained by minimizing the following latent diffusion model loss function:",
+        "bbox": [
+            511,
+            825,
+            903,
+            869
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {l d m}} = \\mathbb {E} _ {z _ {t}, \\eta , \\psi , \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\eta , \\psi , \\zeta)) \\| _ {2} ^ {2} \\right], \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            553,
+            883,
+            903,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "26392",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/66efb9bdfdf378e1471d23c74712a4482299aa59ce3a5ebfeba941baa74bac58.jpg",
+        "image_caption": [
+            "Figure 4. Illustration of the proposed Multi-view Spatial Attention. Query (Q): multi-view features  $\\mathbf{F}^1$ ; Key (K) and Value (V): concatenation of  $\\mathbf{F}^1$  and garment features  $F_{f}^{l}$  and  $F_{b}^{l}$ . The attention score between viewpoints  $i$  and  $j$  is modulated by a weight  $C_{ij}$ , determined by the cosine of the angle between them."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            93,
+            88,
+            486,
+            294
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\eta = [\\mathcal{E}(g_f); \\mathcal{E}(g_b); \\mathcal{E}(\\mathbf{N_i})_{i=1}^m]$  represents the input latent garment images and latent normal maps;  $\\zeta = [\\mathcal{E}'(\\mathbf{A_i})_{i=1}^m]$  denotes the input latent clothing-agnostic images;  $\\psi = \\mathbf{Y}$  is the proposed multi-view CLIP embedding.",
+        "bbox": [
+            89,
+            380,
+            483,
+            443
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            455,
+            225,
+            474
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            89,
+            481,
+            282,
+            500
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Datasets. We conduct experiments on two public datasets: Thuman2.0 [55] and MVHumanNet [51]. Thuman2.0 comprises 526 reconstructed clothed human scans, from which we render multi-view input images. Of these samples, 426 are used for training, while the remaining 100 are set aside for testing. To further evaluate the effectiveness and robustness of our method, we also perform experiments on MVHumanNet, a large-scale dataset of multi-view human images that encompasses a diverse range of subjects, daily outfits and motion sequences. The images in MVHumanNet are captured using a multi-view system with either 48 or 24 cameras. We use 4,990 subjects from this dataset, allocating 4,790 to training and 200 for tests. For each subject, we randomly select two frames of multi-view images from its entire motion sequence. While MVHumanNet provides multi-view images directly for editing and reconstruction, we render uniformly distributed views around each human subject in Thuman2.0 to ensure consistent input.",
+        "bbox": [
+            88,
+            503,
+            482,
+            775
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Baselines. We primarily compare our method with three existing methods: DreamWaltz [23], GaussCtrl [48], and TIP-Editor [63]. DreamWaltz is a method designed for directly generating 3D human bodies based on textual descriptions, while GaussCtrl and TIP-Editor are two radiance-based editing methods. GaussCtrl is based on Stable Diffusion, using a description-like prompt to edit the scene. TIP-Editor accepts both text and image prompts. We configure it by",
+        "bbox": [
+            89,
+            780,
+            482,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "specifying the human body as the editing region and the desired garment as the image prompt. We use ChatGPT to generate the text prompts corresponding to the clothing images.",
+        "bbox": [
+            511,
+            90,
+            903,
+            151
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metrics. For quantitative evaluation, we assess garment-to-person alignment between the edited person and the reference image. Following [63], we calculate the average DINO similarity [34] between the reference image and the rendered multi-view images of the edited 3D scene. Additionally, to evaluate multi-view consistency, we compute the CLIP Directional Consistency Score as outlined in [17]. Given the large scale of experiments (repeated 3DGS reconstruction), we selected a subset of examples from the dataset for metric evaluation. Specifically, from the test sets of Thuman and MVHumanNet, we randomly sampled 10 human scans each, performing virtual try-on with 6 randomly chosen garments per human scan.",
+        "bbox": [
+            511,
+            155,
+            906,
+            351
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We further conducted a user study involving 50 participants who rated the results of our method and three baseline methods based on two criteria: overall \"Quality\" and \"Alignment\" with the reference image. Each evaluation consisted of two questions: (1) Which method produces the highest quality of the edited 3D human? and (2) Which method achieves the most consistent alignment with the target clothing? Participants viewed the VTON results as rotating randomized video sequences.",
+        "bbox": [
+            511,
+            351,
+            908,
+            488
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. During pre-processing, we crop the multi-view images to the bounding box around the person and resize them to a resolution of  $768 \\times 576$ . The front view and the back view of garment images are obtained from the corresponding clothed human images. After editing, we pad the images back to their original size. The data processing pipeline is the same for both Thuman2.0 and MVHumanNet datasets.",
+        "bbox": [
+            511,
+            491,
+            908,
+            611
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The Main UNet and the GarmentNet are initialized by the Stable Diffusion V1.5 [39]. The training process is divided into two stages. In the first stage, each view is trained independently, during which we establish the feature extraction capabilities of both the PoseEncoder and GarmentNet, as well as the generative capability of the Main UNet. The second stage involves multi-view training, where we randomly select  $M$  views for each human subject. This stage is focused on training the proposed MVAttention module to enhance the network's ability to maintain consistency across views. Due to memory constraints, we set  $M = 8$  for the training phase. During the testing phase, we uniformly sampled 32 views from a 360-degree rotation around the subject. The editing of these 32 views is conducted in two batches, with each batch processing  $M = 16$  views.",
+        "bbox": [
+            511,
+            612,
+            908,
+            839
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2. Comparisons with State-of-the-Art Methods",
+        "text_level": 1,
+        "bbox": [
+            511,
+            848,
+            888,
+            864
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Qualitative Evaluation. Fig. 5 shows visual comparisons between our method and the baselines. DreamWaltz [23]",
+        "bbox": [
+            511,
+            869,
+            906,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "26393",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/16b7e1376295cee170255b3f0d8c6c37631ea1766a2dacc28c5a68023f4e5152.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"4\">Thuman2.0 [55]</td><td colspan=\"4\">MVHumanNet [51]</td></tr><tr><td>CLIPcons↑</td><td>DINOsim↑</td><td>Votequality</td><td>Votealign</td><td>CLIPcons↑</td><td>DINOsim↑</td><td>Votequality</td><td>Votealign</td></tr><tr><td>DreamWaltz [23]</td><td>0.887</td><td>0.556</td><td>0.46%</td><td>1.54%</td><td>0.935</td><td>0.495</td><td>0.46%</td><td>0.46%</td></tr><tr><td>TIP-Editor [63]</td><td>0.939</td><td>0.569</td><td>0.92%</td><td>0.62%</td><td>0.948</td><td>0.512</td><td>2.15%</td><td>1.38%</td></tr><tr><td>GaussCtrl [48]</td><td>0.931</td><td>0.577</td><td>1.08%</td><td>1.38%</td><td>0.938</td><td>0.521</td><td>1.69%</td><td>1.23%</td></tr><tr><td>Ours</td><td>0.923</td><td>0.633</td><td>97.54%</td><td>96.46%</td><td>0.933</td><td>0.623</td><td>95.69%</td><td>96.92%</td></tr></table>",
+        "bbox": [
+            96,
+            88,
+            901,
+            180
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Quantitative comparisons.  ${\\mathrm{{CLIP}}}_{\\text{cons }}$  denotes the CLIP Direction Consistency Score.  ${\\mathrm{{DINO}}}_{\\text{sim }}$  is the DINO similarity.",
+        "bbox": [
+            114,
+            185,
+            879,
+            200
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/6b71985790ddfa2ccbcfe297ed7bf36c95d426c4edc7ae3f9d723d3ce1a7d5c3.jpg",
+        "image_caption": [
+            "Figure 5. Qualitative comparison. The first two rows show the results on Thuman2.0 dataset while the last two rows show the results on MVHumaNet dataset. Our method achieves good texture preservation (highlighted by the blue boxes), while three baseline methods mostly fail."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            133,
+            203,
+            867,
+            612
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "regenerates 3D clothed humans from text prompts but struggles to accurately retain both body and clothing characteristics. GaussCtrl [48], lacking support for image prompts, fails to maintain detailed clothing textures. While TipEditor [63] leverages Lora [21] for personalization, it encounters difficulties in consistently mapping clothing inputs from two views into the 3D human because the personalized concept are semantic in 2D space. In contrast, our method effectively preserves intricate clothing details, such as text, stripes, and logos.",
+        "bbox": [
+            88,
+            678,
+            480,
+            829
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Quantitative Evaluation. Tab. 1 shows the results for the CLIP Directional Consistency Score and DINO similarity on Thuman2.0 and MVHumanNet datasets. Our approach surpasses other methods on  $\\mathrm{DINO}_{sim}$ , clearly illustrating",
+        "bbox": [
+            88,
+            839,
+            482,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "the superiority of our method in terms of garment texture preservation. While our results on  $\\mathrm{CLIP}_{\\text {cons }}$  are comparable to those of other methods, it is important to note that these methods incorporate the SDS loss, which to some extent smoothes the representation of humans in 3D space. Additionally, the \"flatter\" textures of other methods could also result in artificially higher consistency scores. Furthermore, user studies have shown that our method significantly exceeds baselines in terms of edited 3D human quality and the alignment of clothing details.",
+        "bbox": [
+            511,
+            678,
+            906,
+            829
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3. Visual Results using E-commerce Garment",
+        "text_level": 1,
+        "bbox": [
+            511,
+            845,
+            879,
+            862
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Fig. 6 showcases VTON results using garments from the MVG dataset [46], whose images are from e-commerce",
+        "bbox": [
+            511,
+            869,
+            906,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "26394",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7737d81fbc34085b8a538160848be10370bc9d5255819ab87e19ccf3b3907f3f.jpg",
+        "image_caption": [
+            "Figure 6. Generalization to e-commerce garments (the MVG dataset). Our method, trained on the THuman2.0 dataset, shows strong generalizability when applied to e-commerce garments. For clarity in visualization, we display garment images on human models; however, in the actual VTON process, the garments are segmented from the models using parse maps."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            89,
+            480,
+            340
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "platforms like YOOX NET-A-PORTER, Taobao, and TikTok*, and a model trained on the Thuman2.0 dataset [55]. The results demonstrate that our method effectively preserves intricate garment details and textures. For instance, it accurately maintains the stripe patterns in the first row, the cute tie in the second row, and the buttons in the third row, highlighting the robustness of our approach in handling diverse and realistic clothing items.",
+        "bbox": [
+            89,
+            455,
+            483,
+            575
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            89,
+            590,
+            243,
+            606
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We conduct an ablation study on Thuman2.0 dataset in Tab. 2 and Fig. 7 to evaluate the impact of our three proposed modules in enhancing a typical 2D VTON network with 3D-consistent generation capabilities. Starting with the 2D VTON baseline [53] using DensePose, we progressively replace DensePose with our pseudo-3D pose, incorporate multi-view CLIP embeddings, and ultimately integrate MVAttention in the final configuration. Results in Tab. 2 indicate that each module contributes to metric improvements. Fig. 7 visualizes an example of multi-view image editing. The incorporation of pseudo-3D pose substantially improves limb generation compared to the 2D VTON baseline. Comparing rows 4 and 5, prior to the integration of multi-view CLIP embedding, the model captures limited spatial information, resulting in detail loss at specific angles (columns 3, 4, and 6). Finally, the proposed MVAttention achieves a more coherent generation across views.",
+        "bbox": [
+            89,
+            614,
+            483,
+            869
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5de364b5b01530170b9e9a96febaa852e03b779b915dc80ff95da76bb7a36c53.jpg",
+        "image_caption": [
+            "Figure 7. Visualization of the impact of the three proposed techniques on multi-view consistent editing. The red boxes highlight the artifacts. Starting from the 2D VTON baseline, the pseudo-3D pose improves limb generation quality, multi-view CLIP embedding enhances detail across different viewing directions, and finally, MVAttention further strengthens consistency in the generated images."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            535,
+            88,
+            883,
+            412
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/f705e5297b3e2f00053ee946d9e1e39d2ea67a67e62a277374d728081f02ef49.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>CLIPcons↑</td><td>DINOsim↑</td></tr><tr><td>2D-VTON</td><td>0.892</td><td>0.609</td></tr><tr><td>+ Pseudo-3D Pose</td><td>0.910</td><td>0.626</td></tr><tr><td>+ Multi-view CLIP Embedding</td><td>0.913</td><td>0.631</td></tr><tr><td>+ Multi-view Spatial Attention</td><td>0.923</td><td>0.633</td></tr></table>",
+        "bbox": [
+            517,
+            526,
+            898,
+            619
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Ablation studies. We ablate the impact of the three proposed techniques on Thuman2.0 dataset.",
+        "bbox": [
+            511,
+            623,
+            905,
+            651
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusions",
+        "text_level": 1,
+        "bbox": [
+            513,
+            672,
+            640,
+            689
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this work, we proposed VTON 360, a novel 3D Virtual Try-On (VTON) method that achieves high-fidelity VTON with the ability to render clothing from arbitrary viewing directions. Our method features a novel formulation of 3D VTON as an extension of 2D VTON that ensures 3D consistent results across multiple views. To bridge the gap between 2D VTON models and 3D consistency requirements, we introduce several key innovations, including multi-view inputs, pseudo-3D pose representation, multi-view spatial attention, and multi-view CLIP embedding. Extensive experiments demonstrate the effectiveness of our approach, significantly outperforming prior 3D VTON techniques in both fidelity and any-view rendering.",
+        "bbox": [
+            511,
+            703,
+            906,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_footnote",
+        "text": "* https://net-a-porter.com, www.taobao.com, www.douyin.com",
+        "bbox": [
+            109,
+            886,
+            439,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "26395",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            91,
+            90,
+            250,
+            107
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work is supported in part by the National Key R&D Program of China under Grant No.2024YFB3908503, in part by the National Natural Science Foundation of China under Grant NO. 62322608 and in part by the CCF-Kuaishou Large Model Explorer Fund (NO. CCF-KuaiShou 2024007).",
+        "bbox": [
+            89,
+            114,
+            483,
+            205
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            91,
+            220,
+            187,
+            237
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Easymocap - make human motion capture easier. Github, 2021.5",
+            "[2] Shuai Bai, Huiling Zhou, Zhikang Li, Chang Zhou, and Hongxia Yang. Single stage virtual try-on via deformable attention flows. In European Conference on Computer Vision, pages 409-425. Springer, 2022. 2",
+            "[3] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5420-5430, 2019. 2, 3",
+            "[4] Robert Bridson, Ronald Fedkiw, and John Anderson. Robust treatment of collisions, contact and friction for cloth animation. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 594-603, 2002. 3",
+            "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3",
+            "[6] Haodong Chen, Yongle Huang, Haojian Huang, Xiangsheng Ge, and Dian Shao. Gaussianvton: 3d human virtual try-on via multi-stage gaussian splatting editing with image prompting. arXiv preprint arXiv:2405.07472, 2024. 2, 3, 1",
+            "[7] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024. 3",
+            "[8] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14131-14140, 2021. 1, 2",
+            "[9] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. arXiv preprint arXiv:2403.05139, 2024. 3, 4, 5",
+            "[10] Xin Dong, Fuwei Zhao, Zhenyu Xie, Xijin Zhang, Daniel K Du, Min Zheng, Xiang Long, Xiaodan Liang, and Jianchao Yang. Dressing in the wild by watching dance videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3480-3489, 2022. 2"
+        ],
+        "bbox": [
+            93,
+            247,
+            483,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[11] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8485-8493, 2021. 2",
+            "[12] Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang. Taming the power of diffusion models for high-quality virtual try-on with appearance flow. arXiv preprint arXiv:2308.06101, 2023. 1, 2, 4",
+            "[13] Peng Guan, Loretta Reiss, David A Hirshberg, Alexander Weiss, and Michael J Black. Drape: Dressing any person. ACM Transactions on Graphics (ToG), 31(4):1-10, 2012. 1, 3",
+            "[14] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densepose: Dense human pose estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7297-7306, 2018. 4, 5",
+            "[15] Fabian Hahn, Bernhard Thomaszewski, Stelian Coros, Robert W Sumner, Forrester Cole, Mark Meyer, Tony DeRose, and Markus Gross. Subspace clothing simulation using adaptive bases. ACM Transactions on Graphics (TOG), 33(4):1-9, 2014. 1, 3",
+            "[16] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 2",
+            "[17] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19740–19750, 2023. 3, 6, 1",
+            "[18] Sen He, Yi-Zhe Song, and Tao Xiang. Style-based global appearance flow for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3470-3479, 2022. 2",
+            "[19] Zijian He, Peixin Chen, Guangrun Wang, Guanbin Li, Philip HS Torr, and Liang Lin. Wildvidfit: Video virtual try-on in the wild via image-based controlled diffusion models. In European Conference on Computer Vision, pages 123-139. Springer, 2024. 1, 2",
+            "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3",
+            "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 7",
+            "[22] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023. 5",
+            "[23] Yukun Huang, Jianan Wang, Ailing Zeng, He Cao, Xianbiao Qi, Yukai Shi, Zheng-Jun Zha, and Lei Zhang. Dreamwaltz: Make a scene with complex 3d animatable avatars. Advances in Neural Information Processing Systems, 36, 2024. 3, 6, 7",
+            "[24] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "26396",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "reconstruction of lifelike clothed humans. In 2024 International Conference on 3D Vision (3DV), pages 1531-1542. IEEE, 2024. 3",
+            "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 3, 1",
+            "[26] Jeongho Kim, Gyojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. arXiv preprint arXiv:2312.01725, 2023. 3, 4, 5",
+            "[27] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3",
+            "[28] Zorah Lahner, Daniel Cremers, and Tony Tung. Deepwrinkles: Accurate and realistic clothing modeling. In Proceedings of the European conference on computer vision (ECCV), pages 667-684, 2018. 1, 3",
+            "[29] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In Proceedings of the European conference on computer vision (ECCV), 2022. 2",
+            "[30] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851-866. 2023. 3",
+            "[31] Yifang Men, Yiming Mao, Yuning Jiang, Wei-Ying Ma, and Zhouhui Lian. Controllable person image synthesis with attribute-decomposed gan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5084-5093, 2020. 2",
+            "[32] Aymen Mir, Thiemo Alldieck, and Gerard Pons-Moll. Learning to transfer texture from clothing images to 3d humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7023-7034, 2020. 2, 3",
+            "[33] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. Ladi-vton: Latent diffusion textual-inversion enhanced virtual try-on. arXiv preprint arXiv:2305.13501, 2023. 1, 2",
+            "[34] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6",
+            "[35] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 5",
+            "[36] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael J Black. Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics (ToG), 36(4):1-15, 2017. 3",
+            "[37] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3"
+        ],
+        "bbox": [
+            91,
+            90,
+            482,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[38] Yurui Ren, Xiaoqing Fan, Ge Li, Shan Liu, and Thomas H Li. Neural texture extraction and distribution for controllable person image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13535-13544, 2022. 2",
+            "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 3, 4, 6",
+            "[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3",
+            "[41] Igor Santesteban, Nils Thuerey, Miguel A Otaduy, and Dan Casas. Self-supervised collision handling via generative 3d garment models for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11763-11773, 2021. 2",
+            "[42] Igor Santesteban, Miguel Otaduy, Nils Thuerey, and Dan Casas. Ulnef: Untangled layered neural fields for mix-and-match virtual try-on. Advances in Neural Information Processing Systems, 35:12110-12125, 2022. 2",
+            "[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3",
+            "[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3",
+            "[45] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 3",
+            "[46] Haoyu Wang, Zhilu Zhang, Donglin Di, Shiliang Zhang, and Wangmeng Zuo. Mv-vton: Multi-view virtual try-on with diffusion models. arXiv preprint arXiv:2404.17364, 2024. 7",
+            "[47] Junjie Wang, Jiemin Fang, Xiaopeng Zhang, Lingxi Xie, and Qi Tian. Gaussianeditor: Editing 3d gaussians delicately with text instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20902-20911, 2024. 3",
+            "[48] Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. Gaussctrl: multi-view consistent text-driven 3d gaussian splattering editing. arXiv preprint arXiv:2403.08733, 2024. 3, 6, 7",
+            "[49] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 5",
+            "[50] Zhenyu Xie, Haoye Dong, Yufei Gao, Zehua Ma, and Xiaodan Liang. Dreamvton: Customizing 3d virtual try-on with personalized diffusion models. arXiv preprint arXiv:2407.16511, 2024. 2, 3",
+            "[51] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu,"
+        ],
+        "bbox": [
+            516,
+            90,
+            906,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "26397",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Chongjie Wang, Shijie Wang, et al. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19801-19811, 2024. 2, 6, 7",
+            "[52] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. 5",
+            "[53] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. arXiv preprint arXiv:2403.01779, 2024. 3, 4, 8",
+            "[54] Han Yang, Xinrui Yu, and Ziwei Liu. Full-range virtual try-on with recurrent tri-level transform. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3460-3469, 2022. 2",
+            "[55] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5746-5756, 2021. 2, 6, 7, 8",
+            "[56] Jinsong Zhang, Kun Li, Yu-Kun Lai, and Jingyu Yang. Pise: Person image synthesis and editing with decoupled gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7982-7990, 2021. 2",
+            "[57] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 3",
+            "[58] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 5",
+            "[59] Fuwei Zhao, Zhenyu Xie, Michael Kampffmeyer, Haoye Dong, Songfang Han, Tianxiang Zheng, Tao Zhang, and Xiaodan Liang. M3d-vton: A monocular-to-3d virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13239-13249, 2021. 3",
+            "[60] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4606-4615, 2023. 2",
+            "[61] Jingyu Zhuang, Chen Wang, Liang Lin, Lingjie Liu, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. In SIGGRAPH Asia 2023 Conference Papers, pages 1-10, 2023. 3",
+            "[62] Jingyu Zhuang, Di Kang, Linchao Bao, Liang Lin, and Guanbin Li. Dagensm: Disentangled avatar generation with gs-enhanced mesh. arXiv preprint arXiv:2411.15205, 2024.3",
+            "[63] Jingyu Zhuang, Di Kang, Yan-Pei Cao, Guanbin Li, Liang Lin, and Ying Shan. Tip-editor: An accurate 3d editor following both text-prompts and image-prompts. arXiv preprint arXiv:2401.14828, 2024. 3, 6, 7, 1"
+        ],
+        "bbox": [
+            91,
+            90,
+            482,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "26398",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2025/VTON 360_ High-Fidelity Virtual Try-On from Any Viewing Direction/4af7cfba-2431-4854-85f1-f3f911dcc811_model.json

@@ -0,0 +1,2103 @@
+[
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.107,
+                0.003,
+                0.182,
+                0.043
+            ],
+            "angle": 0,
+            "content": "CVF"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.237,
+                0.001,
+                0.808,
+                0.047
+            ],
+            "angle": 0,
+            "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.143,
+                0.131,
+                0.854,
+                0.153
+            ],
+            "angle": 0,
+            "content": "VTON 360: High-Fidelity Virtual Try-On from Any Viewing Direction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.1,
+                0.16,
+                0.9,
+                0.251
+            ],
+            "angle": 0,
+            "content": "Zijian He<sup>1</sup> Yuwei Ning<sup>1</sup> Yipeng Qin<sup>2</sup> Guangrun Wang<sup>1</sup> Sibei Yang<sup>3</sup> Liang Lin<sup>1,4,5</sup> Guanbin Li<sup>1,4,5*</sup>  \n<sup>1</sup>Sun Yat-sen University  \n<sup>2</sup>Cardiff University  \n<sup>3</sup>ShanghaiTech University  \n<sup>4</sup>Guangdong Key Laboratory of Big Data Analysis and Processing  \n<sup>5</sup>Peng Cheng Laboratory  \nhezj39@mail2.sysu.edu.cn, yuwei.ning@hust.edu.cn, {qinyipengl991, wanggrun}@gmail.com, yangsb@shanghaiitech.edu.cn linliang@ieee.org, liguanbin@mail.sysu.edu.cn"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.092,
+                0.253,
+                0.511,
+                0.475
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.518,
+                0.253,
+                0.71,
+                0.476
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.715,
+                0.253,
+                0.907,
+                0.476
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.089,
+                0.481,
+                0.908,
+                0.54
+            ],
+            "angle": 0,
+            "content": "Figure 1. Results of VTON 360. Our VTON 360 enables high-fidelity 3D Virtual Try-On (VTON) by seamlessly adapting E-commerce garments onto a clothed 3D human model, supporting full \\(360^{\\circ}\\) view rendering. The highlighted bounding boxes (dashed line) demonstrate our method's ability to preserve intricate clothing details and patterns (e.g., collar accessories, horizontal line patterns, logos, texts, numbers) across diverse garment types."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.248,
+                0.544,
+                0.327,
+                0.559
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.57,
+                0.486,
+                0.873
+            ],
+            "angle": 0,
+            "content": "Virtual Try-On (VTON) is a transformative technology in e-commerce and fashion design, enabling realistic digital visualization of clothing on individuals. In this work, we propose VTON 360, a novel 3D VTON method that addresses the open challenge of achieving high-fidelity VTON that supports any-view rendering. Specifically, we leverage the equivalence between a 3D model and its rendered multiview 2D images, and reformulate 3D VTON as an extension of 2D VTON that ensures 3D consistent results across multiple views. To achieve this, we extend 2D VTON models to include multi-view garments and clothing-agnostic human body images as input, and propose several novel techniques to enhance them, including: i) a pseudo-3D pose representation using normal maps derived from the SMPL-X 3D human model, ii) a multi-view spatial attention mechanism that models the correlations between features from different viewing angles, and iii) a multi-view CLIP embedding that enhances the garment CLIP features used in 2D VTON with camera information. Extensive experiments on large-scale real datasets and clothing images from e-commerce"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.545,
+                0.907,
+                0.59
+            ],
+            "angle": 0,
+            "content": "platforms demonstrate the effectiveness of our approach.  \nProject page: https://scnuhealthy.github.io/  \nVTON360."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.602,
+                0.645,
+                0.617
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.511,
+                0.628,
+                0.907,
+                0.749
+            ],
+            "angle": 0,
+            "content": "Virtual Try-On (VTON) enables realistic digital visualization of clothing on individuals and has emerged as a transformative technology in e-commerce and fashion design. While significant research efforts have been made on 2D VTON solutions [8, 12, 19, 33, 39], these approaches are inherently limited in their representation of view-related features. To overcome this limitation and enable high-fidelity any-view rendering, 3D VTON methods were introduced."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.511,
+                0.75,
+                0.909,
+                0.902
+            ],
+            "angle": 0,
+            "content": "3D VTON requires accurate garment transfer onto a 3D human body while ensuring realistic garment fitting, texture preservation, and 3D consistency. The two primary aims of 3D VTON are i) achieving high-fidelity and ii) supporting any-view rendering. Leveraging the inherent capability of 3D models for any-view rendering, early 3D VTON methods [13, 15, 28] make clothing simulation on synthetic human bodies. Specifically, these methods utilized 3D scanners to capture clothing meshes, followed by the development of specialized dressing algorithms. Although effec"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.11,
+                0.887,
+                0.312,
+                0.9
+            ],
+            "angle": 0,
+            "content": "* Corresponding author is Guanbin Li."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "26388"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.092,
+                0.482,
+                0.304
+            ],
+            "angle": 0,
+            "content": "tive, these methods rely on costly 3D scanning equipment and the physical presence of the human body/clothing (i.e., not fully virtual), restricting their practicality in real-world applications. As a byproduct, most early methods focused on developing geometrically correct dressing algorithms using standard templates of human body and clothing models. Addressing this limitation, researchers extended 3D VTON by introducing algorithms that reconstruct 3D clothing models from input images, enabling the use of image-based clothing inputs [3, 32, 41, 42]. However, since input clothing images (usually frontal) are inherently 2D and lack multi-view information, this approach struggles to reconstruct high-fidelity clothing models that can be rendered well from all viewing directions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.093,
+                0.308,
+                0.483,
+                0.716
+            ],
+            "angle": 0,
+            "content": "To complement this missing information, DreamVTON [50] introduces a novel approach that leverages Text-to-Image (T2I) diffusion models to reconstruct both the human body and clothing from input images. Its key insight is that T2I models learned view-agnostic \"concepts\" of both bodies and garments during their training, and that the corresponding concepts for the input body and clothing images can be obtained using LoRA [21]. By utilizing Score Distillation Sampling (SDS) [37], DreamVTON can generate visual-pleasing 3D VTON results by ensuring consistency between renderings from arbitrary viewpoints and the concepts. Nonetheless, DreamVTON's high flexibility comes at the cost of low fidelity. This limitation stems from the fact that the concepts learned by T2I models are semantic in nature, thus lacking 3D geometric consistency and pixel-level accuracy with respect to the input body and clothing images. Recently, a concurrent work, namely GaussianVTON [6], partially addressed this limitation by formulating 3D VTON as a 3D scene editing task, where a given 3D human model is edited using multi-view images generated by 2D VTON methods. While it significantly enhances the fidelity of the human body, the fidelity and 3D consistency of clothing remain problematic, as there are no 2D VTON methods that can generate multi-view images with 3D consistency. Therefore, to the best of our knowledge, achieving high-fidelity 3D VTON that supports any-view rendering remains an open challenge."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.72,
+                0.483,
+                0.902
+            ],
+            "angle": 0,
+            "content": "In this work, we address the above-mentioned challenge via proposing VTON 360, a novel 3D VTON method that achieves high-fidelity VTON from arbitrary viewing directions. Similar to GaussianVTON [6], our method edits a given 3D human model by inpainting the rendered images using a latent diffusion model. However, we set ourselves apart through our novel garment fidelity preservation strategy that can generate high-fidelity on-body garments in all viewing directions. Specifically, we first extend both the garment and clothing-agnostic human body inputs to typical 2D VTON models to leverage multi-view information, including paired front and back view garment images as well"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.092,
+                0.905,
+                0.483
+            ],
+            "angle": 0,
+            "content": "as a set of multi-view clothing-agnostic human body images sampled from random azimuth angles. Then, we propose several novel enhancements to bridge the gap between typical 2D VTON methods and our multi-view 3D consistency requirements: i) We propose a pseudo-3D pose representation using normal maps derived from the SMPL-X 3D human model, which captures fine-grained surface orientation details and provides more consistent geometry across views compared to the 2D pose representations (semantic segmentation maps) used in 2D VTON models. ii) We design a Multi-view Spatial Attention mechanism that models the correlations between features from different viewing angles, featuring a novel \"correlation\" matrix modeling the relationships among different input views. iii) We propose a multi-view CLIP embedding that enhances the garment CLIP embedding used in 2D VTON methods with camera information, thereby facilitating network learning of features relevant to a particular view. Together, these innovations enable our 2D VTON model to generate high-quality, multi-view and 3D-consistent virtual try-on results. Extensive experiments on Thuman2.0 [55] and MVHumanNet [51] datasets demonstrate that our method achieves high fidelity 3D VTON which supports any-view rendering. In addition, we show the effectiveness and generalizability of our methodology by testing it using garments from e-commerce platforms. Our conclusions include:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.484,
+                0.905,
+                0.529
+            ],
+            "angle": 0,
+            "content": "- We propose a novel 3D Virtual Try-On (VTON) method, namely VTON 360, which achieves high-fidelity VTON from arbitrary viewing directions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.53,
+                0.905,
+                0.664
+            ],
+            "angle": 0,
+            "content": "- Leveraging the equivalence between a 3D model and its rendered multi-view 2D images, we reformulate 3D VTON as an extension of 2D VTON that ensures consistent results across multiple views. Specifically, we introduce several novel techniques, including: (i) pseudo-3D pose representation; (ii) multi-view spatial attention; and (iii) multi-view CLIP embedding. These innovations enhance traditional 2D VTON models to generate multiview and 3D-consistent results."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.665,
+                0.904,
+                0.711
+            ],
+            "angle": 0,
+            "content": "- Extensive experimental results on two large real datasets as well as real clothing images from e-commerce platforms demonstrate the effectiveness of our approach."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.514,
+                0.484,
+                0.905,
+                0.711
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.724,
+                0.655,
+                0.74
+            ],
+            "angle": 0,
+            "content": "2. Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.75,
+                0.905,
+                0.901
+            ],
+            "angle": 0,
+            "content": "2D Virtual Try-On. 2D Virtual Try-On (VTON) aims to transfer a desired garment to the corresponding region of a target human image while preserving the human pose and identity. Early methods [2, 8, 10, 11, 16, 18, 29, 31, 38, 54, 56] use Generative Adversarial Networks (GANs) to deform the garments to match the target body shape, which a critical step for achieving realistic VTON. However, accurately adapting to diverse real-world conditions remains a significant challenge. Addressing this issue, recent methods [12, 19, 33, 60] reframe 2D VTON as a conditioned in-"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "26389"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.092,
+                0.482,
+                0.168
+            ],
+            "angle": 0,
+            "content": "painting task, leveraging the strong priors provided by diffusion models [20, 43, 45] to achieve promising results. This strategy is further improved by [9, 26, 53], which introduce a ReferenceNet to extract hierarchical garment features and apply attention mechanisms to condition the Main UNet."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.093,
+                0.179,
+                0.485,
+                0.709
+            ],
+            "angle": 0,
+            "content": "3D Virutal Try-On. For 3D Virtual Try-On (VTON), traditional methods [4, 13, 15, 28, 36] rely on 3D scanning or cloth simulation to generate highly precise body and garment geometry. These methods were then extended by learning-based methods [3, 32] that employ differentiable rendering to dress the SMPL [30] model with a desired garment mesh. Despite their effectiveness, such methods rely on costly 3D scanning and the physical presence of human body/clothing, limiting their application in the real world. Addressing this limitation, M3D-VTON [59] proposes a depth-based 3D VTON framework to reconstruct 3D clothed human models from 2D human and garment images, but the results often suffer from explicit warping artifacts. To improve 3D VTON results, recent methods [23, 24, 50, 62] resort to text-to-image (T2I) diffusion models and employ the Score Distillation Sampling (SDS) loss [40] to ensure consistency among different viewing directions. Specifically, TeCH [24] adapts the generative priors of T2I diffusion model to the specific person and clothes by training descriptive text prompts with DreamBooth [40]. DreamWaltz [23] leverages Pose ControlNet [57] to attain clothed human body models. DreamVTON [50] introduces a multi-concept LoRA [21] to personalize the T2I diffusion model, and uses a template-based optimization mechanism that combines with SDS loss to better preserve patterns on the garment. Although effective, these methods often produce results lacking in fidelity, as the concepts learned by T2I models are semantic rather than at the pixel level. Concurrent to our work, GaussianVTON [6] proposes an alternative approach by combining Gaussian Splatting [25] with pre-trained 2D VTON models and formulate it as an editing task. However, since there are no 2D VTON methods that can generate multi-view images with 3D consistency, the fidelity and 3D consistency of the clothing generated remain problematic."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.72,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Radiance Field-based 3D Human or Scene Editing. Recently, radiance field-based editing has attracted significant interest due to its efficient differentiable rendering capabilities, sparking substantial advancements in text-driven 3D editing. For example, InstructN2N [17] employ an image-based diffusion model InstructP2P [5] to modify the rendered image by the user's text description with a variant of the score distillation sampling (SDS) [37] loss. GaussianEditor [7] applies Gaussian Splatting [25] as 3D representation instead of NeRF, adopting Gaussian semantic tracking to track target Gaussian values, significantly improving editing speed and controllability. To enable accu"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.092,
+                0.908,
+                0.289
+            ],
+            "angle": 0,
+            "content": "rate location and appearance control, subsequent works [47, 61] specify the editing region using the attention score or with a segmentation model. TIP-Editor [63] proposes a content personalization step dedicated to the reference image based on LoRA, achieving the editing following a hybrid text-image prompt. GaussCtrl [48] leverage depth conditions and attention-based latent code alignment to achieve 3D-aware multi-view consistent editing instead of iteratively editing single views using SDS loss. However, these works primarily focus on global appearance modifications within a text-driven pipeline, while our approach emphasizes preserving fine textural details from different viewing directions throughout the editing process."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.301,
+                0.64,
+                0.318
+            ],
+            "angle": 0,
+            "content": "3. Preliminary"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.326,
+                0.907,
+                0.448
+            ],
+            "angle": 0,
+            "content": "Latent Diffusion Model. Latent Diffusion Model [39] is a variant of diffusion models that performs denoising within the latent space of a Variational Autoencoder (VAE) [27]. During training, given a fixed encoder \\(\\mathcal{E}\\), an input image \\(x\\) is transformed into its latent representation \\(z_0 = \\mathcal{E}(x)\\). A conditional diffusion model \\(\\hat{\\epsilon}_{\\theta}\\), typically implemented with a UNet architecture, is then trained using a weighted denoising score matching objective:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.584,
+                0.457,
+                0.907,
+                0.476
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {L D M} = \\mathbb {E} _ {\\mathbf {z}, \\mathbf {c}, \\boldsymbol {\\epsilon}, t} [ \\boldsymbol {\\epsilon} - \\| \\hat {\\epsilon} _ {\\theta} (\\mathbf {z} _ {\\mathbf {t}}; \\mathbf {c}, t) \\| _ {2} ^ {2} ] \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.485,
+                0.909,
+                0.592
+            ],
+            "angle": 0,
+            "content": "where \\(\\mathbf{z}_t\\coloneqq \\alpha_t\\mathbf{x} + \\sigma_t\\boldsymbol {\\epsilon}\\) denotes the forward diffusion process at timestep \\(t\\) . \\(\\alpha_{t},\\sigma_{t}\\) are time-dependent functions defined by the diffusion model formulation; c denotes the conditional input and \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{1})\\) is Gaussian noise. During inference, data samples are generated by initiating from Gaussian noise \\(\\mathbf{z}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{1})\\) and iteratively refining it using a DDIM [44] sampler."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.604,
+                0.605,
+                0.619
+            ],
+            "angle": 0,
+            "content": "4. Method"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.629,
+                0.909,
+                0.825
+            ],
+            "angle": 0,
+            "content": "Our method leverages the equivalence between a 3D model and its rendered multi-view 2D images to achieve high-fidelity, any-view 3D VTON. Specifically, as Fig. 2 shows, given an input 3D human model and a garment image, our method 1) renders the 3D model into multi-view 2D images and 2) formulates 3D VTON as a consistent, unified 2D VTON process across these rendered views; 3) By reconstructing the edited images into a 3D model using existing 3D reconstruction methods, we ensure visual coherence and precise garment alignment from any viewing angle. Among them, the second step is crucial as existing 2D VTON methods [9, 26, 53] lack 3D knowledge, preventing them from generating multi-view images with 3D consistency."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.826,
+                0.908,
+                0.903
+            ],
+            "angle": 0,
+            "content": "To address this challenge, we propose several novel techniques (Sec. 4.2) that equip a typical 2D VTON network (Sec. 4.1), which is built on a latent diffusion model [39], with the capability to generate 3D-consistent results. We use Gaussian Splatting [25] as our 3D representation."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "26390"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.091,
+                0.089,
+                0.905,
+                0.433
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.09,
+                0.442,
+                0.908,
+                0.498
+            ],
+            "angle": 0,
+            "content": "Figure 2. Overview of VTON 360. Given an input 3D human model \\(\\mathbf{G}_{\\mathrm{src}}\\) and a pair of garment images \\((g_{f}, g_{b})\\), our method 1) renders \\(\\mathbf{G}_{\\mathrm{src}}\\) into multi-view 2D images (left) and 2) edits the rendered multi-view 2D images (middle); 3) reconstructs the edited images into a 3D model \\(\\mathbf{G}_{\\mathrm{VTON}}\\) (right). In the crucial step 2), we propose three novel techniques to equip a typical 2D VTON network with the capability to generate 3D-consistent results: 1) Pseudo-3D Pose Input, 2) Multi-view Spatial Attention, and 3) Multi-view CLIP Embedding."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.093,
+                0.504,
+                0.482,
+                0.68
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.09,
+                0.687,
+                0.483,
+                0.771
+            ],
+            "angle": 0,
+            "content": "Figure 3. DensePose (2D) vs. SMPL-X normal map (pseudo-3D) representations. DensePose applies uniform labels per body part, lacking 3D consistency across views and causing artifacts and temporal inconsistencies (highlighted with red boxes). In contrast, SMPL-X normal maps capture fine surface details, ensuring geometric coherence and stable, realistic shading across views."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.785,
+                0.373,
+                0.801
+            ],
+            "angle": 0,
+            "content": "4.1. Recap of 2D VTON Framework"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.81,
+                0.484,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Following previous works [12, 26, 53], we formulate 2D VTON as an exemplar-based inpainting problem, aiming to fill an input clothing-agnostic image \\(\\mathbf{A}\\) with a given garment image \\(g\\), where \\(\\mathbf{A}\\) is obtained following the method used in [53]. As illustrated in Fig. 2 (middle), the network architecture is based on the latent diffusion model [39] with"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.507,
+                0.818,
+                0.521
+            ],
+            "angle": 0,
+            "content": "an encoder \\(\\mathcal{E}\\) and comprises two components:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.521,
+                0.89,
+                0.536
+            ],
+            "angle": 0,
+            "content": "- A GarmentNet [9, 53] that extracts features from \\(\\mathcal{E}(g)\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.536,
+                0.905,
+                0.641
+            ],
+            "angle": 0,
+            "content": "- A Main UNet that inpaints \\(\\mathbf{A}\\) according to i) detailed garment features extracted by the GarmentNet; ii) the 2D pose of \\(\\mathbf{A}\\) represented by semantic labels using Dense-Pose [14]; iii) CLIP embeddings of input garment \\(g\\). Among them, i) and ii) together with noise are input to the self-attention layers, while iii) is input to the cross-attention layers of the Main UNet."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.513,
+                0.521,
+                0.905,
+                0.641
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.642,
+                0.905,
+                0.671
+            ],
+            "angle": 0,
+            "content": "Both the GarmentNet and the Main UNet share the same network architecture."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.682,
+                0.878,
+                0.699
+            ],
+            "angle": 0,
+            "content": "4.2. Multi-view 2D VTON with 3D Consistency"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.705,
+                0.905,
+                0.75
+            ],
+            "angle": 0,
+            "content": "To enable the aforementioned 2D VTON model to generate multi-view and 3D-consistent results, we propose the following novel enhancements to its design:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.75,
+                0.888,
+                0.765
+            ],
+            "angle": 0,
+            "content": "Multi-view Inputs. We extend both inputs to the model:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.515,
+                0.766,
+                0.906,
+                0.886
+            ],
+            "angle": 0,
+            "content": "- Multi-view Garment Inputs: We extend the input garment representation from a single image \\( g \\) to paired front and back view images \\( g_{f} \\), \\( g_{b} \\), providing comprehensive garment information across all viewing angles. Accordingly, we use the encoder \\( \\mathcal{E} \\) to encode \\( g_{f}, g_{b} \\) into their latent representations \\( \\mathcal{E}(g_{f}), \\mathcal{E}(g_{b}) \\) and feed them into GarmentNet to obtain their multi-layer features \\( F_{f}^{l} \\) and \\( F_{b}^{l} \\) at layer \\( l \\), respectively."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.886,
+                0.905,
+                0.902
+            ],
+            "angle": 0,
+            "content": "- Multi-view Clothing-agnostic Image Inputs: Based on"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.514,
+                0.766,
+                0.906,
+                0.902
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.518,
+                0.957
+            ],
+            "angle": 0,
+            "content": "26391"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.104,
+                0.092,
+                0.483,
+                0.211
+            ],
+            "angle": 0,
+            "content": "the equivalence between a 3D human model and its rendered multi-view 2D images, we extend the input human body representation from a single, clothing-agnostic image, \\(\\mathbf{A}\\), to a set of \\(m\\) multi-view images, denoted as \\(\\mathbf{A}_1, \\mathbf{A}_2, \\dots, \\mathbf{A}_m\\). These images are sampled from random azimuth angles, allowing the 2D VTON model to access comprehensive, multi-view information from the input 3D human model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.213,
+                0.483,
+                0.47
+            ],
+            "angle": 0,
+            "content": "Pseudo-3D Pose Input. As shown in Fig. 3, the 2D Dense-Pose representations [14] commonly used in state-of-the-art 2D VTON methods [9, 26] assign a uniform semantic label to all pixels within each body part (e.g., thigh), inherently lack 3D geometric consistency across multiple views, and often introduce artifacts and temporal inconsistencies. To address these limitations, we propose a novel pseudo-3D pose representation: the normal maps \\(\\mathbf{N}\\) derived from the SMPL-X [35] model of the input body. These normal maps capture fine-grained surface orientation details, providing a more consistent representation across views by preserving geometric structure in the 3D space. Furthermore, they enable smoother, temporally stable transitions and realistic shading effects across multi-view scenarios. In practice, we employ a lightweight PoseEncoder \\(\\mathcal{E}'\\) [22] and feed \\(\\mathcal{E}'(\\mathbf{N})\\) into the Main UNet. We obtain the SMPL-X model from the multi-view images of the input body using EasyMoCap [1]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.471,
+                0.483,
+                0.59
+            ],
+            "angle": 0,
+            "content": "Accordingly, we concatenate three components as the enhanced input to the Main UNet: i) a noise latent \\( z_{t} \\); ii) the encoded pseudo-3D poses \\( \\mathcal{E}'(\\mathbf{N_1}),\\mathcal{E}'(\\mathbf{N_2}),\\dots,\\mathcal{E}'(\\mathbf{N_m}) \\); and iii) the encoded multi-view clothing-agnostic images \\( \\mathcal{E}(\\mathbf{A_1}),\\mathcal{E}(\\mathbf{A_2}),\\dots,\\mathcal{E}(\\mathbf{A_m}) \\). Let \\( F_{1}^{l},F_{2}^{l},\\dots,F_{m}^{l} \\) be the feature representations at layer \\( l \\) of the Main UNet, and recall the garment features \\( F_{f}^{l} \\) and \\( F_{b}^{l} \\) defined above, we enhance the self-attention layers of the Main UNet as:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.592,
+                0.483,
+                0.818
+            ],
+            "angle": 0,
+            "content": "Multi-view Spatial Attention. To cope with the aforementioned multi-view input features and ensure their consistency, we draw insights from the temporal attention layer commonly used in video generation and editing [49, 58] and extend it to our multi-view spatial attention layer, denoted as MVAttention. The key distinction of our MVAttention is that its input multi-view features \\( F_{1}^{l}, F_{2}^{l}, \\ldots, F_{m}^{l} \\) are from images captured from non-uniform spatial intervals, with the viewing angles varying randomly. Consequently, features from similar views exhibit a higher correlation, while those from distinct views are largely independent. To model this relationship, we construct a \"correlation\" matrix \\( C \\) based on the angular disparity obtained from camera rotation matrices of the input multi-view images, and define our MVAttention as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.113,
+                0.827,
+                0.482,
+                0.906
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} \\mathbf {F} ^ {\\mathbf {l}} = \\left[ F _ {1} ^ {l} \\oplus F _ {2} ^ {l} \\dots \\oplus F _ {m} ^ {l} \\right], \\hat {\\mathbf {F}} ^ {\\mathbf {l}} = \\left[ \\mathbf {F} ^ {\\mathbf {l}} \\oplus F _ {f} ^ {l} \\oplus F _ {b} ^ {l} \\right] \\\\ Q = W ^ {Q} \\mathbf {F} ^ {1}, K = W ^ {K} \\hat {\\mathbf {F}} ^ {1}, V = W ^ {V} \\hat {\\mathbf {F}} ^ {1} \\tag {2} \\\\ A _ {i} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {i} \\times \\left(C _ {i} \\cdot K ^ {T}\\right)}{\\sqrt {d}}\\right), \\mathbf {H} _ {\\mathbf {i}} ^ {\\mathbf {l}} = A _ {i} \\times V _ {i} \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.092,
+                0.905,
+                0.228
+            ],
+            "angle": 0,
+            "content": "where \\(i\\in \\{1,2,\\dots,m\\}\\) denotes \\(i\\) -th view; the Query \\(Q\\) comes directly from \\(\\mathbf{F}^1\\) and the concatenation of \\([\\mathbf{F}^1,F_f^l,F_b^l ]\\) serves as the key \\(K\\) and the value \\(V\\) . \\(\\oplus\\) indicates matrix concatenation along the token axis; \\(d\\) denotes the dimension; \\(W^{Q},W^{K},W^{V}\\) represent the linear transformation matrices; we omitted the \\(l\\) of the attention matrices and parameters for simplicity; \\(C\\in \\mathbb{R}^{m\\times m}\\) \\(C_i\\) represents \\(i\\) -th row in \\(C\\) , and its \"correlation\" value between \\(i\\) -th and \\(j\\) -th features is \\(C_{ij}\\) .."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.584,
+                0.24,
+                0.905,
+                0.259
+            ],
+            "angle": 0,
+            "content": "\\[\nC _ {i j} = \\left(\\left(\\operatorname {t r a c e} \\left(R _ {i} ^ {T} R _ {j}\\right) - 1\\right) / 2 + 1\\right) / 2 \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.27,
+                0.905,
+                0.315
+            ],
+            "angle": 0,
+            "content": "where \\(R_{i}\\) and \\(R_{j}\\) are the extrinsic rotation matrices of the corresponding camera views, \\((\\mathrm{trace}(R_i^T R_j) - 1) / 2\\) is the cosine value of the angle between these camera views."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.316,
+                0.906,
+                0.528
+            ],
+            "angle": 0,
+            "content": "Multi-view CLIP Embedding. Camera viewpoints can serve as an effective condition signal to enhance 3D consistency in video content generation [52]. Building on this insight, we incorporate camera condition within our try-on network by encoding camera parameters as an additional token, enabling the generation of more consistent multi-view images. Specifically, we define a world coordinate system in which the camera faces the subject directly. For each input image (view) \\(\\mathbf{A}_{\\mathrm{i}}\\), \\(1 \\leq i \\leq m\\), we extract the rotation matrix from the camera's corresponding extrinsic matrix. This rotation matrix is then reshaped into a 9-dimensional tensor \\(\\mathbf{r}_{\\mathrm{i}}\\), which undergoes positional encoding to effectively integrate the camera parameters into the feature representation \\(F_{i}^{c}\\)."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.596,
+                0.54,
+                0.905,
+                0.579
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} F _ {i} ^ {c} = \\left(\\sin \\left(2 ^ {0} \\pi \\mathbf {r} _ {\\mathbf {i}}\\right), \\cos \\left(2 ^ {0} \\pi \\mathbf {r} _ {\\mathbf {i}}\\right), \\dots , \\right. \\tag {4} \\\\ \\sin (2 ^ {L - 1} \\pi \\mathbf {r _ {i}}), \\cos (2 ^ {L - 1} \\pi \\mathbf {r _ {i}})) \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.591,
+                0.906,
+                0.68
+            ],
+            "angle": 0,
+            "content": "where \\(L\\) is the length of positional embedding. We then project \\(F_{i}^{c}\\) to match the dimensionality of the garment CLIP image embedding \\(F^{g}\\) via an MLP and concatenate them along the token axis to form \\(Y_{i}\\). This combined representation, \\(Y_{i}\\), is subsequently used in the key \\(K_{\\mathrm{x}}\\) and value \\(V_{\\mathrm{x}}\\) of the cross-attention layers of the Main UNet:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.636,
+                0.694,
+                0.778,
+                0.71
+            ],
+            "angle": 0,
+            "content": "\\[\nY _ {i} = F ^ {g} \\oplus \\operatorname {M L P} \\left(F _ {i} ^ {c}\\right)\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.561,
+                0.713,
+                0.905,
+                0.737
+            ],
+            "angle": 0,
+            "content": "\\[\nQ _ {\\mathrm {x}} = W _ {\\mathrm {x}} ^ {Q} \\mathbf {H} _ {\\mathrm {i}} ^ {1}, K _ {\\mathrm {x}} = W _ {\\mathrm {x}} ^ {K} Y _ {i}, V _ {\\mathrm {x}} = W _ {\\mathrm {x}} ^ {V} Y _ {i} \\tag {5}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.588,
+                0.735,
+                0.795,
+                0.769
+            ],
+            "angle": 0,
+            "content": "\\[\nF _ {i} ^ {(l + 1)} = \\mathrm {s o f t m a x} (\\frac {Q _ {\\mathrm {x}} K _ {\\mathrm {x}} ^ {T}}{\\sqrt {d _ {\\mathrm {x}}}}) V _ {\\mathrm {x}},\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.78,
+                0.905,
+                0.826
+            ],
+            "angle": 0,
+            "content": "where \\(\\mathbf{H}^1\\) is the output of the MVAttention of the \\(l\\)-th layer; we omitted the \\(l\\) of the cross attention matrices and parameters for simplicity."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.827,
+                0.905,
+                0.871
+            ],
+            "angle": 0,
+            "content": "Training. Our enhanced multi-view 2D VTON network can be trained by minimizing the following latent diffusion model loss function:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.555,
+                0.884,
+                0.905,
+                0.904
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\mathrm {l d m}} = \\mathbb {E} _ {z _ {t}, \\eta , \\psi , \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\eta , \\psi , \\zeta)) \\| _ {2} ^ {2} \\right], \\tag {6}\n\\]"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "26392"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.094,
+                0.089,
+                0.487,
+                0.295
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.09,
+                0.302,
+                0.487,
+                0.375
+            ],
+            "angle": 0,
+            "content": "Figure 4. Illustration of the proposed Multi-view Spatial Attention. Query (Q): multi-view features \\(\\mathbf{F}^1\\); Key (K) and Value (V): concatenation of \\(\\mathbf{F}^1\\) and garment features \\(F_{f}^{l}\\) and \\(F_{b}^{l}\\). The attention score between viewpoints \\(i\\) and \\(j\\) is modulated by a weight \\(C_{ij}\\), determined by the cosine of the angle between them."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.381,
+                0.484,
+                0.444
+            ],
+            "angle": 0,
+            "content": "where \\(\\eta = [\\mathcal{E}(g_f); \\mathcal{E}(g_b); \\mathcal{E}(\\mathbf{N_i})_{i=1}^m]\\) represents the input latent garment images and latent normal maps; \\(\\zeta = [\\mathcal{E}'(\\mathbf{A_i})_{i=1}^m]\\) denotes the input latent clothing-agnostic images; \\(\\psi = \\mathbf{Y}\\) is the proposed multi-view CLIP embedding."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.457,
+                0.226,
+                0.475
+            ],
+            "angle": 0,
+            "content": "5. Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.482,
+                0.283,
+                0.5
+            ],
+            "angle": 0,
+            "content": "5.1. Experimental Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.504,
+                0.483,
+                0.776
+            ],
+            "angle": 0,
+            "content": "Datasets. We conduct experiments on two public datasets: Thuman2.0 [55] and MVHumanNet [51]. Thuman2.0 comprises 526 reconstructed clothed human scans, from which we render multi-view input images. Of these samples, 426 are used for training, while the remaining 100 are set aside for testing. To further evaluate the effectiveness and robustness of our method, we also perform experiments on MVHumanNet, a large-scale dataset of multi-view human images that encompasses a diverse range of subjects, daily outfits and motion sequences. The images in MVHumanNet are captured using a multi-view system with either 48 or 24 cameras. We use 4,990 subjects from this dataset, allocating 4,790 to training and 200 for tests. For each subject, we randomly select two frames of multi-view images from its entire motion sequence. While MVHumanNet provides multi-view images directly for editing and reconstruction, we render uniformly distributed views around each human subject in Thuman2.0 to ensure consistent input."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.781,
+                0.483,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Baselines. We primarily compare our method with three existing methods: DreamWaltz [23], GaussCtrl [48], and TIP-Editor [63]. DreamWaltz is a method designed for directly generating 3D human bodies based on textual descriptions, while GaussCtrl and TIP-Editor are two radiance-based editing methods. GaussCtrl is based on Stable Diffusion, using a description-like prompt to edit the scene. TIP-Editor accepts both text and image prompts. We configure it by"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.092,
+                0.905,
+                0.152
+            ],
+            "angle": 0,
+            "content": "specifying the human body as the editing region and the desired garment as the image prompt. We use ChatGPT to generate the text prompts corresponding to the clothing images."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.156,
+                0.907,
+                0.352
+            ],
+            "angle": 0,
+            "content": "Evaluation Metrics. For quantitative evaluation, we assess garment-to-person alignment between the edited person and the reference image. Following [63], we calculate the average DINO similarity [34] between the reference image and the rendered multi-view images of the edited 3D scene. Additionally, to evaluate multi-view consistency, we compute the CLIP Directional Consistency Score as outlined in [17]. Given the large scale of experiments (repeated 3DGS reconstruction), we selected a subset of examples from the dataset for metric evaluation. Specifically, from the test sets of Thuman and MVHumanNet, we randomly sampled 10 human scans each, performing virtual try-on with 6 randomly chosen garments per human scan."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.352,
+                0.909,
+                0.489
+            ],
+            "angle": 0,
+            "content": "We further conducted a user study involving 50 participants who rated the results of our method and three baseline methods based on two criteria: overall \"Quality\" and \"Alignment\" with the reference image. Each evaluation consisted of two questions: (1) Which method produces the highest quality of the edited 3D human? and (2) Which method achieves the most consistent alignment with the target clothing? Participants viewed the VTON results as rotating randomized video sequences."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.492,
+                0.909,
+                0.612
+            ],
+            "angle": 0,
+            "content": "Implementation Details. During pre-processing, we crop the multi-view images to the bounding box around the person and resize them to a resolution of \\(768 \\times 576\\). The front view and the back view of garment images are obtained from the corresponding clothed human images. After editing, we pad the images back to their original size. The data processing pipeline is the same for both Thuman2.0 and MVHumanNet datasets."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.613,
+                0.909,
+                0.84
+            ],
+            "angle": 0,
+            "content": "The Main UNet and the GarmentNet are initialized by the Stable Diffusion V1.5 [39]. The training process is divided into two stages. In the first stage, each view is trained independently, during which we establish the feature extraction capabilities of both the PoseEncoder and GarmentNet, as well as the generative capability of the Main UNet. The second stage involves multi-view training, where we randomly select \\( M \\) views for each human subject. This stage is focused on training the proposed MVAttention module to enhance the network's ability to maintain consistency across views. Due to memory constraints, we set \\( M = 8 \\) for the training phase. During the testing phase, we uniformly sampled 32 views from a 360-degree rotation around the subject. The editing of these 32 views is conducted in two batches, with each batch processing \\( M = 16 \\) views."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.849,
+                0.89,
+                0.866
+            ],
+            "angle": 0,
+            "content": "5.2. Comparisons with State-of-the-Art Methods"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.871,
+                0.908,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Qualitative Evaluation. Fig. 5 shows visual comparisons between our method and the baselines. DreamWaltz [23]"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "26393"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.097,
+                0.089,
+                0.903,
+                0.181
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"4\">Thuman2.0 [55]</td><td colspan=\"4\">MVHumanNet [51]</td></tr><tr><td>CLIPcons↑</td><td>DINOsim↑</td><td>Votequality</td><td>Votealign</td><td>CLIPcons↑</td><td>DINOsim↑</td><td>Votequality</td><td>Votealign</td></tr><tr><td>DreamWaltz [23]</td><td>0.887</td><td>0.556</td><td>0.46%</td><td>1.54%</td><td>0.935</td><td>0.495</td><td>0.46%</td><td>0.46%</td></tr><tr><td>TIP-Editor [63]</td><td>0.939</td><td>0.569</td><td>0.92%</td><td>0.62%</td><td>0.948</td><td>0.512</td><td>2.15%</td><td>1.38%</td></tr><tr><td>GaussCtrl [48]</td><td>0.931</td><td>0.577</td><td>1.08%</td><td>1.38%</td><td>0.938</td><td>0.521</td><td>1.69%</td><td>1.23%</td></tr><tr><td>Ours</td><td>0.923</td><td>0.633</td><td>97.54%</td><td>96.46%</td><td>0.933</td><td>0.623</td><td>95.69%</td><td>96.92%</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.116,
+                0.186,
+                0.88,
+                0.201
+            ],
+            "angle": 0,
+            "content": "Table 1. Quantitative comparisons. \\( {\\mathrm{{CLIP}}}_{\\text{cons }} \\) denotes the CLIP Direction Consistency Score. \\( {\\mathrm{{DINO}}}_{\\text{sim }} \\) is the DINO similarity."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.134,
+                0.204,
+                0.868,
+                0.613
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.089,
+                0.625,
+                0.908,
+                0.668
+            ],
+            "angle": 0,
+            "content": "Figure 5. Qualitative comparison. The first two rows show the results on Thuman2.0 dataset while the last two rows show the results on MVHumaNet dataset. Our method achieves good texture preservation (highlighted by the blue boxes), while three baseline methods mostly fail."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.679,
+                0.482,
+                0.83
+            ],
+            "angle": 0,
+            "content": "regenerates 3D clothed humans from text prompts but struggles to accurately retain both body and clothing characteristics. GaussCtrl [48], lacking support for image prompts, fails to maintain detailed clothing textures. While TipEditor [63] leverages Lora [21] for personalization, it encounters difficulties in consistently mapping clothing inputs from two views into the 3D human because the personalized concept are semantic in 2D space. In contrast, our method effectively preserves intricate clothing details, such as text, stripes, and logos."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.84,
+                0.483,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Quantitative Evaluation. Tab. 1 shows the results for the CLIP Directional Consistency Score and DINO similarity on Thuman2.0 and MVHumanNet datasets. Our approach surpasses other methods on \\(\\mathrm{DINO}_{sim}\\), clearly illustrating"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.679,
+                0.907,
+                0.83
+            ],
+            "angle": 0,
+            "content": "the superiority of our method in terms of garment texture preservation. While our results on \\(\\mathrm{CLIP}_{\\text {cons }}\\) are comparable to those of other methods, it is important to note that these methods incorporate the SDS loss, which to some extent smoothes the representation of humans in 3D space. Additionally, the \"flatter\" textures of other methods could also result in artificially higher consistency scores. Furthermore, user studies have shown that our method significantly exceeds baselines in terms of edited 3D human quality and the alignment of clothing details."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.846,
+                0.88,
+                0.863
+            ],
+            "angle": 0,
+            "content": "5.3. Visual Results using E-commerce Garment"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.871,
+                0.908,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Fig. 6 showcases VTON results using garments from the MVG dataset [46], whose images are from e-commerce"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "26394"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.095,
+                0.09,
+                0.482,
+                0.341
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.09,
+                0.361,
+                0.485,
+                0.446
+            ],
+            "angle": 0,
+            "content": "Figure 6. Generalization to e-commerce garments (the MVG dataset). Our method, trained on the THuman2.0 dataset, shows strong generalizability when applied to e-commerce garments. For clarity in visualization, we display garment images on human models; however, in the actual VTON process, the garments are segmented from the models using parse maps."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.456,
+                0.484,
+                0.577
+            ],
+            "angle": 0,
+            "content": "platforms like YOOX NET-A-PORTER, Taobao, and TikTok*, and a model trained on the Thuman2.0 dataset [55]. The results demonstrate that our method effectively preserves intricate garment details and textures. For instance, it accurately maintains the stripe patterns in the first row, the cute tie in the second row, and the buttons in the third row, highlighting the robustness of our approach in handling diverse and realistic clothing items."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.591,
+                0.245,
+                0.607
+            ],
+            "angle": 0,
+            "content": "5.4. Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.615,
+                0.484,
+                0.871
+            ],
+            "angle": 0,
+            "content": "We conduct an ablation study on Thuman2.0 dataset in Tab. 2 and Fig. 7 to evaluate the impact of our three proposed modules in enhancing a typical 2D VTON network with 3D-consistent generation capabilities. Starting with the 2D VTON baseline [53] using DensePose, we progressively replace DensePose with our pseudo-3D pose, incorporate multi-view CLIP embeddings, and ultimately integrate MVAttention in the final configuration. Results in Tab. 2 indicate that each module contributes to metric improvements. Fig. 7 visualizes an example of multi-view image editing. The incorporation of pseudo-3D pose substantially improves limb generation compared to the 2D VTON baseline. Comparing rows 4 and 5, prior to the integration of multi-view CLIP embedding, the model captures limited spatial information, resulting in detail loss at specific angles (columns 3, 4, and 6). Finally, the proposed MVAttention achieves a more coherent generation across views."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.536,
+                0.089,
+                0.885,
+                0.413
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.513,
+                0.417,
+                0.908,
+                0.515
+            ],
+            "angle": 0,
+            "content": "Figure 7. Visualization of the impact of the three proposed techniques on multi-view consistent editing. The red boxes highlight the artifacts. Starting from the 2D VTON baseline, the pseudo-3D pose improves limb generation quality, multi-view CLIP embedding enhances detail across different viewing directions, and finally, MVAttention further strengthens consistency in the generated images."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.518,
+                0.527,
+                0.9,
+                0.62
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Methods</td><td>CLIPcons↑</td><td>DINOsim↑</td></tr><tr><td>2D-VTON</td><td>0.892</td><td>0.609</td></tr><tr><td>+ Pseudo-3D Pose</td><td>0.910</td><td>0.626</td></tr><tr><td>+ Multi-view CLIP Embedding</td><td>0.913</td><td>0.631</td></tr><tr><td>+ Multi-view Spatial Attention</td><td>0.923</td><td>0.633</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.513,
+                0.624,
+                0.906,
+                0.652
+            ],
+            "angle": 0,
+            "content": "Table 2. Ablation studies. We ablate the impact of the three proposed techniques on Thuman2.0 dataset."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.673,
+                0.642,
+                0.69
+            ],
+            "angle": 0,
+            "content": "6. Conclusions"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.704,
+                0.907,
+                0.901
+            ],
+            "angle": 0,
+            "content": "In this work, we proposed VTON 360, a novel 3D Virtual Try-On (VTON) method that achieves high-fidelity VTON with the ability to render clothing from arbitrary viewing directions. Our method features a novel formulation of 3D VTON as an extension of 2D VTON that ensures 3D consistent results across multiple views. To bridge the gap between 2D VTON models and 3D consistency requirements, we introduce several key innovations, including multi-view inputs, pseudo-3D pose representation, multi-view spatial attention, and multi-view CLIP embedding. Extensive experiments demonstrate the effectiveness of our approach, significantly outperforming prior 3D VTON techniques in both fidelity and any-view rendering."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.11,
+                0.887,
+                0.441,
+                0.901
+            ],
+            "angle": 0,
+            "content": "* https://net-a-porter.com, www.taobao.com, www.douyin.com"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "26395"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.092,
+                0.091,
+                0.251,
+                0.108
+            ],
+            "angle": 0,
+            "content": "Acknowledgement"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.116,
+                0.484,
+                0.207
+            ],
+            "angle": 0,
+            "content": "This work is supported in part by the National Key R&D Program of China under Grant No.2024YFB3908503, in part by the National Natural Science Foundation of China under Grant NO. 62322608 and in part by the CCF-Kuaishou Large Model Explorer Fund (NO. CCF-KuaiShou 2024007)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.093,
+                0.222,
+                0.188,
+                0.238
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.101,
+                0.248,
+                0.483,
+                0.274
+            ],
+            "angle": 0,
+            "content": "[1] Easymocap - make human motion capture easier. Github, 2021.5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.101,
+                0.277,
+                0.484,
+                0.332
+            ],
+            "angle": 0,
+            "content": "[2] Shuai Bai, Huiling Zhou, Zhikang Li, Chang Zhou, and Hongxia Yang. Single stage virtual try-on via deformable attention flows. In European Conference on Computer Vision, pages 409-425. Springer, 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.335,
+                0.484,
+                0.403
+            ],
+            "angle": 0,
+            "content": "[3] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5420-5430, 2019. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.406,
+                0.483,
+                0.473
+            ],
+            "angle": 0,
+            "content": "[4] Robert Bridson, Ronald Fedkiw, and John Anderson. Robust treatment of collisions, contact and friction for cloth animation. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 594-603, 2002. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.477,
+                0.484,
+                0.544
+            ],
+            "angle": 0,
+            "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.548,
+                0.484,
+                0.615
+            ],
+            "angle": 0,
+            "content": "[6] Haodong Chen, Yongle Huang, Haojian Huang, Xiangsheng Ge, and Dian Shao. Gaussianvton: 3d human virtual try-on via multi-stage gaussian splatting editing with image prompting. arXiv preprint arXiv:2405.07472, 2024. 2, 3, 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.619,
+                0.484,
+                0.701
+            ],
+            "angle": 0,
+            "content": "[7] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.704,
+                0.484,
+                0.773
+            ],
+            "angle": 0,
+            "content": "[8] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14131-14140, 2021. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.775,
+                0.484,
+                0.829
+            ],
+            "angle": 0,
+            "content": "[9] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. arXiv preprint arXiv:2403.05139, 2024. 3, 4, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.832,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[10] Xin Dong, Fuwei Zhao, Zhenyu Xie, Xijin Zhang, Daniel K Du, Min Zheng, Xiang Long, Xiaodan Liang, and Jianchao Yang. Dressing in the wild by watching dance videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3480-3489, 2022. 2"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.094,
+                0.248,
+                0.484,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.093,
+                0.905,
+                0.161
+            ],
+            "angle": 0,
+            "content": "[11] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8485-8493, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.164,
+                0.905,
+                0.219
+            ],
+            "angle": 0,
+            "content": "[12] Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang. Taming the power of diffusion models for high-quality virtual try-on with appearance flow. arXiv preprint arXiv:2308.06101, 2023. 1, 2, 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.221,
+                0.905,
+                0.274
+            ],
+            "angle": 0,
+            "content": "[13] Peng Guan, Loretta Reiss, David A Hirshberg, Alexander Weiss, and Michael J Black. Drape: Dressing any person. ACM Transactions on Graphics (ToG), 31(4):1-10, 2012. 1, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.277,
+                0.905,
+                0.332
+            ],
+            "angle": 0,
+            "content": "[14] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densepose: Dense human pose estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7297-7306, 2018. 4, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.334,
+                0.905,
+                0.403
+            ],
+            "angle": 0,
+            "content": "[15] Fabian Hahn, Bernhard Thomaszewski, Stelian Coros, Robert W Sumner, Forrester Cole, Mark Meyer, Tony DeRose, and Markus Gross. Subspace clothing simulation using adaptive bases. ACM Transactions on Graphics (TOG), 33(4):1-9, 2014. 1, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.406,
+                0.905,
+                0.46
+            ],
+            "angle": 0,
+            "content": "[16] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.462,
+                0.905,
+                0.531
+            ],
+            "angle": 0,
+            "content": "[17] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19740–19750, 2023. 3, 6, 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.533,
+                0.905,
+                0.587
+            ],
+            "angle": 0,
+            "content": "[18] Sen He, Yi-Zhe Song, and Tao Xiang. Style-based global appearance flow for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3470-3479, 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.59,
+                0.905,
+                0.658
+            ],
+            "angle": 0,
+            "content": "[19] Zijian He, Peixin Chen, Guangrun Wang, Guanbin Li, Philip HS Torr, and Liang Lin. Wildvidfit: Video virtual try-on in the wild via image-based controlled diffusion models. In European Conference on Computer Vision, pages 123-139. Springer, 2024. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.66,
+                0.905,
+                0.701
+            ],
+            "angle": 0,
+            "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.703,
+                0.905,
+                0.758
+            ],
+            "angle": 0,
+            "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.76,
+                0.905,
+                0.815
+            ],
+            "angle": 0,
+            "content": "[22] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.817,
+                0.905,
+                0.871
+            ],
+            "angle": 0,
+            "content": "[23] Yukun Huang, Jianan Wang, Ailing Zeng, He Cao, Xianbiao Qi, Yukai Shi, Zheng-Jun Zha, and Lei Zhang. Dreamwaltz: Make a scene with complex 3d animatable avatars. Advances in Neural Information Processing Systems, 36, 2024. 3, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.874,
+                0.905,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[24] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.517,
+                0.093,
+                0.905,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.945,
+                0.521,
+                0.957
+            ],
+            "angle": 0,
+            "content": "26396"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.125,
+                0.092,
+                0.482,
+                0.133
+            ],
+            "angle": 0,
+            "content": "reconstruction of lifelike clothed humans. In 2024 International Conference on 3D Vision (3DV), pages 1531-1542. IEEE, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.134,
+                0.483,
+                0.189
+            ],
+            "angle": 0,
+            "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 3, 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.19,
+                0.482,
+                0.245
+            ],
+            "angle": 0,
+            "content": "[26] Jeongho Kim, Gyojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. arXiv preprint arXiv:2312.01725, 2023. 3, 4, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.247,
+                0.482,
+                0.273
+            ],
+            "angle": 0,
+            "content": "[27] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.274,
+                0.482,
+                0.328
+            ],
+            "angle": 0,
+            "content": "[28] Zorah Lahner, Daniel Cremers, and Tony Tung. Deepwrinkles: Accurate and realistic clothing modeling. In Proceedings of the European conference on computer vision (ECCV), pages 667-684, 2018. 1, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.33,
+                0.482,
+                0.398
+            ],
+            "angle": 0,
+            "content": "[29] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In Proceedings of the European conference on computer vision (ECCV), 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.4,
+                0.482,
+                0.455
+            ],
+            "angle": 0,
+            "content": "[30] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851-866. 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.457,
+                0.482,
+                0.524
+            ],
+            "angle": 0,
+            "content": "[31] Yifang Men, Yiming Mao, Yuning Jiang, Wei-Ying Ma, and Zhouhui Lian. Controllable person image synthesis with attribute-decomposed gan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5084-5093, 2020. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.525,
+                0.482,
+                0.593
+            ],
+            "angle": 0,
+            "content": "[32] Aymen Mir, Thiemo Alldieck, and Gerard Pons-Moll. Learning to transfer texture from clothing images to 3d humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7023-7034, 2020. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.595,
+                0.482,
+                0.65
+            ],
+            "angle": 0,
+            "content": "[33] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. Ladi-vton: Latent diffusion textual-inversion enhanced virtual try-on. arXiv preprint arXiv:2305.13501, 2023. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.651,
+                0.482,
+                0.719
+            ],
+            "angle": 0,
+            "content": "[34] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.72,
+                0.483,
+                0.803
+            ],
+            "angle": 0,
+            "content": "[35] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.804,
+                0.483,
+                0.858
+            ],
+            "angle": 0,
+            "content": "[36] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael J Black. Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics (ToG), 36(4):1-15, 2017. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.859,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[37] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.093,
+                0.092,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.092,
+                0.906,
+                0.161
+            ],
+            "angle": 0,
+            "content": "[38] Yurui Ren, Xiaoqing Fan, Ge Li, Shan Liu, and Thomas H Li. Neural texture extraction and distribution for controllable person image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13535-13544, 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.163,
+                0.907,
+                0.232
+            ],
+            "angle": 0,
+            "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 3, 4, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.233,
+                0.906,
+                0.287
+            ],
+            "angle": 0,
+            "content": "[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.288,
+                0.907,
+                0.357
+            ],
+            "angle": 0,
+            "content": "[41] Igor Santesteban, Nils Thuerey, Miguel A Otaduy, and Dan Casas. Self-supervised collision handling via generative 3d garment models for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11763-11773, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.358,
+                0.906,
+                0.412
+            ],
+            "angle": 0,
+            "content": "[42] Igor Santesteban, Miguel Otaduy, Nils Thuerey, and Dan Casas. Ulnef: Untangled layered neural fields for mix-and-match virtual try-on. Advances in Neural Information Processing Systems, 35:12110-12125, 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.413,
+                0.906,
+                0.48
+            ],
+            "angle": 0,
+            "content": "[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.482,
+                0.906,
+                0.524
+            ],
+            "angle": 0,
+            "content": "[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.525,
+                0.906,
+                0.566
+            ],
+            "angle": 0,
+            "content": "[45] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.567,
+                0.906,
+                0.608
+            ],
+            "angle": 0,
+            "content": "[46] Haoyu Wang, Zhilu Zhang, Donglin Di, Shiliang Zhang, and Wangmeng Zuo. Mv-vton: Multi-view virtual try-on with diffusion models. arXiv preprint arXiv:2404.17364, 2024. 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.609,
+                0.906,
+                0.676
+            ],
+            "angle": 0,
+            "content": "[47] Junjie Wang, Jiemin Fang, Xiaopeng Zhang, Lingxi Xie, and Qi Tian. Gaussianeditor: Editing 3d gaussians delicately with text instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20902-20911, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.678,
+                0.906,
+                0.734
+            ],
+            "angle": 0,
+            "content": "[48] Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. Gaussctrl: multi-view consistent text-driven 3d gaussian splattering editing. arXiv preprint arXiv:2403.08733, 2024. 3, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.735,
+                0.906,
+                0.816
+            ],
+            "angle": 0,
+            "content": "[49] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.818,
+                0.906,
+                0.871
+            ],
+            "angle": 0,
+            "content": "[50] Zhenyu Xie, Haoye Dong, Yufei Gao, Zehua Ma, and Xiaodan Liang. Dreamvton: Customizing 3d virtual try-on with personalized diffusion models. arXiv preprint arXiv:2407.16511, 2024. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.873,
+                0.906,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[51] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu,"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.517,
+                0.092,
+                0.907,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "26397"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.125,
+                0.092,
+                0.483,
+                0.161
+            ],
+            "angle": 0,
+            "content": "Chongjie Wang, Shijie Wang, et al. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19801-19811, 2024. 2, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.162,
+                0.483,
+                0.218
+            ],
+            "angle": 0,
+            "content": "[52] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.219,
+                0.482,
+                0.272
+            ],
+            "angle": 0,
+            "content": "[53] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. arXiv preprint arXiv:2403.01779, 2024. 3, 4, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.274,
+                0.483,
+                0.328
+            ],
+            "angle": 0,
+            "content": "[54] Han Yang, Xinrui Yu, and Ziwei Liu. Full-range virtual try-on with recurrent tri-level transform. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3460-3469, 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.33,
+                0.482,
+                0.411
+            ],
+            "angle": 0,
+            "content": "[55] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5746-5756, 2021. 2, 6, 7, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.413,
+                0.482,
+                0.468
+            ],
+            "angle": 0,
+            "content": "[56] Jinsong Zhang, Kun Li, Yu-Kun Lai, and Jingyu Yang. Pise: Person image synthesis and editing with decoupled gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7982-7990, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.47,
+                0.482,
+                0.509
+            ],
+            "angle": 0,
+            "content": "[57] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.511,
+                0.482,
+                0.566
+            ],
+            "angle": 0,
+            "content": "[58] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.567,
+                0.482,
+                0.648
+            ],
+            "angle": 0,
+            "content": "[59] Fuwei Zhao, Zhenyu Xie, Michael Kampffmeyer, Haoye Dong, Songfang Han, Tianxiang Zheng, Tao Zhang, and Xiaodan Liang. M3d-vton: A monocular-to-3d virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13239-13249, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.65,
+                0.482,
+                0.732
+            ],
+            "angle": 0,
+            "content": "[60] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4606-4615, 2023. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.734,
+                0.482,
+                0.789
+            ],
+            "angle": 0,
+            "content": "[61] Jingyu Zhuang, Chen Wang, Liang Lin, Lingjie Liu, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. In SIGGRAPH Asia 2023 Conference Papers, pages 1-10, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.79,
+                0.482,
+                0.843
+            ],
+            "angle": 0,
+            "content": "[62] Jingyu Zhuang, Di Kang, Linchao Bao, Liang Lin, and Guanbin Li. Dagensm: Disentangled avatar generation with gs-enhanced mesh. arXiv preprint arXiv:2411.15205, 2024.3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.845,
+                0.482,
+                0.9
+            ],
+            "angle": 0,
+            "content": "[63] Jingyu Zhuang, Di Kang, Yan-Pei Cao, Guanbin Li, Liang Lin, and Ying Shan. Tip-editor: An accurate 3d editor following both text-prompts and image-prompts. arXiv preprint arXiv:2401.14828, 2024. 3, 6, 7, 1"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.093,
+                0.092,
+                0.483,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "26398"
+        }
+    ]
+]

2025/VTON 360_ High-Fidelity Virtual Try-On from Any Viewing Direction/4af7cfba-2431-4854-85f1-f3f911dcc811_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8beaca2d741a0c68bff5d3478c50c4d6e1879052f576cc81d91835be2382bf22
+size 2082731

2025/VTON 360_ High-Fidelity Virtual Try-On from Any Viewing Direction/full.md

@@ -0,0 +1,284 @@
+# VTON 360: High-Fidelity Virtual Try-On from Any Viewing Direction
+
+Zijian He<sup>1</sup> Yuwei Ning<sup>1</sup> Yipeng Qin<sup>2</sup> Guangrun Wang<sup>1</sup> Sibei Yang<sup>3</sup> Liang Lin<sup>1,4,5</sup> Guanbin Li<sup>1,4,5*</sup>  
+<sup>1</sup>Sun Yat-sen University  
+<sup>2</sup>Cardiff University  
+<sup>3</sup>ShanghaiTech University  
+<sup>4</sup>Guangdong Key Laboratory of Big Data Analysis and Processing  
+<sup>5</sup>Peng Cheng Laboratory  
+hezj39@mail2.sysu.edu.cn, yuwei.ning@hust.edu.cn, {qinyipengl991, wanggrun}@gmail.com, yangsb@shanghaiitech.edu.cn linliang@ieee.org, liguanbin@mail.sysu.edu.cn
+
+![](images/3e4ed0ca9fe9c0466bbb1e4f4019077b1eea386c4b063ea02f3897e6315ec9c9.jpg)  
+Figure 1. Results of VTON 360. Our VTON 360 enables high-fidelity 3D Virtual Try-On (VTON) by seamlessly adapting E-commerce garments onto a clothed 3D human model, supporting full  $360^{\circ}$  view rendering. The highlighted bounding boxes (dashed line) demonstrate our method's ability to preserve intricate clothing details and patterns (e.g., collar accessories, horizontal line patterns, logos, texts, numbers) across diverse garment types.
+
+![](images/18ca224d94844f9dde5541e3e34a54b63839bcf1d69569779a98633de60a2a1f.jpg)
+
+![](images/a5889a307af41522f81b308115e028b9c3f02dce5c03771e67bcdd87075b042d.jpg)
+
+# Abstract
+
+Virtual Try-On (VTON) is a transformative technology in e-commerce and fashion design, enabling realistic digital visualization of clothing on individuals. In this work, we propose VTON 360, a novel 3D VTON method that addresses the open challenge of achieving high-fidelity VTON that supports any-view rendering. Specifically, we leverage the equivalence between a 3D model and its rendered multiview 2D images, and reformulate 3D VTON as an extension of 2D VTON that ensures 3D consistent results across multiple views. To achieve this, we extend 2D VTON models to include multi-view garments and clothing-agnostic human body images as input, and propose several novel techniques to enhance them, including: i) a pseudo-3D pose representation using normal maps derived from the SMPL-X 3D human model, ii) a multi-view spatial attention mechanism that models the correlations between features from different viewing angles, and iii) a multi-view CLIP embedding that enhances the garment CLIP features used in 2D VTON with camera information. Extensive experiments on large-scale real datasets and clothing images from e-commerce
+
+platforms demonstrate the effectiveness of our approach.  
+Project page: https://scnuhealthy.github.io/  
+VTON360.
+
+# 1. Introduction
+
+Virtual Try-On (VTON) enables realistic digital visualization of clothing on individuals and has emerged as a transformative technology in e-commerce and fashion design. While significant research efforts have been made on 2D VTON solutions [8, 12, 19, 33, 39], these approaches are inherently limited in their representation of view-related features. To overcome this limitation and enable high-fidelity any-view rendering, 3D VTON methods were introduced.
+
+3D VTON requires accurate garment transfer onto a 3D human body while ensuring realistic garment fitting, texture preservation, and 3D consistency. The two primary aims of 3D VTON are i) achieving high-fidelity and ii) supporting any-view rendering. Leveraging the inherent capability of 3D models for any-view rendering, early 3D VTON methods [13, 15, 28] make clothing simulation on synthetic human bodies. Specifically, these methods utilized 3D scanners to capture clothing meshes, followed by the development of specialized dressing algorithms. Although effec
+
+tive, these methods rely on costly 3D scanning equipment and the physical presence of the human body/clothing (i.e., not fully virtual), restricting their practicality in real-world applications. As a byproduct, most early methods focused on developing geometrically correct dressing algorithms using standard templates of human body and clothing models. Addressing this limitation, researchers extended 3D VTON by introducing algorithms that reconstruct 3D clothing models from input images, enabling the use of image-based clothing inputs [3, 32, 41, 42]. However, since input clothing images (usually frontal) are inherently 2D and lack multi-view information, this approach struggles to reconstruct high-fidelity clothing models that can be rendered well from all viewing directions.
+
+To complement this missing information, DreamVTON [50] introduces a novel approach that leverages Text-to-Image (T2I) diffusion models to reconstruct both the human body and clothing from input images. Its key insight is that T2I models learned view-agnostic "concepts" of both bodies and garments during their training, and that the corresponding concepts for the input body and clothing images can be obtained using LoRA [21]. By utilizing Score Distillation Sampling (SDS) [37], DreamVTON can generate visual-pleasing 3D VTON results by ensuring consistency between renderings from arbitrary viewpoints and the concepts. Nonetheless, DreamVTON's high flexibility comes at the cost of low fidelity. This limitation stems from the fact that the concepts learned by T2I models are semantic in nature, thus lacking 3D geometric consistency and pixel-level accuracy with respect to the input body and clothing images. Recently, a concurrent work, namely GaussianVTON [6], partially addressed this limitation by formulating 3D VTON as a 3D scene editing task, where a given 3D human model is edited using multi-view images generated by 2D VTON methods. While it significantly enhances the fidelity of the human body, the fidelity and 3D consistency of clothing remain problematic, as there are no 2D VTON methods that can generate multi-view images with 3D consistency. Therefore, to the best of our knowledge, achieving high-fidelity 3D VTON that supports any-view rendering remains an open challenge.
+
+In this work, we address the above-mentioned challenge via proposing VTON 360, a novel 3D VTON method that achieves high-fidelity VTON from arbitrary viewing directions. Similar to GaussianVTON [6], our method edits a given 3D human model by inpainting the rendered images using a latent diffusion model. However, we set ourselves apart through our novel garment fidelity preservation strategy that can generate high-fidelity on-body garments in all viewing directions. Specifically, we first extend both the garment and clothing-agnostic human body inputs to typical 2D VTON models to leverage multi-view information, including paired front and back view garment images as well
+
+as a set of multi-view clothing-agnostic human body images sampled from random azimuth angles. Then, we propose several novel enhancements to bridge the gap between typical 2D VTON methods and our multi-view 3D consistency requirements: i) We propose a pseudo-3D pose representation using normal maps derived from the SMPL-X 3D human model, which captures fine-grained surface orientation details and provides more consistent geometry across views compared to the 2D pose representations (semantic segmentation maps) used in 2D VTON models. ii) We design a Multi-view Spatial Attention mechanism that models the correlations between features from different viewing angles, featuring a novel "correlation" matrix modeling the relationships among different input views. iii) We propose a multi-view CLIP embedding that enhances the garment CLIP embedding used in 2D VTON methods with camera information, thereby facilitating network learning of features relevant to a particular view. Together, these innovations enable our 2D VTON model to generate high-quality, multi-view and 3D-consistent virtual try-on results. Extensive experiments on Thuman2.0 [55] and MVHumanNet [51] datasets demonstrate that our method achieves high fidelity 3D VTON which supports any-view rendering. In addition, we show the effectiveness and generalizability of our methodology by testing it using garments from e-commerce platforms. Our conclusions include:
+
+- We propose a novel 3D Virtual Try-On (VTON) method, namely VTON 360, which achieves high-fidelity VTON from arbitrary viewing directions.  
+- Leveraging the equivalence between a 3D model and its rendered multi-view 2D images, we reformulate 3D VTON as an extension of 2D VTON that ensures consistent results across multiple views. Specifically, we introduce several novel techniques, including: (i) pseudo-3D pose representation; (ii) multi-view spatial attention; and (iii) multi-view CLIP embedding. These innovations enhance traditional 2D VTON models to generate multiview and 3D-consistent results.  
+- Extensive experimental results on two large real datasets as well as real clothing images from e-commerce platforms demonstrate the effectiveness of our approach.
+
+# 2. Related Work
+
+2D Virtual Try-On. 2D Virtual Try-On (VTON) aims to transfer a desired garment to the corresponding region of a target human image while preserving the human pose and identity. Early methods [2, 8, 10, 11, 16, 18, 29, 31, 38, 54, 56] use Generative Adversarial Networks (GANs) to deform the garments to match the target body shape, which a critical step for achieving realistic VTON. However, accurately adapting to diverse real-world conditions remains a significant challenge. Addressing this issue, recent methods [12, 19, 33, 60] reframe 2D VTON as a conditioned in-
+
+painting task, leveraging the strong priors provided by diffusion models [20, 43, 45] to achieve promising results. This strategy is further improved by [9, 26, 53], which introduce a ReferenceNet to extract hierarchical garment features and apply attention mechanisms to condition the Main UNet.
+
+3D Virutal Try-On. For 3D Virtual Try-On (VTON), traditional methods [4, 13, 15, 28, 36] rely on 3D scanning or cloth simulation to generate highly precise body and garment geometry. These methods were then extended by learning-based methods [3, 32] that employ differentiable rendering to dress the SMPL [30] model with a desired garment mesh. Despite their effectiveness, such methods rely on costly 3D scanning and the physical presence of human body/clothing, limiting their application in the real world. Addressing this limitation, M3D-VTON [59] proposes a depth-based 3D VTON framework to reconstruct 3D clothed human models from 2D human and garment images, but the results often suffer from explicit warping artifacts. To improve 3D VTON results, recent methods [23, 24, 50, 62] resort to text-to-image (T2I) diffusion models and employ the Score Distillation Sampling (SDS) loss [40] to ensure consistency among different viewing directions. Specifically, TeCH [24] adapts the generative priors of T2I diffusion model to the specific person and clothes by training descriptive text prompts with DreamBooth [40]. DreamWaltz [23] leverages Pose ControlNet [57] to attain clothed human body models. DreamVTON [50] introduces a multi-concept LoRA [21] to personalize the T2I diffusion model, and uses a template-based optimization mechanism that combines with SDS loss to better preserve patterns on the garment. Although effective, these methods often produce results lacking in fidelity, as the concepts learned by T2I models are semantic rather than at the pixel level. Concurrent to our work, GaussianVTON [6] proposes an alternative approach by combining Gaussian Splatting [25] with pre-trained 2D VTON models and formulate it as an editing task. However, since there are no 2D VTON methods that can generate multi-view images with 3D consistency, the fidelity and 3D consistency of the clothing generated remain problematic.
+
+Radiance Field-based 3D Human or Scene Editing. Recently, radiance field-based editing has attracted significant interest due to its efficient differentiable rendering capabilities, sparking substantial advancements in text-driven 3D editing. For example, InstructN2N [17] employ an image-based diffusion model InstructP2P [5] to modify the rendered image by the user's text description with a variant of the score distillation sampling (SDS) [37] loss. GaussianEditor [7] applies Gaussian Splatting [25] as 3D representation instead of NeRF, adopting Gaussian semantic tracking to track target Gaussian values, significantly improving editing speed and controllability. To enable accu
+
+rate location and appearance control, subsequent works [47, 61] specify the editing region using the attention score or with a segmentation model. TIP-Editor [63] proposes a content personalization step dedicated to the reference image based on LoRA, achieving the editing following a hybrid text-image prompt. GaussCtrl [48] leverage depth conditions and attention-based latent code alignment to achieve 3D-aware multi-view consistent editing instead of iteratively editing single views using SDS loss. However, these works primarily focus on global appearance modifications within a text-driven pipeline, while our approach emphasizes preserving fine textural details from different viewing directions throughout the editing process.
+
+# 3. Preliminary
+
+Latent Diffusion Model. Latent Diffusion Model [39] is a variant of diffusion models that performs denoising within the latent space of a Variational Autoencoder (VAE) [27]. During training, given a fixed encoder  $\mathcal{E}$ , an input image  $x$  is transformed into its latent representation  $z_0 = \mathcal{E}(x)$ . A conditional diffusion model  $\hat{\epsilon}_{\theta}$ , typically implemented with a UNet architecture, is then trained using a weighted denoising score matching objective:
+
+$$
+\mathcal {L} _ {L D M} = \mathbb {E} _ {\mathbf {z}, \mathbf {c}, \boldsymbol {\epsilon}, t} [ \boldsymbol {\epsilon} - \| \hat {\epsilon} _ {\theta} (\mathbf {z} _ {\mathbf {t}}; \mathbf {c}, t) \| _ {2} ^ {2} ] \tag {1}
+$$
+
+where  $\mathbf{z}_t\coloneqq \alpha_t\mathbf{x} + \sigma_t\boldsymbol {\epsilon}$  denotes the forward diffusion process at timestep  $t$  .  $\alpha_{t},\sigma_{t}$  are time-dependent functions defined by the diffusion model formulation; c denotes the conditional input and  $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{1})$  is Gaussian noise. During inference, data samples are generated by initiating from Gaussian noise  $\mathbf{z}_T\sim \mathcal{N}(\mathbf{0},\mathbf{1})$  and iteratively refining it using a DDIM [44] sampler.
+
+# 4. Method
+
+Our method leverages the equivalence between a 3D model and its rendered multi-view 2D images to achieve high-fidelity, any-view 3D VTON. Specifically, as Fig. 2 shows, given an input 3D human model and a garment image, our method 1) renders the 3D model into multi-view 2D images and 2) formulates 3D VTON as a consistent, unified 2D VTON process across these rendered views; 3) By reconstructing the edited images into a 3D model using existing 3D reconstruction methods, we ensure visual coherence and precise garment alignment from any viewing angle. Among them, the second step is crucial as existing 2D VTON methods [9, 26, 53] lack 3D knowledge, preventing them from generating multi-view images with 3D consistency.
+
+To address this challenge, we propose several novel techniques (Sec. 4.2) that equip a typical 2D VTON network (Sec. 4.1), which is built on a latent diffusion model [39], with the capability to generate 3D-consistent results. We use Gaussian Splatting [25] as our 3D representation.
+
+![](images/50456d17e5ccbb76f3e583fe2b0ebac3fa363e2d03f09e2a13d08f82ca40471e.jpg)  
+Figure 2. Overview of VTON 360. Given an input 3D human model  $\mathbf{G}_{\mathrm{src}}$  and a pair of garment images  $(g_{f}, g_{b})$ , our method 1) renders  $\mathbf{G}_{\mathrm{src}}$  into multi-view 2D images (left) and 2) edits the rendered multi-view 2D images (middle); 3) reconstructs the edited images into a 3D model  $\mathbf{G}_{\mathrm{VTON}}$  (right). In the crucial step 2), we propose three novel techniques to equip a typical 2D VTON network with the capability to generate 3D-consistent results: 1) Pseudo-3D Pose Input, 2) Multi-view Spatial Attention, and 3) Multi-view CLIP Embedding.
+
+![](images/c3dd68bc18957e8ae4d2aa15c28ee2931189de321d4ce4c838cc8d0a92c0acbb.jpg)  
+Figure 3. DensePose (2D) vs. SMPL-X normal map (pseudo-3D) representations. DensePose applies uniform labels per body part, lacking 3D consistency across views and causing artifacts and temporal inconsistencies (highlighted with red boxes). In contrast, SMPL-X normal maps capture fine surface details, ensuring geometric coherence and stable, realistic shading across views.
+
+# 4.1. Recap of 2D VTON Framework
+
+Following previous works [12, 26, 53], we formulate 2D VTON as an exemplar-based inpainting problem, aiming to fill an input clothing-agnostic image  $\mathbf{A}$  with a given garment image  $g$ , where  $\mathbf{A}$  is obtained following the method used in [53]. As illustrated in Fig. 2 (middle), the network architecture is based on the latent diffusion model [39] with
+
+an encoder  $\mathcal{E}$  and comprises two components:
+
+- A GarmentNet [9, 53] that extracts features from  $\mathcal{E}(g)$ .  
+- A Main UNet that inpaints  $\mathbf{A}$  according to i) detailed garment features extracted by the GarmentNet; ii) the 2D pose of  $\mathbf{A}$  represented by semantic labels using Dense-Pose [14]; iii) CLIP embeddings of input garment  $g$ . Among them, i) and ii) together with noise are input to the self-attention layers, while iii) is input to the cross-attention layers of the Main UNet.
+
+Both the GarmentNet and the Main UNet share the same network architecture.
+
+# 4.2. Multi-view 2D VTON with 3D Consistency
+
+To enable the aforementioned 2D VTON model to generate multi-view and 3D-consistent results, we propose the following novel enhancements to its design:
+
+Multi-view Inputs. We extend both inputs to the model:
+
+- Multi-view Garment Inputs: We extend the input garment representation from a single image  $g$  to paired front and back view images  $g_{f}$ ,  $g_{b}$ , providing comprehensive garment information across all viewing angles. Accordingly, we use the encoder  $\mathcal{E}$  to encode  $g_{f}, g_{b}$  into their latent representations  $\mathcal{E}(g_{f}), \mathcal{E}(g_{b})$  and feed them into GarmentNet to obtain their multi-layer features  $F_{f}^{l}$  and  $F_{b}^{l}$  at layer  $l$ , respectively.  
+- Multi-view Clothing-agnostic Image Inputs: Based on
+
+the equivalence between a 3D human model and its rendered multi-view 2D images, we extend the input human body representation from a single, clothing-agnostic image,  $\mathbf{A}$ , to a set of  $m$  multi-view images, denoted as  $\mathbf{A}_1, \mathbf{A}_2, \dots, \mathbf{A}_m$ . These images are sampled from random azimuth angles, allowing the 2D VTON model to access comprehensive, multi-view information from the input 3D human model.
+
+Pseudo-3D Pose Input. As shown in Fig. 3, the 2D Dense-Pose representations [14] commonly used in state-of-the-art 2D VTON methods [9, 26] assign a uniform semantic label to all pixels within each body part (e.g., thigh), inherently lack 3D geometric consistency across multiple views, and often introduce artifacts and temporal inconsistencies. To address these limitations, we propose a novel pseudo-3D pose representation: the normal maps  $\mathbf{N}$  derived from the SMPL-X [35] model of the input body. These normal maps capture fine-grained surface orientation details, providing a more consistent representation across views by preserving geometric structure in the 3D space. Furthermore, they enable smoother, temporally stable transitions and realistic shading effects across multi-view scenarios. In practice, we employ a lightweight PoseEncoder  $\mathcal{E}'$  [22] and feed  $\mathcal{E}'(\mathbf{N})$  into the Main UNet. We obtain the SMPL-X model from the multi-view images of the input body using EasyMoCap [1].
+
+Accordingly, we concatenate three components as the enhanced input to the Main UNet: i) a noise latent  $z_{t}$ ; ii) the encoded pseudo-3D poses  $\mathcal{E}'(\mathbf{N_1}),\mathcal{E}'(\mathbf{N_2}),\dots,\mathcal{E}'(\mathbf{N_m})$ ; and iii) the encoded multi-view clothing-agnostic images  $\mathcal{E}(\mathbf{A_1}),\mathcal{E}(\mathbf{A_2}),\dots,\mathcal{E}(\mathbf{A_m})$ . Let  $F_{1}^{l},F_{2}^{l},\dots,F_{m}^{l}$  be the feature representations at layer  $l$  of the Main UNet, and recall the garment features  $F_{f}^{l}$  and  $F_{b}^{l}$  defined above, we enhance the self-attention layers of the Main UNet as:
+
+Multi-view Spatial Attention. To cope with the aforementioned multi-view input features and ensure their consistency, we draw insights from the temporal attention layer commonly used in video generation and editing [49, 58] and extend it to our multi-view spatial attention layer, denoted as MVAttention. The key distinction of our MVAttention is that its input multi-view features  $F_{1}^{l}, F_{2}^{l}, \ldots, F_{m}^{l}$  are from images captured from non-uniform spatial intervals, with the viewing angles varying randomly. Consequently, features from similar views exhibit a higher correlation, while those from distinct views are largely independent. To model this relationship, we construct a "correlation" matrix  $C$  based on the angular disparity obtained from camera rotation matrices of the input multi-view images, and define our MVAttention as follows:
+
+$$
+\begin{array}{l} \mathbf {F} ^ {\mathbf {l}} = \left[ F _ {1} ^ {l} \oplus F _ {2} ^ {l} \dots \oplus F _ {m} ^ {l} \right], \hat {\mathbf {F}} ^ {\mathbf {l}} = \left[ \mathbf {F} ^ {\mathbf {l}} \oplus F _ {f} ^ {l} \oplus F _ {b} ^ {l} \right] \\ Q = W ^ {Q} \mathbf {F} ^ {1}, K = W ^ {K} \hat {\mathbf {F}} ^ {1}, V = W ^ {V} \hat {\mathbf {F}} ^ {1} \tag {2} \\ A _ {i} = \operatorname {s o f t m a x} \left(\frac {Q _ {i} \times \left(C _ {i} \cdot K ^ {T}\right)}{\sqrt {d}}\right), \mathbf {H} _ {\mathbf {i}} ^ {\mathbf {l}} = A _ {i} \times V _ {i} \\ \end{array}
+$$
+
+where  $i\in \{1,2,\dots,m\}$  denotes  $i$  -th view; the Query  $Q$  comes directly from  $\mathbf{F}^1$  and the concatenation of  $[\mathbf{F}^1,F_f^l,F_b^l ]$  serves as the key  $K$  and the value  $V$  .  $\oplus$  indicates matrix concatenation along the token axis;  $d$  denotes the dimension;  $W^{Q},W^{K},W^{V}$  represent the linear transformation matrices; we omitted the  $l$  of the attention matrices and parameters for simplicity;  $C\in \mathbb{R}^{m\times m}$ $C_i$  represents  $i$  -th row in  $C$  , and its "correlation" value between  $i$  -th and  $j$  -th features is  $C_{ij}$  ..
+
+$$
+C _ {i j} = \left(\left(\operatorname {t r a c e} \left(R _ {i} ^ {T} R _ {j}\right) - 1\right) / 2 + 1\right) / 2 \tag {3}
+$$
+
+where  $R_{i}$  and  $R_{j}$  are the extrinsic rotation matrices of the corresponding camera views,  $(\mathrm{trace}(R_i^T R_j) - 1) / 2$  is the cosine value of the angle between these camera views.
+
+Multi-view CLIP Embedding. Camera viewpoints can serve as an effective condition signal to enhance 3D consistency in video content generation [52]. Building on this insight, we incorporate camera condition within our try-on network by encoding camera parameters as an additional token, enabling the generation of more consistent multi-view images. Specifically, we define a world coordinate system in which the camera faces the subject directly. For each input image (view)  $\mathbf{A}_{\mathrm{i}}$ ,  $1 \leq i \leq m$ , we extract the rotation matrix from the camera's corresponding extrinsic matrix. This rotation matrix is then reshaped into a 9-dimensional tensor  $\mathbf{r}_{\mathrm{i}}$ , which undergoes positional encoding to effectively integrate the camera parameters into the feature representation  $F_{i}^{c}$ .
+
+$$
+\begin{array}{l} F _ {i} ^ {c} = \left(\sin \left(2 ^ {0} \pi \mathbf {r} _ {\mathbf {i}}\right), \cos \left(2 ^ {0} \pi \mathbf {r} _ {\mathbf {i}}\right), \dots , \right. \tag {4} \\ \sin (2 ^ {L - 1} \pi \mathbf {r _ {i}}), \cos (2 ^ {L - 1} \pi \mathbf {r _ {i}})) \\ \end{array}
+$$
+
+where  $L$  is the length of positional embedding. We then project  $F_{i}^{c}$  to match the dimensionality of the garment CLIP image embedding  $F^{g}$  via an MLP and concatenate them along the token axis to form  $Y_{i}$ . This combined representation,  $Y_{i}$ , is subsequently used in the key  $K_{\mathrm{x}}$  and value  $V_{\mathrm{x}}$  of the cross-attention layers of the Main UNet:
+
+$$
+Y _ {i} = F ^ {g} \oplus \operatorname {M L P} \left(F _ {i} ^ {c}\right)
+$$
+
+$$
+Q _ {\mathrm {x}} = W _ {\mathrm {x}} ^ {Q} \mathbf {H} _ {\mathrm {i}} ^ {1}, K _ {\mathrm {x}} = W _ {\mathrm {x}} ^ {K} Y _ {i}, V _ {\mathrm {x}} = W _ {\mathrm {x}} ^ {V} Y _ {i} \tag {5}
+$$
+
+$$
+F _ {i} ^ {(l + 1)} = \mathrm {s o f t m a x} (\frac {Q _ {\mathrm {x}} K _ {\mathrm {x}} ^ {T}}{\sqrt {d _ {\mathrm {x}}}}) V _ {\mathrm {x}},
+$$
+
+where  $\mathbf{H}^1$  is the output of the MVAttention of the  $l$ -th layer; we omitted the  $l$  of the cross attention matrices and parameters for simplicity.
+
+Training. Our enhanced multi-view 2D VTON network can be trained by minimizing the following latent diffusion model loss function:
+
+$$
+\mathcal {L} _ {\mathrm {l d m}} = \mathbb {E} _ {z _ {t}, \eta , \psi , \epsilon , t} \left[ \| \epsilon - \epsilon_ {\theta} (z _ {t}, t, \eta , \psi , \zeta)) \| _ {2} ^ {2} \right], \tag {6}
+$$
+
+![](images/66efb9bdfdf378e1471d23c74712a4482299aa59ce3a5ebfeba941baa74bac58.jpg)  
+Figure 4. Illustration of the proposed Multi-view Spatial Attention. Query (Q): multi-view features  $\mathbf{F}^1$ ; Key (K) and Value (V): concatenation of  $\mathbf{F}^1$  and garment features  $F_{f}^{l}$  and  $F_{b}^{l}$ . The attention score between viewpoints  $i$  and  $j$  is modulated by a weight  $C_{ij}$ , determined by the cosine of the angle between them.
+
+where  $\eta = [\mathcal{E}(g_f); \mathcal{E}(g_b); \mathcal{E}(\mathbf{N_i})_{i=1}^m]$  represents the input latent garment images and latent normal maps;  $\zeta = [\mathcal{E}'(\mathbf{A_i})_{i=1}^m]$  denotes the input latent clothing-agnostic images;  $\psi = \mathbf{Y}$  is the proposed multi-view CLIP embedding.
+
+# 5. Experiments
+
+# 5.1. Experimental Setup
+
+Datasets. We conduct experiments on two public datasets: Thuman2.0 [55] and MVHumanNet [51]. Thuman2.0 comprises 526 reconstructed clothed human scans, from which we render multi-view input images. Of these samples, 426 are used for training, while the remaining 100 are set aside for testing. To further evaluate the effectiveness and robustness of our method, we also perform experiments on MVHumanNet, a large-scale dataset of multi-view human images that encompasses a diverse range of subjects, daily outfits and motion sequences. The images in MVHumanNet are captured using a multi-view system with either 48 or 24 cameras. We use 4,990 subjects from this dataset, allocating 4,790 to training and 200 for tests. For each subject, we randomly select two frames of multi-view images from its entire motion sequence. While MVHumanNet provides multi-view images directly for editing and reconstruction, we render uniformly distributed views around each human subject in Thuman2.0 to ensure consistent input.
+
+Baselines. We primarily compare our method with three existing methods: DreamWaltz [23], GaussCtrl [48], and TIP-Editor [63]. DreamWaltz is a method designed for directly generating 3D human bodies based on textual descriptions, while GaussCtrl and TIP-Editor are two radiance-based editing methods. GaussCtrl is based on Stable Diffusion, using a description-like prompt to edit the scene. TIP-Editor accepts both text and image prompts. We configure it by
+
+specifying the human body as the editing region and the desired garment as the image prompt. We use ChatGPT to generate the text prompts corresponding to the clothing images.
+
+Evaluation Metrics. For quantitative evaluation, we assess garment-to-person alignment between the edited person and the reference image. Following [63], we calculate the average DINO similarity [34] between the reference image and the rendered multi-view images of the edited 3D scene. Additionally, to evaluate multi-view consistency, we compute the CLIP Directional Consistency Score as outlined in [17]. Given the large scale of experiments (repeated 3DGS reconstruction), we selected a subset of examples from the dataset for metric evaluation. Specifically, from the test sets of Thuman and MVHumanNet, we randomly sampled 10 human scans each, performing virtual try-on with 6 randomly chosen garments per human scan.
+
+We further conducted a user study involving 50 participants who rated the results of our method and three baseline methods based on two criteria: overall "Quality" and "Alignment" with the reference image. Each evaluation consisted of two questions: (1) Which method produces the highest quality of the edited 3D human? and (2) Which method achieves the most consistent alignment with the target clothing? Participants viewed the VTON results as rotating randomized video sequences.
+
+Implementation Details. During pre-processing, we crop the multi-view images to the bounding box around the person and resize them to a resolution of  $768 \times 576$ . The front view and the back view of garment images are obtained from the corresponding clothed human images. After editing, we pad the images back to their original size. The data processing pipeline is the same for both Thuman2.0 and MVHumanNet datasets.
+
+The Main UNet and the GarmentNet are initialized by the Stable Diffusion V1.5 [39]. The training process is divided into two stages. In the first stage, each view is trained independently, during which we establish the feature extraction capabilities of both the PoseEncoder and GarmentNet, as well as the generative capability of the Main UNet. The second stage involves multi-view training, where we randomly select  $M$  views for each human subject. This stage is focused on training the proposed MVAttention module to enhance the network's ability to maintain consistency across views. Due to memory constraints, we set  $M = 8$  for the training phase. During the testing phase, we uniformly sampled 32 views from a 360-degree rotation around the subject. The editing of these 32 views is conducted in two batches, with each batch processing  $M = 16$  views.
+
+# 5.2. Comparisons with State-of-the-Art Methods
+
+Qualitative Evaluation. Fig. 5 shows visual comparisons between our method and the baselines. DreamWaltz [23]
+
+<table><tr><td rowspan="2">Method</td><td colspan="4">Thuman2.0 [55]</td><td colspan="4">MVHumanNet [51]</td></tr><tr><td>CLIPcons↑</td><td>DINOsim↑</td><td>Votequality</td><td>Votealign</td><td>CLIPcons↑</td><td>DINOsim↑</td><td>Votequality</td><td>Votealign</td></tr><tr><td>DreamWaltz [23]</td><td>0.887</td><td>0.556</td><td>0.46%</td><td>1.54%</td><td>0.935</td><td>0.495</td><td>0.46%</td><td>0.46%</td></tr><tr><td>TIP-Editor [63]</td><td>0.939</td><td>0.569</td><td>0.92%</td><td>0.62%</td><td>0.948</td><td>0.512</td><td>2.15%</td><td>1.38%</td></tr><tr><td>GaussCtrl [48]</td><td>0.931</td><td>0.577</td><td>1.08%</td><td>1.38%</td><td>0.938</td><td>0.521</td><td>1.69%</td><td>1.23%</td></tr><tr><td>Ours</td><td>0.923</td><td>0.633</td><td>97.54%</td><td>96.46%</td><td>0.933</td><td>0.623</td><td>95.69%</td><td>96.92%</td></tr></table>
+
+Table 1. Quantitative comparisons.  ${\mathrm{{CLIP}}}_{\text{cons }}$  denotes the CLIP Direction Consistency Score.  ${\mathrm{{DINO}}}_{\text{sim }}$  is the DINO similarity.
+
+![](images/6b71985790ddfa2ccbcfe297ed7bf36c95d426c4edc7ae3f9d723d3ce1a7d5c3.jpg)  
+Figure 5. Qualitative comparison. The first two rows show the results on Thuman2.0 dataset while the last two rows show the results on MVHumaNet dataset. Our method achieves good texture preservation (highlighted by the blue boxes), while three baseline methods mostly fail.
+
+regenerates 3D clothed humans from text prompts but struggles to accurately retain both body and clothing characteristics. GaussCtrl [48], lacking support for image prompts, fails to maintain detailed clothing textures. While TipEditor [63] leverages Lora [21] for personalization, it encounters difficulties in consistently mapping clothing inputs from two views into the 3D human because the personalized concept are semantic in 2D space. In contrast, our method effectively preserves intricate clothing details, such as text, stripes, and logos.
+
+Quantitative Evaluation. Tab. 1 shows the results for the CLIP Directional Consistency Score and DINO similarity on Thuman2.0 and MVHumanNet datasets. Our approach surpasses other methods on  $\mathrm{DINO}_{sim}$ , clearly illustrating
+
+the superiority of our method in terms of garment texture preservation. While our results on  $\mathrm{CLIP}_{\text {cons }}$  are comparable to those of other methods, it is important to note that these methods incorporate the SDS loss, which to some extent smoothes the representation of humans in 3D space. Additionally, the "flatter" textures of other methods could also result in artificially higher consistency scores. Furthermore, user studies have shown that our method significantly exceeds baselines in terms of edited 3D human quality and the alignment of clothing details.
+
+# 5.3. Visual Results using E-commerce Garment
+
+Fig. 6 showcases VTON results using garments from the MVG dataset [46], whose images are from e-commerce
+
+![](images/7737d81fbc34085b8a538160848be10370bc9d5255819ab87e19ccf3b3907f3f.jpg)  
+Figure 6. Generalization to e-commerce garments (the MVG dataset). Our method, trained on the THuman2.0 dataset, shows strong generalizability when applied to e-commerce garments. For clarity in visualization, we display garment images on human models; however, in the actual VTON process, the garments are segmented from the models using parse maps.
+
+platforms like YOOX NET-A-PORTER, Taobao, and TikTok*, and a model trained on the Thuman2.0 dataset [55]. The results demonstrate that our method effectively preserves intricate garment details and textures. For instance, it accurately maintains the stripe patterns in the first row, the cute tie in the second row, and the buttons in the third row, highlighting the robustness of our approach in handling diverse and realistic clothing items.
+
+# 5.4. Ablation Study
+
+We conduct an ablation study on Thuman2.0 dataset in Tab. 2 and Fig. 7 to evaluate the impact of our three proposed modules in enhancing a typical 2D VTON network with 3D-consistent generation capabilities. Starting with the 2D VTON baseline [53] using DensePose, we progressively replace DensePose with our pseudo-3D pose, incorporate multi-view CLIP embeddings, and ultimately integrate MVAttention in the final configuration. Results in Tab. 2 indicate that each module contributes to metric improvements. Fig. 7 visualizes an example of multi-view image editing. The incorporation of pseudo-3D pose substantially improves limb generation compared to the 2D VTON baseline. Comparing rows 4 and 5, prior to the integration of multi-view CLIP embedding, the model captures limited spatial information, resulting in detail loss at specific angles (columns 3, 4, and 6). Finally, the proposed MVAttention achieves a more coherent generation across views.
+
+![](images/5de364b5b01530170b9e9a96febaa852e03b779b915dc80ff95da76bb7a36c53.jpg)  
+Figure 7. Visualization of the impact of the three proposed techniques on multi-view consistent editing. The red boxes highlight the artifacts. Starting from the 2D VTON baseline, the pseudo-3D pose improves limb generation quality, multi-view CLIP embedding enhances detail across different viewing directions, and finally, MVAttention further strengthens consistency in the generated images.
+
+<table><tr><td>Methods</td><td>CLIPcons↑</td><td>DINOsim↑</td></tr><tr><td>2D-VTON</td><td>0.892</td><td>0.609</td></tr><tr><td>+ Pseudo-3D Pose</td><td>0.910</td><td>0.626</td></tr><tr><td>+ Multi-view CLIP Embedding</td><td>0.913</td><td>0.631</td></tr><tr><td>+ Multi-view Spatial Attention</td><td>0.923</td><td>0.633</td></tr></table>
+
+Table 2. Ablation studies. We ablate the impact of the three proposed techniques on Thuman2.0 dataset.
+
+# 6. Conclusions
+
+In this work, we proposed VTON 360, a novel 3D Virtual Try-On (VTON) method that achieves high-fidelity VTON with the ability to render clothing from arbitrary viewing directions. Our method features a novel formulation of 3D VTON as an extension of 2D VTON that ensures 3D consistent results across multiple views. To bridge the gap between 2D VTON models and 3D consistency requirements, we introduce several key innovations, including multi-view inputs, pseudo-3D pose representation, multi-view spatial attention, and multi-view CLIP embedding. Extensive experiments demonstrate the effectiveness of our approach, significantly outperforming prior 3D VTON techniques in both fidelity and any-view rendering.
+
+# Acknowledgement
+
+This work is supported in part by the National Key R&D Program of China under Grant No.2024YFB3908503, in part by the National Natural Science Foundation of China under Grant NO. 62322608 and in part by the CCF-Kuaishou Large Model Explorer Fund (NO. CCF-KuaiShou 2024007).
+
+# References
+
+[1] Easymocap - make human motion capture easier. Github, 2021.5  
+[2] Shuai Bai, Huiling Zhou, Zhikang Li, Chang Zhou, and Hongxia Yang. Single stage virtual try-on via deformable attention flows. In European Conference on Computer Vision, pages 409-425. Springer, 2022. 2  
+[3] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5420-5430, 2019. 2, 3  
+[4] Robert Bridson, Ronald Fedkiw, and John Anderson. Robust treatment of collisions, contact and friction for cloth animation. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 594-603, 2002. 3  
+[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3  
+[6] Haodong Chen, Yongle Huang, Haojian Huang, Xiangsheng Ge, and Dian Shao. Gaussianvton: 3d human virtual try-on via multi-stage gaussian splatting editing with image prompting. arXiv preprint arXiv:2405.07472, 2024. 2, 3, 1  
+[7] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024. 3  
+[8] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14131-14140, 2021. 1, 2  
+[9] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. arXiv preprint arXiv:2403.05139, 2024. 3, 4, 5  
+[10] Xin Dong, Fuwei Zhao, Zhenyu Xie, Xijin Zhang, Daniel K Du, Min Zheng, Xiang Long, Xiaodan Liang, and Jianchao Yang. Dressing in the wild by watching dance videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3480-3489, 2022. 2
+
+[11] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8485-8493, 2021. 2  
+[12] Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang. Taming the power of diffusion models for high-quality virtual try-on with appearance flow. arXiv preprint arXiv:2308.06101, 2023. 1, 2, 4  
+[13] Peng Guan, Loretta Reiss, David A Hirshberg, Alexander Weiss, and Michael J Black. Drape: Dressing any person. ACM Transactions on Graphics (ToG), 31(4):1-10, 2012. 1, 3  
+[14] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densepose: Dense human pose estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7297-7306, 2018. 4, 5  
+[15] Fabian Hahn, Bernhard Thomaszewski, Stelian Coros, Robert W Sumner, Forrester Cole, Mark Meyer, Tony DeRose, and Markus Gross. Subspace clothing simulation using adaptive bases. ACM Transactions on Graphics (TOG), 33(4):1-9, 2014. 1, 3  
+[16] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 2  
+[17] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19740–19750, 2023. 3, 6, 1  
+[18] Sen He, Yi-Zhe Song, and Tao Xiang. Style-based global appearance flow for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3470-3479, 2022. 2  
+[19] Zijian He, Peixin Chen, Guangrun Wang, Guanbin Li, Philip HS Torr, and Liang Lin. Wildvidfit: Video virtual try-on in the wild via image-based controlled diffusion models. In European Conference on Computer Vision, pages 123-139. Springer, 2024. 1, 2  
+[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3  
+[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 7  
+[22] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023. 5  
+[23] Yukun Huang, Jianan Wang, Ailing Zeng, He Cao, Xianbiao Qi, Yukai Shi, Zheng-Jun Zha, and Lei Zhang. Dreamwaltz: Make a scene with complex 3d animatable avatars. Advances in Neural Information Processing Systems, 36, 2024. 3, 6, 7  
+[24] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided
+
+reconstruction of lifelike clothed humans. In 2024 International Conference on 3D Vision (3DV), pages 1531-1542. IEEE, 2024. 3  
+[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 3, 1  
+[26] Jeongho Kim, Gyojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. arXiv preprint arXiv:2312.01725, 2023. 3, 4, 5  
+[27] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3  
+[28] Zorah Lahner, Daniel Cremers, and Tony Tung. Deepwrinkles: Accurate and realistic clothing modeling. In Proceedings of the European conference on computer vision (ECCV), pages 667-684, 2018. 1, 3  
+[29] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In Proceedings of the European conference on computer vision (ECCV), 2022. 2  
+[30] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851-866. 2023. 3  
+[31] Yifang Men, Yiming Mao, Yuning Jiang, Wei-Ying Ma, and Zhouhui Lian. Controllable person image synthesis with attribute-decomposed gan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5084-5093, 2020. 2  
+[32] Aymen Mir, Thiemo Alldieck, and Gerard Pons-Moll. Learning to transfer texture from clothing images to 3d humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7023-7034, 2020. 2, 3  
+[33] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. Ladi-vton: Latent diffusion textual-inversion enhanced virtual try-on. arXiv preprint arXiv:2305.13501, 2023. 1, 2  
+[34] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6  
+[35] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 5  
+[36] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael J Black. Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics (ToG), 36(4):1-15, 2017. 3  
+[37] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3
+
+[38] Yurui Ren, Xiaoqing Fan, Ge Li, Shan Liu, and Thomas H Li. Neural texture extraction and distribution for controllable person image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13535-13544, 2022. 2  
+[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 3, 4, 6  
+[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3  
+[41] Igor Santesteban, Nils Thuerey, Miguel A Otaduy, and Dan Casas. Self-supervised collision handling via generative 3d garment models for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11763-11773, 2021. 2  
+[42] Igor Santesteban, Miguel Otaduy, Nils Thuerey, and Dan Casas. Ulnef: Untangled layered neural fields for mix-and-match virtual try-on. Advances in Neural Information Processing Systems, 35:12110-12125, 2022. 2  
+[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3  
+[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3  
+[45] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 3  
+[46] Haoyu Wang, Zhilu Zhang, Donglin Di, Shiliang Zhang, and Wangmeng Zuo. Mv-vton: Multi-view virtual try-on with diffusion models. arXiv preprint arXiv:2404.17364, 2024. 7  
+[47] Junjie Wang, Jiemin Fang, Xiaopeng Zhang, Lingxi Xie, and Qi Tian. Gaussianeditor: Editing 3d gaussians delicately with text instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20902-20911, 2024. 3  
+[48] Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. Gaussctrl: multi-view consistent text-driven 3d gaussian splattering editing. arXiv preprint arXiv:2403.08733, 2024. 3, 6, 7  
+[49] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 5  
+[50] Zhenyu Xie, Haoye Dong, Yufei Gao, Zehua Ma, and Xiaodan Liang. Dreamvton: Customizing 3d virtual try-on with personalized diffusion models. arXiv preprint arXiv:2407.16511, 2024. 2, 3  
+[51] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu,
+
+Chongjie Wang, Shijie Wang, et al. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19801-19811, 2024. 2, 6, 7  
+[52] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. 5  
+[53] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. arXiv preprint arXiv:2403.01779, 2024. 3, 4, 8  
+[54] Han Yang, Xinrui Yu, and Ziwei Liu. Full-range virtual try-on with recurrent tri-level transform. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3460-3469, 2022. 2  
+[55] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5746-5756, 2021. 2, 6, 7, 8  
+[56] Jinsong Zhang, Kun Li, Yu-Kun Lai, and Jingyu Yang. Pise: Person image synthesis and editing with decoupled gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7982-7990, 2021. 2  
+[57] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 3  
+[58] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 5  
+[59] Fuwei Zhao, Zhenyu Xie, Michael Kampffmeyer, Haoye Dong, Songfang Han, Tianxiang Zheng, Tao Zhang, and Xiaodan Liang. M3d-vton: A monocular-to-3d virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13239-13249, 2021. 3  
+[60] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4606-4615, 2023. 2  
+[61] Jingyu Zhuang, Chen Wang, Liang Lin, Lingjie Liu, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. In SIGGRAPH Asia 2023 Conference Papers, pages 1-10, 2023. 3  
+[62] Jingyu Zhuang, Di Kang, Linchao Bao, Liang Lin, and Guanbin Li. Dagensm: Disentangled avatar generation with gs-enhanced mesh. arXiv preprint arXiv:2411.15205, 2024.3  
+[63] Jingyu Zhuang, Di Kang, Yan-Pei Cao, Guanbin Li, Liang Lin, and Ying Shan. Tip-editor: An accurate 3d editor following both text-prompts and image-prompts. arXiv preprint arXiv:2401.14828, 2024. 3, 6, 7, 1

2025/VTON 360_ High-Fidelity Virtual Try-On from Any Viewing Direction/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e50aebfd5dd0c87141ae553f24ad8a64757c81dea9d2b42631a40a528571e483
+size 650200

2025/VTON-HandFit_ Virtual Try-on for Arbitrary Hand Pose Guided by Hand Priors Embedding/00462254-bb4b-4c24-9d0a-9a8f2054757a_content_list.json

@@ -0,0 +1,1251 @@
+[
+    {
+        "type": "text",
+        "text": "VTON-HandFit: Virtual Try-on for Arbitrary Hand Pose Guided by Hand Priors Embedding",
+        "text_level": 1,
+        "bbox": [
+            124,
+            128,
+            875,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yujie Liang $^{1,2*}$ , Xiaobin Hu $^{3*}$ , Boyuan Jiang $^{3}$ , Donghao Luo $^{3\\dagger}$ , Xu Peng $^{3}$ , Kai Wu $^{3}$ , Chengming Xu $^{3}$ , Wenhui Han $^{3}$ , Taisong Jin $^{1,2\\dagger}$ , Chengjie Wang $^{3}$ , Rongrong Ji $^{1,2}$ $^{1}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, China.  \n $^{2}$ School of Informatics, Xiamen University, China.  $^{3}$ Tencent Youtu Lab.",
+        "bbox": [
+            179,
+            202,
+            820,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "liangyujie@stu.xmu.edu.cn,{xiaobinhu,byronjiang,michaelluo,ppxupeng, chengmingxu,jasoncjwang}@tencent.com,wukaiwork,outlook.com, whhan18@gmail.com,{jintaisong,rrji}@xmu.edu.cn",
+        "bbox": [
+            197,
+            294,
+            797,
+            345
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            378,
+            326,
+            395
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Although diffusion-based image virtual try-on has made considerable progress, emerging approaches still struggle to effectively address the issue of hand occlusion (i.e., clothing regions occluded by the hand part), leading to a notable degradation of the try-on performance. To tackle this issue widely existing in real-world scenarios, we propose VTON-HandFit, leveraging the power of hand priors to reconstruct the appearance and structure for hand occlusion cases. Firstly, we tailor a Hand-Pose Aggregation Net using the ControlNet-based structure explicitly and adaptively encoding the global hand and pose priors. Besides, to fully exploit the hand-related structure and appearance information, we propose Hand-feature Disentanglement Embedding module to disentangle the hand priors into the hand structure-parametric and visual-appearance features, and customize a masked cross attention for further decoupled feature embedding. Lastly, we customize a hand-canny constraint loss to better learn the structure edge knowledge from the hand template of model image. VTON-HandFit outperforms the baselines in qualitative and quantitative evaluations on the public dataset and our self-collected hand-occlusion Handfit-3K dataset particularly for the arbitrary hand pose occlusion cases in real-world scenarios. Our project page is at: https://vton-handfit.github.io.",
+        "bbox": [
+            88,
+            411,
+            485,
+            773
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            89,
+            799,
+            223,
+            814
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Image-based virtual try-on (VTON) [26] is a widely adopted and highly promising image synthesis technology in the e-commerce industry. Its primary goal is to enhance",
+        "bbox": [
+            89,
+            824,
+            483,
+            869
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "the shopping experience for consumers and minimize advertising expenses for clothing merchants. The VTON task involves generating an image of a human model wearing a specific garment. Over the past few years, countless researchers [9-11, 50] have dedicated significant efforts to achieve more realistic and precise virtual try-on results.",
+        "bbox": [
+            511,
+            380,
+            906,
+            470
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Most recent studies on virtual try-on predominantly utilize generative adversarial networks (GANs) [13] or latent diffusion models (LDMs) [39] for image synthesis. Traditional GAN-based approaches [15-17] often struggle with accurately generating garment folds, realistic lighting and shadows, and lifelike human body representations. Consequently, more recent research has shifted towards LDM-based methods [61], which effectively improve the authenticity of clothed images. However, these researches focus on generating more realistic and natural try-on images while keeping the detailed features of garments. The hands of try-on models are usually strictly on the sides and garments are not occluded by the hands. Such a strict constraint hinders the virtual try-on paving the way toward real-world applications. For instance, as a highly flexible editing task with a wide range of degrees of freedom, digital humans usually meet the occasions that the garments are occluded by the hands, which severely deteriorates the performance of virtual try-on.",
+        "bbox": [
+            511,
+            474,
+            908,
+            762
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Motivated by the above challenge, basically, there mainly exist two solutions to tackle this arbitrary hand problem: 1). perfect hand parsing model to accurately segment the tiny area among fingers. 2). involving the hand-reconstruction into the try-on processing. As shown in Fig. 1, CatVTON[8] use the parsing model to keep the hand part, leading to unavoidably the remaining artifacts and backgrounds from the model image. The existing parsing models fail to accurately segment the small",
+        "bbox": [
+            511,
+            763,
+            910,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* equal contribution, † corresponding author",
+        "bbox": [
+            93,
+            875,
+            333,
+            888
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "22616",
+        "bbox": [
+            478,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/ffb33d5ded6172475685713deedb9529d89316456ef43ff78a83006a4782f928.jpg",
+        "image_caption": [
+            "Figure 1. Comparison of different models on virtual try-on with hand occlusion. The leftmost three images are reference images, showing the target clothing, the model, and a close-up of the model's hand. Our model excels in preserving hand details and achieving realistic clothing transfer, as highlighted in the red boxes. CatVTON utilizes a parsing model to retain the hand segment, resulting in the inevitable persistence of residual artifacts and backgrounds from the model image. IDM-VTON* indicates post-processing using HandRefiner, which fails to address missing hands and mistakenly alters correctly depicted ones."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            109,
+            88,
+            883,
+            404
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "and fragmented finger areas, thus we embrace the latter strategy to involve the hand-reconstruction by disentangling the hand into the appearance and structure knowledge processing in the try-on stage. Then, we propose a novel diffusion-based virtual try-on method, termed as VTON-HandFit, to generate realistic and natural images for arbitrary hand position or hand-occlusion cases with the aid of hand 3D structure-parametric and visual-appearance priors. To embed the pose- and hand-related knowledge, we build a Hand-Pose Net to exploit the body and hand semantic and structure information by getting the body (e.g., DWPose, DensPose) and 3D hand HaMeR-based [61] reconstruction information. Secondly, considering the over-small character of hand that is easily ignored after down-sampling operation, we use the DINO [55] to dig the local appearance and texture knowledge of hand as the hand-visual embedding features. For the structure of hand, we adopt these structure-parametric priors (e.g., 3D MANO-related parameters, handtype and hand-box coordinates) as hand-structure embedding features. After getting structure-parametric and visual-appearance features, we aim to enhance the model's ability to respectively learn hand-related patterns. To achieve this, we inject the hand-structure embedding into the Hand-Pose Net with a primary focus on enhancing the latent features related to hand structure, and the hand-visual embedding into the mainstream U-Net specifically aiming to enhance texture-related features, respec",
+        "bbox": [
+            91,
+            489,
+            483,
+            896
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "tively. The disentanglement learning of structure and visual appearance enhances the network's ability to efficiently and respectively acquire hand-related knowledge. To prevent the interference of hand-related priors with garment features, we impose a constraint that restricts the hand-related features to operate solely within the hand mask area of the cross-attention mechanism. Finally, we tailor a handcanny constraint loss to more effectively acquire the structural edge knowledge from the hand template of the model image. Our contributions are summarized as follows:",
+        "bbox": [
+            509,
+            489,
+            906,
+            641
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose VTON-HandFit, a novel virtual try-on architecture considering to solve the Hand pose occlusion problem popularly existing in real-world scenarios.",
+            "- We propose a Hand-Pose Aggregation Net to encode the hand and pose structure priors and also adaptively to adjust the weight of each component. Meanwhile, hand-canny constraint loss is designed to learn the hand structural-edge knowledge of model images.",
+            "- We propose a novel hand-feature disentanglement embedding module disentangling the hand knowledge into the structure-parametric priors (e.g., MANO parameters) and visual-appearance priors (e.g., color). A mask cross-attention mechanism is proposed to constrain the hand-related features only working on the hand mask area to avoid the garment feature disturbance.",
+            "- We self-collect Handfit-3K test dataset from online retail"
+        ],
+        "bbox": [
+            506,
+            647,
+            908,
+            898
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "22617",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/74037f3537d914e6a0e9804d1cc30bb05894de5764f3a14727b1395c8c65a240.jpg",
+        "image_caption": [
+            "Figure 2. An overview of our VTON-HandFit Network. The network consists of two main components: Hand-feature Disentanglement Embedding and Hand-Pose Aggregation Net. The Hand-feature Disentanglement Embedding module uses the HaMer model to extract hand priors, including hand type  $T_{h}$ , 3D vertices  $V_{h}$ , spatial joint locations  $J_{2d}$ , and joint rotation matrices  $\\theta_{h}$ . These features are processed by the Hand-Struct processor to derive structural features  $c_{struc}$ . Simultaneously, hand images cropped using bounding boxes are processed by DINOv2 and the Hand-Appear processor to obtain visual features  $c_{appear}$ . The structural and visual features are integrated using mask cross attention. The Hand-Pose Aggregation Net module controls body and hand poses by aggregating DWpose, Densepose, and hand depth maps."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            143,
+            89,
+            851,
+            448
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "sites to evaluate hand occlusion performance. Extensive qualitative and quantitative evaluations demonstrate our superiority over state-of-the-art virtual try-on models especially for hand pose occlusion cases, paving the road towards more complicated virtual try-on in real-world.",
+        "bbox": [
+            96,
+            566,
+            483,
+            643
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2. Related Works",
+        "text_level": 1,
+        "bbox": [
+            89,
+            661,
+            240,
+            678
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Image-based virtual try-on. In the context of image-based virtual try-on, the objective is to generate a realistic representation of a target person wearing a specific garment, given a pair of corresponding images. A series of studies [26, 30, 47, 51] have been conducted using Generative Adversarial Networks (GANs) [13]. These studies typically involve the initial deformation of the garment to match the body shape of the person, followed by the utilization of a generator to superimpose the deformed garment onto the person image. Despite efforts made to minimize the discrepancy between the warped garment and the person [12, 21, 26], these approaches often lack the ability to generalize well to diverse person images, particularly those with complex backgrounds or intricate poses. There exists",
+        "bbox": [
+            89,
+            688,
+            483,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "GAN-based try-on method [26] considers occluded problems and designs the warping operation to avoid the garment warping misalignment, which is totally different from diffusion-based algorithms as an inpainting problem.",
+        "bbox": [
+            511,
+            566,
+            906,
+            627
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The remarkable generative capabilities of diffusion models have sparked interest in incorporating them into fashion synthesis, particularly in tasks like visual try-on [2, 4, 5, 22, 27, 43, 59]. TryOnDiffusion [61] employs ParallelUNets to preserve garment details and warp the garment for the try-on process. Subsequent studies have treated virtual try-on as an exemplar-based image inpainting problem [49]. These studies involved fine-tuning the inpainting diffusion models using virtual try-on datasets [24, 32] to generate virtual try-on images of superior quality. For instance, such as LADI-VTON [32] and DCI-VTON [14] have emerged, treating clothing as pseudo-words or employing warping networks to integrate garments into pre-trained diffusion models. IDM-VTON [7] incorporates sophisticated attention modules to encode the high-level semantics of the garment and extract low-level features, thereby preserving fine-grained details. While these diffusion methods address issues related to the nature and realism of synthesized images,",
+        "bbox": [
+            511,
+            628,
+            908,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "22618",
+        "bbox": [
+            478,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "they struggle to meet hard occasions of the hand-based occlusion which very widely exists in the real-world virtual try-on scenarios.",
+        "bbox": [
+            89,
+            90,
+            480,
+            135
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Conditional control of diffusion models. In recent years, latent diffusion models [39] have demonstrated significant success in tasks such as text-to-image generation [19, 25, 37, 40] and image-to-image generation [23, 41]. To achieve more controllable generated results, several studies have introduced conditional control to diffusion models [53, 60]. ControlNet [57] adds spatial conditioning controls (e.g.,anny edges, Hough lines, user scribbles, human key points) to pretrained diffusion models via zero convolutions manner. T2I-Adapter [33] aligns the internal knowledge within T2I models with external control signals, allowing it to be trained based on different conditions. This approach enables the generation of results with rich control and editing effects, particularly in terms of color and structure. In our paper, we focus on the hard occasions of hand-based occlusion in image-based VTON task, and tailor a Hand-Pose Aggregation Net to control the body pose and hand structure. Furthermore, the disengagement knowledge of hand structure and appearance embedding is studied and then these embeddings are respectively injected into the attention mechanism, acting as a mask-based regional control.",
+        "bbox": [
+            91,
+            137,
+            482,
+            452
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Reliable human hand generation. Previous studies [42] have acknowledged the problem of misshaped hands that are generated by existing text-to-image diffusion models. Controlling the hand is considerably more challenging compared to controlling the body pose, primarily due to the relative size of the main subject. Besides, diffusion models exhibit significantly improved capability in generating plausible whole-body poses. This difficulty arises from the intricate nature of learning the physical structure and pose of hands from training images, which encompasses extensive deformations and occlusions. Handrefiner [29] and RHanDS [44] are post-processing solutions adopting a hand mesh reconstruction model to rectify malformed hands. Hand-to-Diffusion [36] uses a two-stage approach, first generating hand effects and then producing the entire image. Hand1000 [56] employs a three-stage training process that integrates hand gesture recognition and optimized text embeddings. VTON-HandFit involves the hand corrections into the generation process not as a post-processing and disentangles the hand priors into the structure and appearance feature embedding for virtual try-on arbitrary hand pose in the real-world.",
+        "bbox": [
+            91,
+            454,
+            482,
+            785
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3. Methodology",
+        "text_level": 1,
+        "bbox": [
+            89,
+            799,
+            225,
+            816
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Data collection of Handfit-3K. Considering that the hand of the current main virtual try-on dataset is separate from the body, there rarely exists a great amount of virtual try-on data for arbitrary hand pose and hard hand-occlusion cases. To meet this urgent need in real-world, we aim to",
+        "bbox": [
+            89,
+            824,
+            480,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "construct a dataset of complex hand poses that occlude the garment. Specifically, we collect an additional dataset of 26,000 images sourced from online retail sites. These images underwent processing using a parsing model and a hand bounding box via HaMeR model [35] to identify instances of occlusion. We devote significant manual effort to inspecting the quality of the data, removing any erroneous or flawed entries, thereby ensuring the high quality of the virtual try-on dataset. As a result, we obtained a specialized hand occlusion test set comprising 3,620 images, termed as Handfit-3K. This dataset facilitates a targeted evaluation of our method's ability to effectively handle occlusions.",
+        "bbox": [
+            511,
+            90,
+            903,
+            272
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1. Preliminary",
+        "text_level": 1,
+        "bbox": [
+            511,
+            281,
+            643,
+            297
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Latent Diffusion Models. Our VTON-HandFit is an extension of stable diffusion model, which is a probabilistic model specifically designed to estimate a data distribution  $\\mathfrak{p}(\\mathbf{x})$  by iteratively reducing noise in a normally distributed variable. Given an input image  $x_0$ , the noise estimation process is:",
+        "bbox": [
+            511,
+            303,
+            903,
+            393
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {n o i s e}} = \\mathbb {E} _ {z \\sim \\mathcal {E} (x), C, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, C) | | _ {2} ^ {2} \\right], \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            521,
+            402,
+            903,
+            422
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where a Variational AutoEncoder  $\\mathcal{E}$  compress input image  $x$  into low-dimension latent space  $z$ . A conditional U-Net denoiser  $\\epsilon_{\\theta}$  aims to estimate noise  $\\theta$  in latent space with the aid of timestep  $t$ ,  $t$ -th noisy latent representation  $z_{t}$ , and other text-prompt conditions  $C$  extracted by text-encoder.",
+        "bbox": [
+            511,
+            430,
+            903,
+            506
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.VTON-HandFit",
+        "text_level": 1,
+        "bbox": [
+            511,
+            513,
+            671,
+            529
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Overview. An overview of our VTON-HandFit is illustrated in Fig. 2. Given a target person image  $x \\in \\mathbb{R}^{H \\times W \\times 3}$  and an input garment image  $g \\in \\mathbb{R}^{H \\times W \\times 3}$ , VTON-HandFit aims to generate a realistic outfitted image  $x_{g} \\in \\mathbb{R}^{H \\times W \\times 3}$ . The garment-agnostic person representation  $x_{m}$  is masked to erase the garment information by the combination of HumanParsing and HumanPose. Thus, this virtual try-on task is transformed into an example-based inpainting problem to recover the missing mask part. For the input of TryonNet, we simply concatenate the noise image  $(z_{t})$  and latent garment-agnostic features  $\\mathcal{E}(x_{m})$  to avoid the other pose-relevant information disturbances. In Hand-Pose Aggregation Net, we also aggregate the pose-related knowledge including the DWPose, DensePose and 3D reconstruction depth map derived by the HaMeR mesh model. These pose-related priors are detached with TryonNet to explicitly be learned in Hand-Pose Net. DensePose provides more bodysemantic information, DWPose contains abundant skeleton structure, and 3D-hand-derived depth maps demonstrate the pixel-level global-aligned hand position and structure knowledge. These modules complement each other by mutually providing information, and working together to constrain the pose and hand generation. In hand-feature disentanglement embedding, we use the HaMeR model for 3D",
+        "bbox": [
+            511,
+            537,
+            903,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "22619",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "hand reconstruction and get the structure-parametric priors. These features are processed by the Hand-Struct processor to derive structural features  $c_{struc}$ . Simultaneously, cropped hand images are processed by DINOv2 and the Hand-Appear processor to obtain visual features  $c_{appear}$ . Besides, GarmentNet aims to extract the garment features into the TryonNet mainly by: 1). encoding the garment image into the latent space  $\\mathcal{E}_g$ , 2). using the CLIP to get the image token features for injection.",
+        "bbox": [
+            89,
+            90,
+            480,
+            226
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Hand-Pose Aggregation Net. To aggregate the hand and pose-relevant priors learning, we tailor a Hand-Pose Net to constrain the semantic and skeleton pose, and global pixel-level hand knowledge. Specifically, dwpose, densepose, and render blocks are designed based on the same zero-multilayer structure with different weights to respectively extract the corresponding features. Zero-multilayer structure consists of seven stacking of conv and SiLU activate function with the zero convolution layer at the last layer.",
+        "bbox": [
+            89,
+            227,
+            480,
+            363
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {F} _ {h p} = \\mathcal {F} _ {d w} + \\mathcal {F} _ {d p} + \\mathcal {F} _ {r h} \\times w _ {\\text {h a n d}}, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            166,
+            375,
+            482,
+            391
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{F}_{dw}$ ,  $\\mathcal{F}_{dp}$ ,  $\\mathcal{F}_{rh}$  are the DWPose feature, Densepose feature, and hand feature after the Zero-multilayer structure, and then the summation is implemented on the pixel-level with the adjustable weight  $w_{\\mathrm{hand}}$  of hand feature  $\\mathcal{F}_{rh}$ . To mitigate the impact of detrimental noise on the hidden states of the neural network layers in the trainable Hand-Pose Net during the initial stages of training, we also adopt the zero convolution learning paradigm of ControlNet.",
+        "bbox": [
+            89,
+            401,
+            482,
+            521
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Hand structural embedding. The HaMeR [35] model is utilized to reconstruct various MANO parameters and other hand-related information from 2D images, including hand mask  $M_h \\in \\mathbb{R}^{768 \\times 1024}$ , hand types  $T_h \\in \\mathbb{R}^{n \\times 1}$ , hand 3D vertices  $V_h \\in \\mathbb{R}^{n \\times 778 \\times 3}$ , spatial joint locations  $J_{2d} \\in \\mathbb{R}^{n \\times 21 \\times 2}$ , and joint rotation matrices  $\\theta \\in \\mathbb{R}^{n \\times 16 \\times 3 \\times 3}$  where  $n$  represents the number of hands. To accommodate variable numbers of hands as input, we pad the hand parameters to a fixed length  $n = 2$  by adding padding at the end. For hand type  $T_h$ , we use 0 to denote left hands, 1 for right hands, and -1 as a filler for images without hands. Similar padding is applied to other parameters to ensure alignment and correspondence across all dimensions. We apply Basis Point Set (BPS) [38] for dimensionality reduction of hand vertices and encode joint rotations as 6D vectors. And then, we import theses information as well as joint rotations and hand types into the hand-structure processor. The structure of hand-structure processor consists of two parallel blocks: three stacked layers of linear and ReLU activation  $(L_r)$  and two stacked layers of linear and layer normalization  $(L_n)$ . These streamlined hand vertices and joint locations are then fed into  $L_r$ . Concurrently, joint rotations and hand types are transformed via  $L_n$ . The outputs from these processes are aggregated to form the Hand Structural Embedding  $c_{struc}$ , encapsulating essential geometric and 3D information for",
+        "bbox": [
+            89,
+            523,
+            482,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "precise hand representation.",
+        "bbox": [
+            513,
+            92,
+            702,
+            106
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nc _ {s t r u c} = \\left[ L _ {r} \\left(\\boldsymbol {V} _ {\\boldsymbol {h}}\\right), L _ {r} \\left(\\boldsymbol {J} _ {\\boldsymbol {2 d}}\\right), L _ {n} \\left(\\boldsymbol {T} _ {\\boldsymbol {h}}\\right), L _ {n} \\left(\\boldsymbol {\\theta} _ {\\boldsymbol {h}}\\right) \\right]. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            550,
+            118,
+            906,
+            135
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Hand appearance embedding. To improve hand visual detail, we extract features  $\\mathcal{F} \\in \\mathbb{R}^{n \\times 1536}$  from the hand Bounding Box regions using DINOv2 [34]. The Hand-Appear processor (HA) uses a linear layer followed by Layer Normalization [1], culminating in the Hand Appearance Embedding  $c_{\\text{appear}}$ . This process ensures the hand's visual fidelity and detail are meticulously preserved in our representation.",
+        "bbox": [
+            511,
+            147,
+            906,
+            267
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nc _ {\\text {a p p e a r}} = \\mathrm {H A} (\\mathcal {F}). \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            643,
+            268,
+            906,
+            284
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Mask-based regional control. Our method introduces a mask-based regional control to enhance hand detail generation and maintain quality in virtual try-on. Utilizing the hand mask  $M_h$ , we modulate Cross-Attention queries, seamlessly integrating Hand Structural Embedding within Hand-Pose Net's Cross-Attention module. Concurrently, Hand Appearance Embedding is adeptly fused into the Try-onNet architecture through mask cross-attention, ensuring precise detail preservation. The mask cross-attention can be represented as:",
+        "bbox": [
+            511,
+            291,
+            906,
+            441
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} Z = \\operatorname {A t t e n t i o n} (Q, K, V, M _ {h}) \\\\ = \\operatorname {S o f t m a x} \\left(\\frac {\\left(Q K ^ {\\top}\\right) \\odot M _ {h}}{\\sqrt {d}}\\right) V, \\tag {5} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            601,
+            455,
+            905,
+            508
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $Q, K$  and  $V$  are the query, key, and value matrices of the attention operation for hand appearance embedding cross-attention,  $M_h$  represents the hand mask.",
+        "bbox": [
+            511,
+            518,
+            905,
+            563
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Hand canny shape constraint. During the training process,  $\\mathcal{L}_{hand}$  is implemented to refine the clarity and precision of hand edges. Specifically, for a small  $R_{t} < T$ , we employ one-step denoising process prior to sampling  $x_0$  for the direct prediction of  $z_0$  from  $z_{t}$  and the model-predicted noise, using Variational Autoencoder for decoder.",
+        "bbox": [
+            511,
+            564,
+            905,
+            654
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nz _ {0} = \\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta}}{\\sqrt {\\bar {\\alpha} _ {t}}}, \\quad t \\leq R _ {t}. \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            596,
+            666,
+            906,
+            699
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Hand regions are extracted from the generated images using bounding box cropping and are processed with the Canny edge detection operator. These regions are then compared to their corresponding areas in the Ground Truth images using Mean Squared Error (MSE) loss. This loss function is specifically designed to sharpen the model's focus on improving edge details and clarity in the hand regions of the generated images.",
+        "bbox": [
+            511,
+            709,
+            905,
+            830
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiment",
+        "text_level": 1,
+        "bbox": [
+            513,
+            844,
+            638,
+            861
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Baselines. To comprehensively evaluate our approach, we selected five state-of-the-art diffusion-based methods in the",
+        "bbox": [
+            511,
+            869,
+            905,
+            898
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "22620",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/87c2702bedbfbc89cfbca50f4f4eb62cfa75b3a211cc8e5d581619b6cc4c3ee6.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            192,
+            90,
+            803,
+            281
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/2331abd65af745254c356932b86a1bea485f1f1ac287398e9bf363e0f013e8f1.jpg",
+        "image_caption": [
+            "(I) Qualitative results on DressCode",
+            "(II) Qualitative results on VITON-HD",
+            "Figure 3. Qualitative comparisons of VTON-HandFit with other methods on public datasets: (I) Dresscode and (II) VITON-HD."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            192,
+            292,
+            803,
+            470
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "virtual try-on domain as baselines for comparison: LaDIVTON [32], StableVTON [24], OOTDiffusion [48], IDM-VTON [7], and CatVTON [54]. While these methods have made substantial contributions to the virtual try-on field, they exhibit certain limitations in handling hand occlusion issues, which motivated us to propose VTON-HandFit.",
+        "bbox": [
+            89,
+            515,
+            482,
+            606
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Dataset. Our experiments utilize two publicly virtual try-on datasets: VITON-HD [6] and DressCode [31]. The VITON-HD dataset consists of 13,679 image pairs, divided into 11,647 for training and 2,032 for testing, featuring front-facing upper-body portraits paired with garments at a resolution of  $1024 \\times 768$ . The DressCode dataset, larger and more varied, contains 48,392 training pairs and 5,400 testing pairs, including full-body portraits and a range of garments (upper-body, lower-body, and dresses) at the same resolution. To enhance our evaluation of the model's capability to address hand occlusion challenges, we build Handfit-3K test with 3,620 images sorted from 26,000 images collected in online retail sites.",
+        "bbox": [
+            89,
+            609,
+            482,
+            806
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation metrics. Following previous works [24, 32], we adopt both paired and unpaired evaluation settings. In the paired setting, the quality of the generated images is assessed using Structural Similarity (SSIM) [45], Learned Perceptual Image Patch Similarity (LPIPS) [58], Frechet Inception Distance (FID) [18], and Kernel Inception Distance",
+        "bbox": [
+            89,
+            810,
+            482,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "(KID) [3]. For the unpaired setting, we employ FID and KID for comparison. Hand generation quality is assessed using FID-H and KID-H for the hand region. Additionally, we use 2D Mean Per Joint Position Error (MPJPE) [20] to evaluate hand joint position accuracy.",
+        "bbox": [
+            511,
+            515,
+            906,
+            590
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation details. Our model is developed using the pre-trained StableDiffusion v1.5 [39]. Detailed body dynamics and structural information are captured using DWpose [52] for pose estimation, Detector2 [46] for Dense-sepose segmentation, and HaMeR for detailed hand depth maps and parameters. We use VITON-HD and DressCode datasets for training. Our training strategy is implemented in two phases. In the first phase, we train the Hand-Pose Aggregation Net and Hand-feature Disentanglement Embedding modules while freezing the other components. In the second phase, we finetune the TryonNet and GarmentNet. All models are trained uniformly with a batch size of 32, employing the AdamW [28] optimizer at a learning rate of 5e-5. Training is conducted on 8 NVIDIA V100 GPUs.",
+        "bbox": [
+            511,
+            592,
+            908,
+            803
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Results on Public Dataset",
+        "text_level": 1,
+        "bbox": [
+            511,
+            816,
+            745,
+            830
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Qualitative results. Fig. 3 shows visual comparisons of the VITON-HD and DressCode datasets between VTON-HandFit and other methods. In previous methods, original hand regions are often reused, leading to artifacts such as",
+        "bbox": [
+            511,
+            839,
+            906,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "22621",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/a0ee4e0f278accf5d87660f54d1f89d64502e161d0a37911326d8051d4ea476d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"3\">Methods</td><td colspan=\"6\">DressCode</td><td colspan=\"6\">VITON-HD</td></tr><tr><td colspan=\"4\">Paired</td><td colspan=\"2\">Unpaired</td><td colspan=\"4\">Paired</td><td colspan=\"2\">Unpaired</td></tr><tr><td>SSIM ↑</td><td>LPIPS ↓</td><td>FID ↓</td><td>KID ↓</td><td>FID ↓</td><td>KID ↓</td><td>SSIM ↑</td><td>LPIPS ↓</td><td>FID ↓</td><td>KID ↓</td><td>FID ↓</td><td>KID ↓</td></tr><tr><td>LaDI-VTON (2023)</td><td>0.9025</td><td>0.0719</td><td>4.8636</td><td>1.5580</td><td>6.8421</td><td>2.3345</td><td>0.8763</td><td>0.0911</td><td>6.6044</td><td>1.0672</td><td>9.4095</td><td>1.6866</td></tr><tr><td>StableVTON (2024)</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>0.8665</td><td>0.0835</td><td>6.8581</td><td>1.2553</td><td>9.5868</td><td>1.4508</td></tr><tr><td>OOTDiffusion (2024)</td><td>0.8975</td><td>0.0725</td><td>3.9497</td><td>0.7198</td><td>6.7019</td><td>1.8630</td><td>0.8513</td><td>0.0964</td><td>6.5186</td><td>0.8961</td><td>9.6733</td><td>1.2061</td></tr><tr><td>IDM-VTON (2024)</td><td>0.9228</td><td>0.0478</td><td>3.8001</td><td>1.2012</td><td>5.6159</td><td>1.5536</td><td>0.8806</td><td>0.0789</td><td>6.3381</td><td>1.3224</td><td>9.6114</td><td>1.6387</td></tr><tr><td>IDM-VTON* (2024)</td><td>0.9007</td><td>0.0624</td><td>4.9170</td><td>2.0176</td><td>6.6867</td><td>2.5793</td><td>0.8316</td><td>0.1058</td><td>7.7491</td><td>2.3381</td><td>9.7712</td><td>1.9323</td></tr><tr><td>CatVTON (2024)</td><td>0.9011</td><td>0.0705</td><td>3.2755</td><td>0.6696</td><td>5.4220</td><td>1.5490</td><td>0.8694</td><td>0.0970</td><td>6.1394</td><td>0.9639</td><td>9.1434</td><td>1.2666</td></tr><tr><td>VTON-HandFit (Ours)</td><td>0.9404</td><td>0.0444</td><td>2.8666</td><td>0.8009</td><td>5.2314</td><td>1.0189</td><td>0.8850</td><td>0.0730</td><td>5.9564</td><td>0.9699</td><td>8.8178</td><td>0.9951</td></tr></table>",
+        "bbox": [
+            91,
+            88,
+            903,
+            243
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Quantitative results on VITON-HD and DressCode datasets. We compare the metrics under both paired (model's clothing is the same as the given cloth image) and unpaired settings (model's clothing differs) with other methods. IDM-VTON* indicates post-processing using HandRefiner.",
+        "bbox": [
+            89,
+            247,
+            906,
+            289
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d9fa97f64b899d23b4070c22fbea7fb4d3ec7188994b33d825325e2d3ed392d4.jpg",
+        "image_caption": [
+            "Figure 4. Qualitative comparisons of VTON-HandFit with other methods on our Handfit-3K dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            150,
+            297,
+            844,
+            513
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "blue outlines around the hands when garment colors significantly differ (third row). In contrast, our model effectively eliminates these residual backgrounds and preserves natural hand poses, enhancing visual fidelity and pose accuracy.",
+        "bbox": [
+            89,
+            545,
+            483,
+            608
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Quantitative results. Tab. 1 presents the quantitative comparisons of VTON-HandFit against other state-of-the-art methods on VITON-HD and DressCode datasets. By enhancing pose accuracy and the fidelity of hand gestures, our approach achieves markedly improved synthesized image quality. Moreover, VTON-HandFit outperforms in FID and KID scores across both datasets in unpaired settings, underscoring the importance of precise hand pose representation for enhancing image quality in practical applications.",
+        "bbox": [
+            89,
+            617,
+            483,
+            755
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Limitations of handrefiner. Despite the strides made in hand reconstruction through HandRefiner [29], as a post-processing operation, HandRefiner falls short in its inability to reconstruct the hands without complete shape and fails to depict the same ones with the original model accurately. These deficiencies not only undermine the model's practicality but also compromise the authenticity of its output, thereby substantially diminishing its performance metrics and overall applicability, as shown in Tab. 1.",
+        "bbox": [
+            89,
+            763,
+            483,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/dfcf5b3138264a54321f89f0ea28afbeb99987b7953a88108154f40c5f9ae89e.jpg",
+        "image_caption": [
+            "Figure 5. Qualitative comparisons in real-world scenarios."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            529,
+            546,
+            890,
+            681
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f6b17a074c0e9c0e6efaf9f8797e2ffd923fadad363685b5bcc2457c536c0594.jpg",
+        "image_caption": [
+            "Figure 6. Qualitative comparisons of VTON-HandFit with variants on VITON-HD dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            527,
+            705,
+            883,
+            768
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2. Results on Handfit-3K Dataset",
+        "text_level": 1,
+        "bbox": [
+            511,
+            816,
+            784,
+            830
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Qualitative results. Fig. 4 presents a qualitative comparison between VTON-HandFit and other diffusion-based methods. Our method outperforms in preserving detailed garment attributes and achieving natural hand poses. The",
+        "bbox": [
+            511,
+            839,
+            906,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "22622",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/6d1d9da32f0888bb5f3c5577af8167412fc47605c96aa3c73fa9eab36246fffc.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"7\">Paired</td><td colspan=\"5\">Unpaired</td></tr><tr><td>SSIM ↑</td><td>LPIPS ↓</td><td>FID ↓</td><td>KID ↓</td><td>MPJPE ↓</td><td>FID-H ↓</td><td>KID-H ↓</td><td>FID ↓</td><td>KID ↓</td><td>MPJPE ↓</td><td>FID-H ↓</td><td>KID-H ↓</td></tr><tr><td>LaDI-VTON (2023)</td><td>0.8597</td><td>0.1480</td><td>12.8430</td><td>4.4112</td><td>107.6315</td><td>23.0199</td><td>17.4558</td><td>16.2506</td><td>7.0065</td><td>111.1547</td><td>21.1371</td><td>14.1960</td></tr><tr><td>StableVTON (2024)</td><td>0.8772</td><td>0.1022</td><td>11.3675</td><td>4.1360</td><td>54.7586</td><td>22.7470</td><td>16.699</td><td>14.0446</td><td>5.1966</td><td>57.3334</td><td>22.8099</td><td>16.3022</td></tr><tr><td>IDM-VTON (2024)</td><td>0.8816</td><td>0.0935</td><td>7.7159</td><td>2.1287</td><td>48.6378</td><td>11.5403</td><td>3.9727</td><td>11.2960</td><td>3.4638</td><td>57.3447</td><td>12.7721</td><td>4.4315</td></tr><tr><td>OOTDiffusion (2024)</td><td>0.8554</td><td>0.1184</td><td>8.7104</td><td>2.1363</td><td>83.9655</td><td>12.1984</td><td>3.6653</td><td>13.1969</td><td>4.4839</td><td>86.1678</td><td>13.8011</td><td>4.4479</td></tr><tr><td>CatVTON (2024)</td><td>0.8761</td><td>0.1123</td><td>10.6966</td><td>4.2581</td><td>60.6608</td><td>13.1264</td><td>4.8723</td><td>15.7081</td><td>6.9544</td><td>63.2245</td><td>15.5603</td><td>5.6134</td></tr><tr><td>VTON-HandFit (Ours)</td><td>0.8827</td><td>0.0819</td><td>5.6027</td><td>0.4932</td><td>25.1376</td><td>7.0574</td><td>1.2474</td><td>10.9786</td><td>3.3248</td><td>26.6892</td><td>9.0003</td><td>2.5362</td></tr></table>",
+        "bbox": [
+            91,
+            88,
+            903,
+            204
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/7ed934ed683f776075d78c6cdfe0ddabd4270fa653ef7d30607c1c6b3db7f1f3.jpg",
+        "table_caption": [
+            "Table 2. Quantitative results on Handfit-3K dataset in the paired and unpaired settings."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>SSIM ↑</td><td>LPIPS ↓</td><td>FID ↓</td><td>KID ↓</td></tr><tr><td>Baseline</td><td>0.8619</td><td>0.0923</td><td>6.9549</td><td>1.2651</td></tr><tr><td>Baseline+HPA(Densepose)</td><td>0.8660</td><td>0.0876</td><td>6.8924</td><td>1.1254</td></tr><tr><td>Baseline+HPA</td><td>0.8738</td><td>0.0812</td><td>6.3390</td><td>0.9502</td></tr><tr><td>Baseline+HPA+Cstruct</td><td>0.8784</td><td>0.0787</td><td>6.2225</td><td>1.1186</td></tr><tr><td>Baseline+HPA+Cstruct+Cappear</td><td>0.8787</td><td>0.0792</td><td>6.0515</td><td>1.0084</td></tr><tr><td>VTON-HandFit</td><td>0.8850</td><td>0.0730</td><td>5.9564</td><td>0.9699</td></tr></table>",
+        "bbox": [
+            91,
+            239,
+            480,
+            330
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Comparison of different variants on VITON-HD dataset.",
+        "bbox": [
+            91,
+            334,
+            480,
+            349
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "images generated by VTON-HandFit demonstrate superior consistency in texture, pattern, and fit relative to baseline methods. Furthermore, the hand poses in our images are more consistent with the original.",
+        "bbox": [
+            89,
+            368,
+            482,
+            428
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Quantitative results. Tab. 2 consolidates the performance of VTON-HandFit against contemporary methods on the Handfit-3K dataset. VTON-HandFit excels in both paired and unpaired settings, showcasing marked advancements in MPJPE, FID-H, and KID-H metrics. This evidences its capability to generate visually compelling images with precise hand pose representation and detailed hand regions. The model's robustness is further affirmed in the unpaired setting, where it consistently leads across all metrics, reinforcing its practical applicability in virtual try-on scenarios with a focus on hand detail fidelity.",
+        "bbox": [
+            88,
+            429,
+            482,
+            594
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Results of real-world scenarios. To validate the performance of our VTON-HandFit in real-world scenarios, we have collected several cases from natural environments, as illustrated in Fig. 5. It is evident that VTON-HandFit continues to perform robustly in real-world settings, with a noticeable enhancement in hand effects.",
+        "bbox": [
+            89,
+            595,
+            482,
+            686
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.3. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            89,
+            696,
+            243,
+            712
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effectiveness of each component. We perform an ablation study to evaluate the contributions of the Hand-Pose Aggregation Net, Hand Structural Embedding  $C_{struct}$ , and Hand Appearance Embedding  $C_{appear}$ . To facilitate a clear comparison, we introduce several model variants to ascertain the effectiveness of each component. The 'Baseline' model excludes all specialized components. The 'Baseline+HPA (Densepose)' variant incorporates only Densepose for pose control. 'Baseline+HPA' utilizes Densepose, Depth, and DWpose for pose control. 'Baseline+HPA+C_{struct}' excludes the appearance embedding. Lastly, 'Baseline+HPA+C_{struct}+C_{appear}' incorpo",
+        "bbox": [
+            89,
+            719,
+            482,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "rates all components. It's important to note that these variants do not include hand-canny constraint loss. Tab. 3 presents a quantitative comparison of our ablation experiments, highlighting the impact of adding different components on the image generation quality. Fig. 6 offers a qualitative analysis, showing the necessity of pose guidance for accurate arm pose replication in the model. By incorporating stronger hand structure and visual guidance, the generation of hands is significantly improved.",
+        "bbox": [
+            511,
+            242,
+            906,
+            378
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            511,
+            393,
+            633,
+            409
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We propose VTON-HandFit, a novel model designed for addressing hand occlusion in virtual try-on scenario. Specifically, we have developed a Hand-Pose Aggregation Net by explicitly and adaptively encoding global hand and pose priors using the ControlNet-based structure. To fully exploit the hand-related structure and appearance information, we propose the Hand-feature Disentanglement Embedding module, which explicitly separates the hand priors into hand structure-parametric and visual-appearance features. Lastly, we have customized a hand-canny constraint loss to enhance the learning of structural edge knowledge from the hand template of the model image. Our approach has been extensively evaluated qualitatively and quantitatively, demonstrating its superiority over state-of-the-art virtual try-on models, particularly in cases involving hand pose occlusion. These results pave the way for more complex virtual try-on applications toward real-world scenarios.",
+        "bbox": [
+            511,
+            419,
+            906,
+            678
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            511,
+            693,
+            671,
+            709
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603) and National Natural Science Foundation of China (No. 62072386), supported by Yunnan Provincial Major S&T Special Plan Project (No. 202402AD080001), Henan Key R&D Project (No. 231111212000), Henan Center for Outstanding Overseas Scientists (No. GZS2022011), Open Foundation of Henan Key Lab of General Aviation Technology (No. ZHKF-230212), Key Lab of Oracle Information Processing of MOE (No. OIP2024E002). We also thank Digital Culture Lab of Tencent SSV for their invaluable contribution during our research.",
+        "bbox": [
+            511,
+            719,
+            908,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "22623",
+        "bbox": [
+            478,
+            944,
+            517,
+            957
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            91,
+            89,
+            187,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5",
+            "[2] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5968-5976, 2023. 3",
+            "[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6",
+            "[4] Shidong Cao, Wenhao Chai, Shengyu Hao, Yanting Zhang, Hangyue Chen, and Gaoang Wang. Diffashion: Reference-based fashion design with structure-aware transfer by diffusion models. IEEE Transactions on Multimedia, 2023. 3",
+            "[5] Mengting Chen, Xi Chen, Zhonghua Zhai, Chen Ju, Xuewen Hong, Jinsong Lan, and Shuai Xiao. Wear-any-way: Manipulable virtual try-on via sparse correspondence alignment. arXiv preprint arXiv:2403.12965, 2024. 3",
+            "[6] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 6",
+            "[7] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. ECCV, 2024. 3, 6",
+            "[8] Zheng Chong, Xiao Dong, Haoxiang Li, Shiyue Zhang, Wenqing Zhang, Xujie Zhang, Hanqing Zhao, and Xiaodan Liang. Catvton: Concatenation is all you need for virtual try-on with diffusion models. arXiv preprint arXiv:2407.15886, 2024. 1",
+            "[9] Morelli Davide, Fincato Matteo, Cornia Marcella, Landi Federico, Cesari Fabio, and Cucchiara Rita. Dress code: High-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1",
+            "[10] Haoye Dong, Xiaodan Liang, Xiaohui Shen, Bochao Wang, Hanjiang Lai, Jia Zhu, Zhiting Hu, and Jian Yin. Towards multi-pose guided virtual try-on network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9026-9035, 2019.",
+            "[11] Chongjian Ge, Yibing Song, Yuying Ge, Han Yang, Wei Liu, and Ping Luo. Disentangled cycle consistency for highly-realistic virtual try-on. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16928-16937, 2021. 1",
+            "[12] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8485-8493, 2021. 3",
+            "[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and"
+        ],
+        "bbox": [
+            93,
+            114,
+            483,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 1, 3",
+            "[14] Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang. Taming the power of diffusion models for high-quality virtual try-on with appearance flow. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7599-7607, 2023. 3",
+            "[15] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 1",
+            "[16] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10471-10480, 2019.",
+            "[17] Sen He, Yi-Zhe Song, and Tao Xiang. Style-based global appearance flow for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3470-3479, 2022. 1",
+            "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6",
+            "[19] Xiaobin Hu, Xu Peng, Donghao Luo, Xiaozhong Ji, Jinlong Peng, Zhengkai Jiang, Jiangning Zhang, Taisong Jin, Chengjie Wang, and Rongrong Ji. Diffumatting: Synthesizing arbitrary objects with matting-level annotation. arXiv preprint arXiv:2403.06168, 2024. 4",
+            "[20] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 6",
+            "[21] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzenes. Do not mask what you do not need to mask: a parser-free virtual try-on. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 619-635. Springer, 2020. 3",
+            "[22] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22623-22633. IEEE, 2023. 3",
+            "[23] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 4",
+            "[24] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8176-8185, 2024. 3, 6",
+            "[25] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF"
+        ],
+        "bbox": [
+            517,
+            92,
+            903,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "22624",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 4",
+            "[26] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In European Conference on Computer Vision, pages 204-219. Springer, 2022. 1, 3",
+            "[27] Yuhan Li, Hao Zhou, Wenxiang Shang, Ran Lin, Xuanhong Chen, and Bingbing Ni. Anyfit: Controllable virtual try-on for any combination of attire across any scenario. arXiv preprint arXiv:2405.18172, 2024. 3",
+            "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6",
+            "[29] Wenquan Lu, Yufei Xu, Jing Zhang, Chaoyue Wang, and Dacheng Tao. Handrefiner: Refining malformed hands in generated images by diffusion-based conditional inpainting. In ACM Multimedia 2024, 2023. 4, 7",
+            "[30] Yifang Men, Yiming Mao, Yuning Jiang, Wei-Ying Ma, and Zhouhui Lian. Controllable person image synthesis with attribute-decomposed gan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5084-5093, 2020. 3",
+            "[31] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: High-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2231-2235, 2022. 6",
+            "[32] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. Ladi-vton: Latent diffusion textual-inversion enhanced virtual try-on. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8580-8589, 2023. 3, 6",
+            "[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 4",
+            "[34] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5",
+            "[35] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 4, 5",
+            "[36] Anton Pelykh, Ozge Mercanoglu Sincan, and Richard Bowden. Giving a hand to diffusion models: a two-stage approach to improving conditional human image generation. arXiv preprint arXiv:2403.10731, 2024. 4",
+            "[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 4"
+        ],
+        "bbox": [
+            91,
+            90,
+            482,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[38] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4332-4341, 2019. 5",
+            "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 4, 6",
+            "[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 4",
+            "[41] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 4",
+            "[42] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4695-4703, 2024. 4",
+            "[43] Ke Sun, Jian Cao, Qi Wang, Linrui Tian, Xindi Zhang, Lian Zhuo, Bang Zhang, Liefeng Bo, Wenbo Zhou, Weiming Zhang, et al. Outfitanyone: Ultra-high quality virtual try-on for any clothing and any person. arXiv preprint arXiv:2407.16224, 2024. 3",
+            "[44] Chengrui Wang, Pengfei Liu, Min Zhou, Ming Zeng, Xubin Li, Tiezheng Ge, et al. Rhand: Refining malformed hands for generated images with decoupled structure and style guidance. arXiv preprint arXiv:2404.13984, 2024. 4",
+            "[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6",
+            "[46] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2. https://github.com/facebookresearch/detectron2, 2019. 6",
+            "[47] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. Gqvton: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23550-23559, 2023. 3",
+            "[48] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. arXiv preprint arXiv:2403.01779, 2024.6",
+            "[49] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 3"
+        ],
+        "bbox": [
+            516,
+            92,
+            906,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "22625",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[50] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wangmeng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating-preserving image content. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7850-7859, 2020. 1",
+            "[51] Zhijing Yang, Junyang Chen, Yukai Shi, Hao Li, Tianshui Chen, and Liang Lin. Occlumix: Towards de-occlusion virtual try-on by semantically-guided mixup. IEEE Transactions on Multimedia, 25:1477–1488, 2023. 3",
+            "[52] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 6",
+            "[53] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023.4",
+            "[54] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and An-An Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8372-8382, 2024. 6",
+            "[55] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. ICLR, 2023. 2",
+            "[56] Haozhuo Zhang, Bin Zhu, Yu Cao, and Yanbin Hao. Hand1000: Generating realistic hands from text with only 1,000 images. arXiv preprint arXiv:2408.15461, 2024. 4",
+            "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4",
+            "[58] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6",
+            "[59] Xujie Zhang, Ente Lin, Xiu Li, Yuxuan Luo, Michael Kampffmeyer, Xin Dong, and Xiaodan Liang. Mmtryon: Multi-modal multi-reference control for high-quality fashion generation. arXiv preprint arXiv:2405.00448, 2024. 3",
+            "[60] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 4",
+            "[61] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4606-4615, 2023. 1, 2, 3"
+        ],
+        "bbox": [
+            91,
+            92,
+            480,
+            838
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "22626",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2025/VTON-HandFit_ Virtual Try-on for Arbitrary Hand Pose Guided by Hand Priors Embedding/00462254-bb4b-4c24-9d0a-9a8f2054757a_model.json

@@ -0,0 +1,1971 @@
+[
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.107,
+                0.003,
+                0.182,
+                0.043
+            ],
+            "angle": 0,
+            "content": "CVF"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.237,
+                0.001,
+                0.812,
+                0.047
+            ],
+            "angle": 0,
+            "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.125,
+                0.13,
+                0.877,
+                0.178
+            ],
+            "angle": 0,
+            "content": "VTON-HandFit: Virtual Try-on for Arbitrary Hand Pose Guided by Hand Priors Embedding"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.18,
+                0.203,
+                0.821,
+                0.293
+            ],
+            "angle": 0,
+            "content": "Yujie Liang\\(^{1,2*}\\), Xiaobin Hu\\(^{3*}\\), Boyuan Jiang\\(^{3}\\), Donghao Luo\\(^{3\\dagger}\\), Xu Peng\\(^{3}\\), Kai Wu\\(^{3}\\), Chengming Xu\\(^{3}\\), Wenhui Han\\(^{3}\\), Taisong Jin\\(^{1,2\\dagger}\\), Chengjie Wang\\(^{3}\\), Rongrong Ji\\(^{1,2}\\)  \n\\(^{1}\\)Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, China.  \n\\(^{2}\\)School of Informatics, Xiamen University, China. \\(^{3}\\)Tencent Youtu Lab."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.198,
+                0.295,
+                0.798,
+                0.346
+            ],
+            "angle": 0,
+            "content": "liangyujie@stu.xmu.edu.cn,{xiaobinhu,byronjiang,michaelluo,ppxupeng, chengmingxu,jasoncjwang}@tencent.com,wukaiwork,outlook.com, whhan18@gmail.com,{jintaisong,rrji}@xmu.edu.cn"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.248,
+                0.38,
+                0.327,
+                0.396
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.412,
+                0.486,
+                0.775
+            ],
+            "angle": 0,
+            "content": "Although diffusion-based image virtual try-on has made considerable progress, emerging approaches still struggle to effectively address the issue of hand occlusion (i.e., clothing regions occluded by the hand part), leading to a notable degradation of the try-on performance. To tackle this issue widely existing in real-world scenarios, we propose VTON-HandFit, leveraging the power of hand priors to reconstruct the appearance and structure for hand occlusion cases. Firstly, we tailor a Hand-Pose Aggregation Net using the ControlNet-based structure explicitly and adaptively encoding the global hand and pose priors. Besides, to fully exploit the hand-related structure and appearance information, we propose Hand-feature Disentanglement Embedding module to disentangle the hand priors into the hand structure-parametric and visual-appearance features, and customize a masked cross attention for further decoupled feature embedding. Lastly, we customize a hand-canny constraint loss to better learn the structure edge knowledge from the hand template of model image. VTON-HandFit outperforms the baselines in qualitative and quantitative evaluations on the public dataset and our self-collected hand-occlusion Handfit-3K dataset particularly for the arbitrary hand pose occlusion cases in real-world scenarios. Our project page is at: https://vton-handfit.github.io."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.8,
+                0.224,
+                0.815
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.825,
+                0.484,
+                0.871
+            ],
+            "angle": 0,
+            "content": "Image-based virtual try-on (VTON) [26] is a widely adopted and highly promising image synthesis technology in the e-commerce industry. Its primary goal is to enhance"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.381,
+                0.907,
+                0.472
+            ],
+            "angle": 0,
+            "content": "the shopping experience for consumers and minimize advertising expenses for clothing merchants. The VTON task involves generating an image of a human model wearing a specific garment. Over the past few years, countless researchers [9-11, 50] have dedicated significant efforts to achieve more realistic and precise virtual try-on results."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.475,
+                0.909,
+                0.763
+            ],
+            "angle": 0,
+            "content": "Most recent studies on virtual try-on predominantly utilize generative adversarial networks (GANs) [13] or latent diffusion models (LDMs) [39] for image synthesis. Traditional GAN-based approaches [15-17] often struggle with accurately generating garment folds, realistic lighting and shadows, and lifelike human body representations. Consequently, more recent research has shifted towards LDM-based methods [61], which effectively improve the authenticity of clothed images. However, these researches focus on generating more realistic and natural try-on images while keeping the detailed features of garments. The hands of try-on models are usually strictly on the sides and garments are not occluded by the hands. Such a strict constraint hinders the virtual try-on paving the way toward real-world applications. For instance, as a highly flexible editing task with a wide range of degrees of freedom, digital humans usually meet the occasions that the garments are occluded by the hands, which severely deteriorates the performance of virtual try-on."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.765,
+                0.911,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Motivated by the above challenge, basically, there mainly exist two solutions to tackle this arbitrary hand problem: 1). perfect hand parsing model to accurately segment the tiny area among fingers. 2). involving the hand-reconstruction into the try-on processing. As shown in Fig. 1, CatVTON[8] use the parsing model to keep the hand part, leading to unavoidably the remaining artifacts and backgrounds from the model image. The existing parsing models fail to accurately segment the small"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.094,
+                0.876,
+                0.334,
+                0.889
+            ],
+            "angle": 0,
+            "content": "* equal contribution, † corresponding author"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.946,
+                0.521,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22616"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.111,
+                0.089,
+                0.885,
+                0.405
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.09,
+                0.406,
+                0.907,
+                0.477
+            ],
+            "angle": 0,
+            "content": "Figure 1. Comparison of different models on virtual try-on with hand occlusion. The leftmost three images are reference images, showing the target clothing, the model, and a close-up of the model's hand. Our model excels in preserving hand details and achieving realistic clothing transfer, as highlighted in the red boxes. CatVTON utilizes a parsing model to retain the hand segment, resulting in the inevitable persistence of residual artifacts and backgrounds from the model image. IDM-VTON* indicates post-processing using HandRefiner, which fails to address missing hands and mistakenly alters correctly depicted ones."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.093,
+                0.49,
+                0.485,
+                0.897
+            ],
+            "angle": 0,
+            "content": "and fragmented finger areas, thus we embrace the latter strategy to involve the hand-reconstruction by disentangling the hand into the appearance and structure knowledge processing in the try-on stage. Then, we propose a novel diffusion-based virtual try-on method, termed as VTON-HandFit, to generate realistic and natural images for arbitrary hand position or hand-occlusion cases with the aid of hand 3D structure-parametric and visual-appearance priors. To embed the pose- and hand-related knowledge, we build a Hand-Pose Net to exploit the body and hand semantic and structure information by getting the body (e.g., DWPose, DensPose) and 3D hand HaMeR-based [61] reconstruction information. Secondly, considering the over-small character of hand that is easily ignored after down-sampling operation, we use the DINO [55] to dig the local appearance and texture knowledge of hand as the hand-visual embedding features. For the structure of hand, we adopt these structure-parametric priors (e.g., 3D MANO-related parameters, handtype and hand-box coordinates) as hand-structure embedding features. After getting structure-parametric and visual-appearance features, we aim to enhance the model's ability to respectively learn hand-related patterns. To achieve this, we inject the hand-structure embedding into the Hand-Pose Net with a primary focus on enhancing the latent features related to hand structure, and the hand-visual embedding into the mainstream U-Net specifically aiming to enhance texture-related features, respec"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.511,
+                0.49,
+                0.907,
+                0.642
+            ],
+            "angle": 0,
+            "content": "tively. The disentanglement learning of structure and visual appearance enhances the network's ability to efficiently and respectively acquire hand-related knowledge. To prevent the interference of hand-related priors with garment features, we impose a constraint that restricts the hand-related features to operate solely within the hand mask area of the cross-attention mechanism. Finally, we tailor a handcanny constraint loss to more effectively acquire the structural edge knowledge from the hand template of the model image. Our contributions are summarized as follows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.648,
+                0.907,
+                0.693
+            ],
+            "angle": 0,
+            "content": "- We propose VTON-HandFit, a novel virtual try-on architecture considering to solve the Hand pose occlusion problem popularly existing in real-world scenarios."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.698,
+                0.907,
+                0.773
+            ],
+            "angle": 0,
+            "content": "- We propose a Hand-Pose Aggregation Net to encode the hand and pose structure priors and also adaptively to adjust the weight of each component. Meanwhile, hand-canny constraint loss is designed to learn the hand structural-edge knowledge of model images."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.777,
+                0.909,
+                0.882
+            ],
+            "angle": 0,
+            "content": "- We propose a novel hand-feature disentanglement embedding module disentangling the hand knowledge into the structure-parametric priors (e.g., MANO parameters) and visual-appearance priors (e.g., color). A mask cross-attention mechanism is proposed to constrain the hand-related features only working on the hand mask area to avoid the garment feature disturbance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.886,
+                0.907,
+                0.9
+            ],
+            "angle": 0,
+            "content": "- We self-collect Handfit-3K test dataset from online retail"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.508,
+                0.648,
+                0.909,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "22617"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.144,
+                0.09,
+                0.852,
+                0.449
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.089,
+                0.456,
+                0.908,
+                0.553
+            ],
+            "angle": 0,
+            "content": "Figure 2. An overview of our VTON-HandFit Network. The network consists of two main components: Hand-feature Disentanglement Embedding and Hand-Pose Aggregation Net. The Hand-feature Disentanglement Embedding module uses the HaMer model to extract hand priors, including hand type \\( T_{h} \\), 3D vertices \\( V_{h} \\), spatial joint locations \\( J_{2d} \\), and joint rotation matrices \\( \\theta_{h} \\). These features are processed by the Hand-Struct processor to derive structural features \\( c_{struc} \\). Simultaneously, hand images cropped using bounding boxes are processed by DINOv2 and the Hand-Appear processor to obtain visual features \\( c_{appear} \\). The structural and visual features are integrated using mask cross attention. The Hand-Pose Aggregation Net module controls body and hand poses by aggregating DWpose, Densepose, and hand depth maps."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.098,
+                0.567,
+                0.484,
+                0.645
+            ],
+            "angle": 0,
+            "content": "sites to evaluate hand occlusion performance. Extensive qualitative and quantitative evaluations demonstrate our superiority over state-of-the-art virtual try-on models especially for hand pose occlusion cases, paving the road towards more complicated virtual try-on in real-world."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.662,
+                0.241,
+                0.679
+            ],
+            "angle": 0,
+            "content": "2. Related Works"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.689,
+                0.484,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Image-based virtual try-on. In the context of image-based virtual try-on, the objective is to generate a realistic representation of a target person wearing a specific garment, given a pair of corresponding images. A series of studies [26, 30, 47, 51] have been conducted using Generative Adversarial Networks (GANs) [13]. These studies typically involve the initial deformation of the garment to match the body shape of the person, followed by the utilization of a generator to superimpose the deformed garment onto the person image. Despite efforts made to minimize the discrepancy between the warped garment and the person [12, 21, 26], these approaches often lack the ability to generalize well to diverse person images, particularly those with complex backgrounds or intricate poses. There exists"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.567,
+                0.907,
+                0.628
+            ],
+            "angle": 0,
+            "content": "GAN-based try-on method [26] considers occluded problems and designs the warping operation to avoid the garment warping misalignment, which is totally different from diffusion-based algorithms as an inpainting problem."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.63,
+                0.909,
+                0.903
+            ],
+            "angle": 0,
+            "content": "The remarkable generative capabilities of diffusion models have sparked interest in incorporating them into fashion synthesis, particularly in tasks like visual try-on [2, 4, 5, 22, 27, 43, 59]. TryOnDiffusion [61] employs ParallelUNets to preserve garment details and warp the garment for the try-on process. Subsequent studies have treated virtual try-on as an exemplar-based image inpainting problem [49]. These studies involved fine-tuning the inpainting diffusion models using virtual try-on datasets [24, 32] to generate virtual try-on images of superior quality. For instance, such as LADI-VTON [32] and DCI-VTON [14] have emerged, treating clothing as pseudo-words or employing warping networks to integrate garments into pre-trained diffusion models. IDM-VTON [7] incorporates sophisticated attention modules to encode the high-level semantics of the garment and extract low-level features, thereby preserving fine-grained details. While these diffusion methods address issues related to the nature and realism of synthesized images,"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.946,
+                0.521,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22618"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.092,
+                0.482,
+                0.136
+            ],
+            "angle": 0,
+            "content": "they struggle to meet hard occasions of the hand-based occlusion which very widely exists in the real-world virtual try-on scenarios."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.093,
+                0.138,
+                0.483,
+                0.453
+            ],
+            "angle": 0,
+            "content": "Conditional control of diffusion models. In recent years, latent diffusion models [39] have demonstrated significant success in tasks such as text-to-image generation [19, 25, 37, 40] and image-to-image generation [23, 41]. To achieve more controllable generated results, several studies have introduced conditional control to diffusion models [53, 60]. ControlNet [57] adds spatial conditioning controls (e.g.,anny edges, Hough lines, user scribbles, human key points) to pretrained diffusion models via zero convolutions manner. T2I-Adapter [33] aligns the internal knowledge within T2I models with external control signals, allowing it to be trained based on different conditions. This approach enables the generation of results with rich control and editing effects, particularly in terms of color and structure. In our paper, we focus on the hard occasions of hand-based occlusion in image-based VTON task, and tailor a Hand-Pose Aggregation Net to control the body pose and hand structure. Furthermore, the disengagement knowledge of hand structure and appearance embedding is studied and then these embeddings are respectively injected into the attention mechanism, acting as a mask-based regional control."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.093,
+                0.455,
+                0.483,
+                0.786
+            ],
+            "angle": 0,
+            "content": "Reliable human hand generation. Previous studies [42] have acknowledged the problem of misshaped hands that are generated by existing text-to-image diffusion models. Controlling the hand is considerably more challenging compared to controlling the body pose, primarily due to the relative size of the main subject. Besides, diffusion models exhibit significantly improved capability in generating plausible whole-body poses. This difficulty arises from the intricate nature of learning the physical structure and pose of hands from training images, which encompasses extensive deformations and occlusions. Handrefiner [29] and RHanDS [44] are post-processing solutions adopting a hand mesh reconstruction model to rectify malformed hands. Hand-to-Diffusion [36] uses a two-stage approach, first generating hand effects and then producing the entire image. Hand1000 [56] employs a three-stage training process that integrates hand gesture recognition and optimized text embeddings. VTON-HandFit involves the hand corrections into the generation process not as a post-processing and disentangles the hand priors into the structure and appearance feature embedding for virtual try-on arbitrary hand pose in the real-world."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.8,
+                0.226,
+                0.817
+            ],
+            "angle": 0,
+            "content": "3. Methodology"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.825,
+                0.482,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Data collection of Handfit-3K. Considering that the hand of the current main virtual try-on dataset is separate from the body, there rarely exists a great amount of virtual try-on data for arbitrary hand pose and hard hand-occlusion cases. To meet this urgent need in real-world, we aim to"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.092,
+                0.905,
+                0.273
+            ],
+            "angle": 0,
+            "content": "construct a dataset of complex hand poses that occlude the garment. Specifically, we collect an additional dataset of 26,000 images sourced from online retail sites. These images underwent processing using a parsing model and a hand bounding box via HaMeR model [35] to identify instances of occlusion. We devote significant manual effort to inspecting the quality of the data, removing any erroneous or flawed entries, thereby ensuring the high quality of the virtual try-on dataset. As a result, we obtained a specialized hand occlusion test set comprising 3,620 images, termed as Handfit-3K. This dataset facilitates a targeted evaluation of our method's ability to effectively handle occlusions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.282,
+                0.644,
+                0.298
+            ],
+            "angle": 0,
+            "content": "3.1. Preliminary"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.304,
+                0.905,
+                0.395
+            ],
+            "angle": 0,
+            "content": "Latent Diffusion Models. Our VTON-HandFit is an extension of stable diffusion model, which is a probabilistic model specifically designed to estimate a data distribution \\( \\mathfrak{p}(\\mathbf{x}) \\) by iteratively reducing noise in a normally distributed variable. Given an input image \\( x_0 \\), the noise estimation process is:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.522,
+                0.404,
+                0.905,
+                0.423
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\text {n o i s e}} = \\mathbb {E} _ {z \\sim \\mathcal {E} (x), C, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, C) | | _ {2} ^ {2} \\right], \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.431,
+                0.905,
+                0.507
+            ],
+            "angle": 0,
+            "content": "where a Variational AutoEncoder \\(\\mathcal{E}\\) compress input image \\(x\\) into low-dimension latent space \\(z\\). A conditional U-Net denoiser \\(\\epsilon_{\\theta}\\) aims to estimate noise \\(\\theta\\) in latent space with the aid of timestep \\(t\\), \\(t\\)-th noisy latent representation \\(z_{t}\\), and other text-prompt conditions \\(C\\) extracted by text-encoder."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.515,
+                0.673,
+                0.53
+            ],
+            "angle": 0,
+            "content": "3.2.VTON-HandFit"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.539,
+                0.905,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Overview. An overview of our VTON-HandFit is illustrated in Fig. 2. Given a target person image \\( x \\in \\mathbb{R}^{H \\times W \\times 3} \\) and an input garment image \\( g \\in \\mathbb{R}^{H \\times W \\times 3} \\), VTON-HandFit aims to generate a realistic outfitted image \\( x_{g} \\in \\mathbb{R}^{H \\times W \\times 3} \\). The garment-agnostic person representation \\( x_{m} \\) is masked to erase the garment information by the combination of HumanParsing and HumanPose. Thus, this virtual try-on task is transformed into an example-based inpainting problem to recover the missing mask part. For the input of TryonNet, we simply concatenate the noise image \\( (z_{t}) \\) and latent garment-agnostic features \\( \\mathcal{E}(x_{m}) \\) to avoid the other pose-relevant information disturbances. In Hand-Pose Aggregation Net, we also aggregate the pose-related knowledge including the DWPose, DensePose and 3D reconstruction depth map derived by the HaMeR mesh model. These pose-related priors are detached with TryonNet to explicitly be learned in Hand-Pose Net. DensePose provides more bodysemantic information, DWPose contains abundant skeleton structure, and 3D-hand-derived depth maps demonstrate the pixel-level global-aligned hand position and structure knowledge. These modules complement each other by mutually providing information, and working together to constrain the pose and hand generation. In hand-feature disentanglement embedding, we use the HaMeR model for 3D"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22619"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.092,
+                0.482,
+                0.227
+            ],
+            "angle": 0,
+            "content": "hand reconstruction and get the structure-parametric priors. These features are processed by the Hand-Struct processor to derive structural features \\( c_{struc} \\). Simultaneously, cropped hand images are processed by DINOv2 and the Hand-Appear processor to obtain visual features \\( c_{appear} \\). Besides, GarmentNet aims to extract the garment features into the TryonNet mainly by: 1). encoding the garment image into the latent space \\( \\mathcal{E}_g \\), 2). using the CLIP to get the image token features for injection."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.228,
+                0.482,
+                0.364
+            ],
+            "angle": 0,
+            "content": "Hand-Pose Aggregation Net. To aggregate the hand and pose-relevant priors learning, we tailor a Hand-Pose Net to constrain the semantic and skeleton pose, and global pixel-level hand knowledge. Specifically, dwpose, densepose, and render blocks are designed based on the same zero-multilayer structure with different weights to respectively extract the corresponding features. Zero-multilayer structure consists of seven stacking of conv and SiLU activate function with the zero convolution layer at the last layer."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.167,
+                0.376,
+                0.483,
+                0.392
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {F} _ {h p} = \\mathcal {F} _ {d w} + \\mathcal {F} _ {d p} + \\mathcal {F} _ {r h} \\times w _ {\\text {h a n d}}, \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.402,
+                0.483,
+                0.522
+            ],
+            "angle": 0,
+            "content": "where \\(\\mathcal{F}_{dw}\\), \\(\\mathcal{F}_{dp}\\), \\(\\mathcal{F}_{rh}\\) are the DWPose feature, Densepose feature, and hand feature after the Zero-multilayer structure, and then the summation is implemented on the pixel-level with the adjustable weight \\(w_{\\mathrm{hand}}\\) of hand feature \\(\\mathcal{F}_{rh}\\). To mitigate the impact of detrimental noise on the hidden states of the neural network layers in the trainable Hand-Pose Net during the initial stages of training, we also adopt the zero convolution learning paradigm of ControlNet."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.524,
+                0.483,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Hand structural embedding. The HaMeR [35] model is utilized to reconstruct various MANO parameters and other hand-related information from 2D images, including hand mask \\( M_h \\in \\mathbb{R}^{768 \\times 1024} \\), hand types \\( T_h \\in \\mathbb{R}^{n \\times 1} \\), hand 3D vertices \\( V_h \\in \\mathbb{R}^{n \\times 778 \\times 3} \\), spatial joint locations \\( J_{2d} \\in \\mathbb{R}^{n \\times 21 \\times 2} \\), and joint rotation matrices \\( \\theta \\in \\mathbb{R}^{n \\times 16 \\times 3 \\times 3} \\) where \\( n \\) represents the number of hands. To accommodate variable numbers of hands as input, we pad the hand parameters to a fixed length \\( n = 2 \\) by adding padding at the end. For hand type \\( T_h \\), we use 0 to denote left hands, 1 for right hands, and -1 as a filler for images without hands. Similar padding is applied to other parameters to ensure alignment and correspondence across all dimensions. We apply Basis Point Set (BPS) [38] for dimensionality reduction of hand vertices and encode joint rotations as 6D vectors. And then, we import theses information as well as joint rotations and hand types into the hand-structure processor. The structure of hand-structure processor consists of two parallel blocks: three stacked layers of linear and ReLU activation \\( (L_r) \\) and two stacked layers of linear and layer normalization \\( (L_n) \\). These streamlined hand vertices and joint locations are then fed into \\( L_r \\). Concurrently, joint rotations and hand types are transformed via \\( L_n \\). The outputs from these processes are aggregated to form the Hand Structural Embedding \\( c_{struc} \\), encapsulating essential geometric and 3D information for"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.093,
+                0.704,
+                0.107
+            ],
+            "angle": 0,
+            "content": "precise hand representation."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.552,
+                0.119,
+                0.907,
+                0.136
+            ],
+            "angle": 0,
+            "content": "\\[\nc _ {s t r u c} = \\left[ L _ {r} \\left(\\boldsymbol {V} _ {\\boldsymbol {h}}\\right), L _ {r} \\left(\\boldsymbol {J} _ {\\boldsymbol {2 d}}\\right), L _ {n} \\left(\\boldsymbol {T} _ {\\boldsymbol {h}}\\right), L _ {n} \\left(\\boldsymbol {\\theta} _ {\\boldsymbol {h}}\\right) \\right]. \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.148,
+                0.907,
+                0.268
+            ],
+            "angle": 0,
+            "content": "Hand appearance embedding. To improve hand visual detail, we extract features \\(\\mathcal{F} \\in \\mathbb{R}^{n \\times 1536}\\) from the hand Bounding Box regions using DINOv2 [34]. The Hand-Appear processor (HA) uses a linear layer followed by Layer Normalization [1], culminating in the Hand Appearance Embedding \\(c_{\\text{appear}}\\). This process ensures the hand's visual fidelity and detail are meticulously preserved in our representation."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.644,
+                0.269,
+                0.907,
+                0.285
+            ],
+            "angle": 0,
+            "content": "\\[\nc _ {\\text {a p p e a r}} = \\mathrm {H A} (\\mathcal {F}). \\tag {4}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.292,
+                0.907,
+                0.443
+            ],
+            "angle": 0,
+            "content": "Mask-based regional control. Our method introduces a mask-based regional control to enhance hand detail generation and maintain quality in virtual try-on. Utilizing the hand mask \\( M_h \\), we modulate Cross-Attention queries, seamlessly integrating Hand Structural Embedding within Hand-Pose Net's Cross-Attention module. Concurrently, Hand Appearance Embedding is adeptly fused into the Try-onNet architecture through mask cross-attention, ensuring precise detail preservation. The mask cross-attention can be represented as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.602,
+                0.456,
+                0.906,
+                0.509
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} Z = \\operatorname {A t t e n t i o n} (Q, K, V, M _ {h}) \\\\ = \\operatorname {S o f t m a x} \\left(\\frac {\\left(Q K ^ {\\top}\\right) \\odot M _ {h}}{\\sqrt {d}}\\right) V, \\tag {5} \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.519,
+                0.906,
+                0.564
+            ],
+            "angle": 0,
+            "content": "where \\( Q, K \\) and \\( V \\) are the query, key, and value matrices of the attention operation for hand appearance embedding cross-attention, \\( M_h \\) represents the hand mask."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.565,
+                0.906,
+                0.655
+            ],
+            "angle": 0,
+            "content": "Hand canny shape constraint. During the training process, \\(\\mathcal{L}_{hand}\\) is implemented to refine the clarity and precision of hand edges. Specifically, for a small \\(R_{t} < T\\), we employ one-step denoising process prior to sampling \\(x_0\\) for the direct prediction of \\(z_0\\) from \\(z_{t}\\) and the model-predicted noise, using Variational Autoencoder for decoder."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.598,
+                0.667,
+                0.907,
+                0.7
+            ],
+            "angle": 0,
+            "content": "\\[\nz _ {0} = \\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta}}{\\sqrt {\\bar {\\alpha} _ {t}}}, \\quad t \\leq R _ {t}. \\tag {6}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.71,
+                0.906,
+                0.831
+            ],
+            "angle": 0,
+            "content": "Hand regions are extracted from the generated images using bounding box cropping and are processed with the Canny edge detection operator. These regions are then compared to their corresponding areas in the Ground Truth images using Mean Squared Error (MSE) loss. This loss function is specifically designed to sharpen the model's focus on improving edge details and clarity in the hand regions of the generated images."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.845,
+                0.64,
+                0.862
+            ],
+            "angle": 0,
+            "content": "4. Experiment"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.871,
+                0.906,
+                0.9
+            ],
+            "angle": 0,
+            "content": "Baselines. To comprehensively evaluate our approach, we selected five state-of-the-art diffusion-based methods in the"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.945,
+                0.52,
+                0.958
+            ],
+            "angle": 0,
+            "content": "22620"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.193,
+                0.092,
+                0.805,
+                0.282
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.415,
+                0.283,
+                0.593,
+                0.294
+            ],
+            "angle": 0,
+            "content": "(I) Qualitative results on DressCode"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.194,
+                0.293,
+                0.804,
+                0.472
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.408,
+                0.473,
+                0.593,
+                0.485
+            ],
+            "angle": 0,
+            "content": "(II) Qualitative results on VITON-HD"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.119,
+                0.489,
+                0.877,
+                0.505
+            ],
+            "angle": 0,
+            "content": "Figure 3. Qualitative comparisons of VTON-HandFit with other methods on public datasets: (I) Dresscode and (II) VITON-HD."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.516,
+                0.483,
+                0.607
+            ],
+            "angle": 0,
+            "content": "virtual try-on domain as baselines for comparison: LaDIVTON [32], StableVTON [24], OOTDiffusion [48], IDM-VTON [7], and CatVTON [54]. While these methods have made substantial contributions to the virtual try-on field, they exhibit certain limitations in handling hand occlusion issues, which motivated us to propose VTON-HandFit."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.611,
+                0.483,
+                0.807
+            ],
+            "angle": 0,
+            "content": "Dataset. Our experiments utilize two publicly virtual try-on datasets: VITON-HD [6] and DressCode [31]. The VITON-HD dataset consists of 13,679 image pairs, divided into 11,647 for training and 2,032 for testing, featuring front-facing upper-body portraits paired with garments at a resolution of \\(1024 \\times 768\\). The DressCode dataset, larger and more varied, contains 48,392 training pairs and 5,400 testing pairs, including full-body portraits and a range of garments (upper-body, lower-body, and dresses) at the same resolution. To enhance our evaluation of the model's capability to address hand occlusion challenges, we build Handfit-3K test with 3,620 images sorted from 26,000 images collected in online retail sites."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.811,
+                0.483,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Evaluation metrics. Following previous works [24, 32], we adopt both paired and unpaired evaluation settings. In the paired setting, the quality of the generated images is assessed using Structural Similarity (SSIM) [45], Learned Perceptual Image Patch Similarity (LPIPS) [58], Frechet Inception Distance (FID) [18], and Kernel Inception Distance"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.516,
+                0.907,
+                0.592
+            ],
+            "angle": 0,
+            "content": "(KID) [3]. For the unpaired setting, we employ FID and KID for comparison. Hand generation quality is assessed using FID-H and KID-H for the hand region. Additionally, we use 2D Mean Per Joint Position Error (MPJPE) [20] to evaluate hand joint position accuracy."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.593,
+                0.909,
+                0.804
+            ],
+            "angle": 0,
+            "content": "Implementation details. Our model is developed using the pre-trained StableDiffusion v1.5 [39]. Detailed body dynamics and structural information are captured using DWpose [52] for pose estimation, Detector2 [46] for Dense-sepose segmentation, and HaMeR for detailed hand depth maps and parameters. We use VITON-HD and DressCode datasets for training. Our training strategy is implemented in two phases. In the first phase, we train the Hand-Pose Aggregation Net and Hand-feature Disentanglement Embedding modules while freezing the other components. In the second phase, we finetune the TryonNet and GarmentNet. All models are trained uniformly with a batch size of 32, employing the AdamW [28] optimizer at a learning rate of 5e-5. Training is conducted on 8 NVIDIA V100 GPUs."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.817,
+                0.746,
+                0.831
+            ],
+            "angle": 0,
+            "content": "4.1. Results on Public Dataset"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.84,
+                0.908,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Qualitative results. Fig. 3 shows visual comparisons of the VITON-HD and DressCode datasets between VTON-HandFit and other methods. In previous methods, original hand regions are often reused, leading to artifacts such as"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.946,
+                0.518,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22621"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.093,
+                0.089,
+                0.904,
+                0.244
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"3\">Methods</td><td colspan=\"6\">DressCode</td><td colspan=\"6\">VITON-HD</td></tr><tr><td colspan=\"4\">Paired</td><td colspan=\"2\">Unpaired</td><td colspan=\"4\">Paired</td><td colspan=\"2\">Unpaired</td></tr><tr><td>SSIM ↑</td><td>LPIPS ↓</td><td>FID ↓</td><td>KID ↓</td><td>FID ↓</td><td>KID ↓</td><td>SSIM ↑</td><td>LPIPS ↓</td><td>FID ↓</td><td>KID ↓</td><td>FID ↓</td><td>KID ↓</td></tr><tr><td>LaDI-VTON (2023)</td><td>0.9025</td><td>0.0719</td><td>4.8636</td><td>1.5580</td><td>6.8421</td><td>2.3345</td><td>0.8763</td><td>0.0911</td><td>6.6044</td><td>1.0672</td><td>9.4095</td><td>1.6866</td></tr><tr><td>StableVTON (2024)</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>0.8665</td><td>0.0835</td><td>6.8581</td><td>1.2553</td><td>9.5868</td><td>1.4508</td></tr><tr><td>OOTDiffusion (2024)</td><td>0.8975</td><td>0.0725</td><td>3.9497</td><td>0.7198</td><td>6.7019</td><td>1.8630</td><td>0.8513</td><td>0.0964</td><td>6.5186</td><td>0.8961</td><td>9.6733</td><td>1.2061</td></tr><tr><td>IDM-VTON (2024)</td><td>0.9228</td><td>0.0478</td><td>3.8001</td><td>1.2012</td><td>5.6159</td><td>1.5536</td><td>0.8806</td><td>0.0789</td><td>6.3381</td><td>1.3224</td><td>9.6114</td><td>1.6387</td></tr><tr><td>IDM-VTON* (2024)</td><td>0.9007</td><td>0.0624</td><td>4.9170</td><td>2.0176</td><td>6.6867</td><td>2.5793</td><td>0.8316</td><td>0.1058</td><td>7.7491</td><td>2.3381</td><td>9.7712</td><td>1.9323</td></tr><tr><td>CatVTON (2024)</td><td>0.9011</td><td>0.0705</td><td>3.2755</td><td>0.6696</td><td>5.4220</td><td>1.5490</td><td>0.8694</td><td>0.0970</td><td>6.1394</td><td>0.9639</td><td>9.1434</td><td>1.2666</td></tr><tr><td>VTON-HandFit (Ours)</td><td>0.9404</td><td>0.0444</td><td>2.8666</td><td>0.8009</td><td>5.2314</td><td>1.0189</td><td>0.8850</td><td>0.0730</td><td>5.9564</td><td>0.9699</td><td>8.8178</td><td>0.9951</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.09,
+                0.248,
+                0.907,
+                0.29
+            ],
+            "angle": 0,
+            "content": "Table 1. Quantitative results on VITON-HD and DressCode datasets. We compare the metrics under both paired (model's clothing is the same as the given cloth image) and unpaired settings (model's clothing differs) with other methods. IDM-VTON* indicates post-processing using HandRefiner."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.151,
+                0.298,
+                0.845,
+                0.515
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.2,
+                0.52,
+                0.796,
+                0.535
+            ],
+            "angle": 0,
+            "content": "Figure 4. Qualitative comparisons of VTON-HandFit with other methods on our Handfit-3K dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.546,
+                0.484,
+                0.609
+            ],
+            "angle": 0,
+            "content": "blue outlines around the hands when garment colors significantly differ (third row). In contrast, our model effectively eliminates these residual backgrounds and preserves natural hand poses, enhancing visual fidelity and pose accuracy."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.618,
+                0.484,
+                0.756
+            ],
+            "angle": 0,
+            "content": "Quantitative results. Tab. 1 presents the quantitative comparisons of VTON-HandFit against other state-of-the-art methods on VITON-HD and DressCode datasets. By enhancing pose accuracy and the fidelity of hand gestures, our approach achieves markedly improved synthesized image quality. Moreover, VTON-HandFit outperforms in FID and KID scores across both datasets in unpaired settings, underscoring the importance of precise hand pose representation for enhancing image quality in practical applications."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.765,
+                0.484,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Limitations of handrefiner. Despite the strides made in hand reconstruction through HandRefiner [29], as a post-processing operation, HandRefiner falls short in its inability to reconstruct the hands without complete shape and fails to depict the same ones with the original model accurately. These deficiencies not only undermine the model's practicality but also compromise the authenticity of its output, thereby substantially diminishing its performance metrics and overall applicability, as shown in Tab. 1."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.53,
+                0.547,
+                0.892,
+                0.683
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.536,
+                0.687,
+                0.882,
+                0.701
+            ],
+            "angle": 0,
+            "content": "Figure 5. Qualitative comparisons in real-world scenarios."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.528,
+                0.707,
+                0.885,
+                0.77
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.513,
+                0.776,
+                0.907,
+                0.804
+            ],
+            "angle": 0,
+            "content": "Figure 6. Qualitative comparisons of VTON-HandFit with variants on VITON-HD dataset."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.817,
+                0.785,
+                0.832
+            ],
+            "angle": 0,
+            "content": "4.2. Results on Handfit-3K Dataset"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.84,
+                0.907,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Qualitative results. Fig. 4 presents a qualitative comparison between VTON-HandFit and other diffusion-based methods. Our method outperforms in preserving detailed garment attributes and achieving natural hand poses. The"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "22622"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.093,
+                0.089,
+                0.905,
+                0.205
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"7\">Paired</td><td colspan=\"5\">Unpaired</td></tr><tr><td>SSIM ↑</td><td>LPIPS ↓</td><td>FID ↓</td><td>KID ↓</td><td>MPJPE ↓</td><td>FID-H ↓</td><td>KID-H ↓</td><td>FID ↓</td><td>KID ↓</td><td>MPJPE ↓</td><td>FID-H ↓</td><td>KID-H ↓</td></tr><tr><td>LaDI-VTON (2023)</td><td>0.8597</td><td>0.1480</td><td>12.8430</td><td>4.4112</td><td>107.6315</td><td>23.0199</td><td>17.4558</td><td>16.2506</td><td>7.0065</td><td>111.1547</td><td>21.1371</td><td>14.1960</td></tr><tr><td>StableVTON (2024)</td><td>0.8772</td><td>0.1022</td><td>11.3675</td><td>4.1360</td><td>54.7586</td><td>22.7470</td><td>16.699</td><td>14.0446</td><td>5.1966</td><td>57.3334</td><td>22.8099</td><td>16.3022</td></tr><tr><td>IDM-VTON (2024)</td><td>0.8816</td><td>0.0935</td><td>7.7159</td><td>2.1287</td><td>48.6378</td><td>11.5403</td><td>3.9727</td><td>11.2960</td><td>3.4638</td><td>57.3447</td><td>12.7721</td><td>4.4315</td></tr><tr><td>OOTDiffusion (2024)</td><td>0.8554</td><td>0.1184</td><td>8.7104</td><td>2.1363</td><td>83.9655</td><td>12.1984</td><td>3.6653</td><td>13.1969</td><td>4.4839</td><td>86.1678</td><td>13.8011</td><td>4.4479</td></tr><tr><td>CatVTON (2024)</td><td>0.8761</td><td>0.1123</td><td>10.6966</td><td>4.2581</td><td>60.6608</td><td>13.1264</td><td>4.8723</td><td>15.7081</td><td>6.9544</td><td>63.2245</td><td>15.5603</td><td>5.6134</td></tr><tr><td>VTON-HandFit (Ours)</td><td>0.8827</td><td>0.0819</td><td>5.6027</td><td>0.4932</td><td>25.1376</td><td>7.0574</td><td>1.2474</td><td>10.9786</td><td>3.3248</td><td>26.6892</td><td>9.0003</td><td>2.5362</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.241,
+                0.209,
+                0.755,
+                0.224
+            ],
+            "angle": 0,
+            "content": "Table 2. Quantitative results on Handfit-3K dataset in the paired and unpaired settings."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.093,
+                0.24,
+                0.481,
+                0.332
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Methods</td><td>SSIM ↑</td><td>LPIPS ↓</td><td>FID ↓</td><td>KID ↓</td></tr><tr><td>Baseline</td><td>0.8619</td><td>0.0923</td><td>6.9549</td><td>1.2651</td></tr><tr><td>Baseline+HPA(Densepose)</td><td>0.8660</td><td>0.0876</td><td>6.8924</td><td>1.1254</td></tr><tr><td>Baseline+HPA</td><td>0.8738</td><td>0.0812</td><td>6.3390</td><td>0.9502</td></tr><tr><td>Baseline+HPA+Cstruct</td><td>0.8784</td><td>0.0787</td><td>6.2225</td><td>1.1186</td></tr><tr><td>Baseline+HPA+Cstruct+Cappear</td><td>0.8787</td><td>0.0792</td><td>6.0515</td><td>1.0084</td></tr><tr><td>VTON-HandFit</td><td>0.8850</td><td>0.0730</td><td>5.9564</td><td>0.9699</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.092,
+                0.335,
+                0.482,
+                0.35
+            ],
+            "angle": 0,
+            "content": "Table 3. Comparison of different variants on VITON-HD dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.369,
+                0.483,
+                0.429
+            ],
+            "angle": 0,
+            "content": "images generated by VTON-HandFit demonstrate superior consistency in texture, pattern, and fit relative to baseline methods. Furthermore, the hand poses in our images are more consistent with the original."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.43,
+                0.483,
+                0.595
+            ],
+            "angle": 0,
+            "content": "Quantitative results. Tab. 2 consolidates the performance of VTON-HandFit against contemporary methods on the Handfit-3K dataset. VTON-HandFit excels in both paired and unpaired settings, showcasing marked advancements in MPJPE, FID-H, and KID-H metrics. This evidences its capability to generate visually compelling images with precise hand pose representation and detailed hand regions. The model's robustness is further affirmed in the unpaired setting, where it consistently leads across all metrics, reinforcing its practical applicability in virtual try-on scenarios with a focus on hand detail fidelity."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.596,
+                0.483,
+                0.687
+            ],
+            "angle": 0,
+            "content": "Results of real-world scenarios. To validate the performance of our VTON-HandFit in real-world scenarios, we have collected several cases from natural environments, as illustrated in Fig. 5. It is evident that VTON-HandFit continues to perform robustly in real-world settings, with a noticeable enhancement in hand effects."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.697,
+                0.245,
+                0.713
+            ],
+            "angle": 0,
+            "content": "4.3. Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.72,
+                0.483,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Effectiveness of each component. We perform an ablation study to evaluate the contributions of the Hand-Pose Aggregation Net, Hand Structural Embedding \\( C_{struct} \\), and Hand Appearance Embedding \\( C_{appear} \\). To facilitate a clear comparison, we introduce several model variants to ascertain the effectiveness of each component. The 'Baseline' model excludes all specialized components. The 'Baseline+HPA (Densepose)' variant incorporates only Densepose for pose control. 'Baseline+HPA' utilizes Densepose, Depth, and DWpose for pose control. 'Baseline+HPA+C_{struct}' excludes the appearance embedding. Lastly, 'Baseline+HPA+C_{struct}+C_{appear}' incorpo"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.243,
+                0.907,
+                0.38
+            ],
+            "angle": 0,
+            "content": "rates all components. It's important to note that these variants do not include hand-canny constraint loss. Tab. 3 presents a quantitative comparison of our ablation experiments, highlighting the impact of adding different components on the image generation quality. Fig. 6 offers a qualitative analysis, showing the necessity of pose guidance for accurate arm pose replication in the model. By incorporating stronger hand structure and visual guidance, the generation of hands is significantly improved."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.395,
+                0.634,
+                0.41
+            ],
+            "angle": 0,
+            "content": "5. Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.42,
+                0.907,
+                0.679
+            ],
+            "angle": 0,
+            "content": "We propose VTON-HandFit, a novel model designed for addressing hand occlusion in virtual try-on scenario. Specifically, we have developed a Hand-Pose Aggregation Net by explicitly and adaptively encoding global hand and pose priors using the ControlNet-based structure. To fully exploit the hand-related structure and appearance information, we propose the Hand-feature Disentanglement Embedding module, which explicitly separates the hand priors into hand structure-parametric and visual-appearance features. Lastly, we have customized a hand-canny constraint loss to enhance the learning of structural edge knowledge from the hand template of the model image. Our approach has been extensively evaluated qualitatively and quantitatively, demonstrating its superiority over state-of-the-art virtual try-on models, particularly in cases involving hand pose occlusion. These results pave the way for more complex virtual try-on applications toward real-world scenarios."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.694,
+                0.673,
+                0.71
+            ],
+            "angle": 0,
+            "content": "Acknowledgments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.72,
+                0.909,
+                0.903
+            ],
+            "angle": 0,
+            "content": "This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603) and National Natural Science Foundation of China (No. 62072386), supported by Yunnan Provincial Major S&T Special Plan Project (No. 202402AD080001), Henan Key R&D Project (No. 231111212000), Henan Center for Outstanding Overseas Scientists (No. GZS2022011), Open Foundation of Henan Key Lab of General Aviation Technology (No. ZHKF-230212), Key Lab of Oracle Information Processing of MOE (No. OIP2024E002). We also thank Digital Culture Lab of Tencent SSV for their invaluable contribution during our research."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.945,
+                0.519,
+                0.958
+            ],
+            "angle": 0,
+            "content": "22623"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.093,
+                0.09,
+                0.188,
+                0.106
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.101,
+                0.115,
+                0.484,
+                0.156
+            ],
+            "angle": 0,
+            "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.101,
+                0.159,
+                0.484,
+                0.242
+            ],
+            "angle": 0,
+            "content": "[2] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5968-5976, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.245,
+                0.483,
+                0.285
+            ],
+            "angle": 0,
+            "content": "[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.288,
+                0.483,
+                0.343
+            ],
+            "angle": 0,
+            "content": "[4] Shidong Cao, Wenhao Chai, Shengyu Hao, Yanting Zhang, Hangyue Chen, and Gaoang Wang. Diffashion: Reference-based fashion design with structure-aware transfer by diffusion models. IEEE Transactions on Multimedia, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.345,
+                0.483,
+                0.4
+            ],
+            "angle": 0,
+            "content": "[5] Mengting Chen, Xi Chen, Zhonghua Zhai, Chen Ju, Xuewen Hong, Jinsong Lan, and Shuai Xiao. Wear-any-way: Manipulable virtual try-on via sparse correspondence alignment. arXiv preprint arXiv:2403.12965, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.403,
+                0.483,
+                0.471
+            ],
+            "angle": 0,
+            "content": "[6] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.474,
+                0.483,
+                0.514
+            ],
+            "angle": 0,
+            "content": "[7] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. ECCV, 2024. 3, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.517,
+                0.483,
+                0.585
+            ],
+            "angle": 0,
+            "content": "[8] Zheng Chong, Xiao Dong, Haoxiang Li, Shiyue Zhang, Wenqing Zhang, Xujie Zhang, Hanqing Zhao, and Xiaodan Liang. Catvton: Concatenation is all you need for virtual try-on with diffusion models. arXiv preprint arXiv:2407.15886, 2024. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.589,
+                0.483,
+                0.657
+            ],
+            "angle": 0,
+            "content": "[9] Morelli Davide, Fincato Matteo, Cornia Marcella, Landi Federico, Cesari Fabio, and Cucchiara Rita. Dress code: High-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.66,
+                0.483,
+                0.728
+            ],
+            "angle": 0,
+            "content": "[10] Haoye Dong, Xiaodan Liang, Xiaohui Shen, Bochao Wang, Hanjiang Lai, Jia Zhu, Zhiting Hu, and Jian Yin. Towards multi-pose guided virtual try-on network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9026-9035, 2019."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.731,
+                0.483,
+                0.799
+            ],
+            "angle": 0,
+            "content": "[11] Chongjian Ge, Yibing Song, Yuying Ge, Han Yang, Wei Liu, and Ping Luo. Disentangled cycle consistency for highly-realistic virtual try-on. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16928-16937, 2021. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.802,
+                0.483,
+                0.87
+            ],
+            "angle": 0,
+            "content": "[12] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8485-8493, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.873,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.094,
+                0.115,
+                0.484,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.548,
+                0.093,
+                0.905,
+                0.12
+            ],
+            "angle": 0,
+            "content": "Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 1, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.123,
+                0.905,
+                0.191
+            ],
+            "angle": 0,
+            "content": "[14] Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang. Taming the power of diffusion models for high-quality virtual try-on with appearance flow. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7599-7607, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.193,
+                0.905,
+                0.248
+            ],
+            "angle": 0,
+            "content": "[15] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.25,
+                0.905,
+                0.305
+            ],
+            "angle": 0,
+            "content": "[16] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10471-10480, 2019."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.307,
+                0.905,
+                0.362
+            ],
+            "angle": 0,
+            "content": "[17] Sen He, Yi-Zhe Song, and Tao Xiang. Style-based global appearance flow for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3470-3479, 2022. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.364,
+                0.905,
+                0.432
+            ],
+            "angle": 0,
+            "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.435,
+                0.905,
+                0.503
+            ],
+            "angle": 0,
+            "content": "[19] Xiaobin Hu, Xu Peng, Donghao Luo, Xiaozhong Ji, Jinlong Peng, Zhengkai Jiang, Jiangning Zhang, Taisong Jin, Chengjie Wang, and Rongrong Ji. Diffumatting: Synthesizing arbitrary objects with matting-level annotation. arXiv preprint arXiv:2403.06168, 2024. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.505,
+                0.905,
+                0.574
+            ],
+            "angle": 0,
+            "content": "[20] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.577,
+                0.905,
+                0.646
+            ],
+            "angle": 0,
+            "content": "[21] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzenes. Do not mask what you do not need to mask: a parser-free virtual try-on. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 619-635. Springer, 2020. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.648,
+                0.905,
+                0.716
+            ],
+            "angle": 0,
+            "content": "[22] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22623-22633. IEEE, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.718,
+                0.905,
+                0.787
+            ],
+            "angle": 0,
+            "content": "[23] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.789,
+                0.905,
+                0.858
+            ],
+            "angle": 0,
+            "content": "[24] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8176-8185, 2024. 3, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.86,
+                0.905,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[25] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.518,
+                0.093,
+                0.905,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22624"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.126,
+                0.092,
+                0.482,
+                0.121
+            ],
+            "angle": 0,
+            "content": "Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.124,
+                0.483,
+                0.193
+            ],
+            "angle": 0,
+            "content": "[26] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In European Conference on Computer Vision, pages 204-219. Springer, 2022. 1, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.195,
+                0.483,
+                0.25
+            ],
+            "angle": 0,
+            "content": "[27] Yuhan Li, Hao Zhou, Wenxiang Shang, Ran Lin, Xuanhong Chen, and Bingbing Ni. Anyfit: Controllable virtual try-on for any combination of attire across any scenario. arXiv preprint arXiv:2405.18172, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.253,
+                0.482,
+                0.281
+            ],
+            "angle": 0,
+            "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.284,
+                0.482,
+                0.338
+            ],
+            "angle": 0,
+            "content": "[29] Wenquan Lu, Yufei Xu, Jing Zhang, Chaoyue Wang, and Dacheng Tao. Handrefiner: Refining malformed hands in generated images by diffusion-based conditional inpainting. In ACM Multimedia 2024, 2023. 4, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.342,
+                0.482,
+                0.411
+            ],
+            "angle": 0,
+            "content": "[30] Yifang Men, Yiming Mao, Yuning Jiang, Wei-Ying Ma, and Zhouhui Lian. Controllable person image synthesis with attribute-decomposed gan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5084-5093, 2020. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.414,
+                0.483,
+                0.483
+            ],
+            "angle": 0,
+            "content": "[31] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: High-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2231-2235, 2022. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.486,
+                0.482,
+                0.555
+            ],
+            "angle": 0,
+            "content": "[32] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. Ladi-vton: Latent diffusion textual-inversion enhanced virtual try-on. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8580-8589, 2023. 3, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.558,
+                0.482,
+                0.627
+            ],
+            "angle": 0,
+            "content": "[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.63,
+                0.482,
+                0.699
+            ],
+            "angle": 0,
+            "content": "[34] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.702,
+                0.483,
+                0.77
+            ],
+            "angle": 0,
+            "content": "[35] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 4, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.774,
+                0.482,
+                0.829
+            ],
+            "angle": 0,
+            "content": "[36] Anton Pelykh, Ozge Mercanoglu Sincan, and Richard Bowden. Giving a hand to diffusion models: a two-stage approach to improving conditional human image generation. arXiv preprint arXiv:2403.10731, 2024. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.832,
+                0.482,
+                0.9
+            ],
+            "angle": 0,
+            "content": "[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 4"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.093,
+                0.092,
+                0.483,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.093,
+                0.905,
+                0.148
+            ],
+            "angle": 0,
+            "content": "[38] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4332-4341, 2019. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.151,
+                0.907,
+                0.219
+            ],
+            "angle": 0,
+            "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 4, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.221,
+                0.905,
+                0.303
+            ],
+            "angle": 0,
+            "content": "[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.306,
+                0.905,
+                0.373
+            ],
+            "angle": 0,
+            "content": "[41] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.376,
+                0.905,
+                0.432
+            ],
+            "angle": 0,
+            "content": "[42] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4695-4703, 2024. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.434,
+                0.905,
+                0.502
+            ],
+            "angle": 0,
+            "content": "[43] Ke Sun, Jian Cao, Qi Wang, Linrui Tian, Xindi Zhang, Lian Zhuo, Bang Zhang, Liefeng Bo, Wenbo Zhou, Weiming Zhang, et al. Outfitanyone: Ultra-high quality virtual try-on for any clothing and any person. arXiv preprint arXiv:2407.16224, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.505,
+                0.905,
+                0.56
+            ],
+            "angle": 0,
+            "content": "[44] Chengrui Wang, Pengfei Liu, Min Zhou, Ming Zeng, Xubin Li, Tiezheng Ge, et al. Rhand: Refining malformed hands for generated images with decoupled structure and style guidance. arXiv preprint arXiv:2404.13984, 2024. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.562,
+                0.905,
+                0.617
+            ],
+            "angle": 0,
+            "content": "[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.619,
+                0.905,
+                0.672
+            ],
+            "angle": 0,
+            "content": "[46] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2. https://github.com/facebookresearch/detectron2, 2019. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.676,
+                0.907,
+                0.759
+            ],
+            "angle": 0,
+            "content": "[47] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. Gqvton: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23550-23559, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.761,
+                0.905,
+                0.815
+            ],
+            "angle": 0,
+            "content": "[48] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. arXiv preprint arXiv:2403.01779, 2024.6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.818,
+                0.905,
+                0.899
+            ],
+            "angle": 0,
+            "content": "[49] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 3"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.517,
+                0.093,
+                0.907,
+                0.899
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.518,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22625"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.093,
+                0.482,
+                0.161
+            ],
+            "angle": 0,
+            "content": "[50] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wangmeng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating-preserving image content. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7850-7859, 2020. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.164,
+                0.482,
+                0.218
+            ],
+            "angle": 0,
+            "content": "[51] Zhijing Yang, Junyang Chen, Yukai Shi, Hao Li, Tianshui Chen, and Liang Lin. Occlumix: Towards de-occlusion virtual try-on by semantically-guided mixup. IEEE Transactions on Multimedia, 25:1477–1488, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.22,
+                0.482,
+                0.275
+            ],
+            "angle": 0,
+            "content": "[52] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.277,
+                0.482,
+                0.33
+            ],
+            "angle": 0,
+            "content": "[53] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023.4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.333,
+                0.482,
+                0.402
+            ],
+            "angle": 0,
+            "content": "[54] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and An-An Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8372-8382, 2024. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.404,
+                0.482,
+                0.458
+            ],
+            "angle": 0,
+            "content": "[55] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. ICLR, 2023. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.461,
+                0.482,
+                0.502
+            ],
+            "angle": 0,
+            "content": "[56] Haozhuo Zhang, Bin Zhu, Yu Cao, and Yanbin Hao. Hand1000: Generating realistic hands from text with only 1,000 images. arXiv preprint arXiv:2408.15461, 2024. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.504,
+                0.482,
+                0.558
+            ],
+            "angle": 0,
+            "content": "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.56,
+                0.482,
+                0.628
+            ],
+            "angle": 0,
+            "content": "[58] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.631,
+                0.482,
+                0.685
+            ],
+            "angle": 0,
+            "content": "[59] Xujie Zhang, Ente Lin, Xiu Li, Yuxuan Luo, Michael Kampffmeyer, Xin Dong, and Xiaodan Liang. Mmtryon: Multi-modal multi-reference control for high-quality fashion generation. arXiv preprint arXiv:2405.00448, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.687,
+                0.482,
+                0.755
+            ],
+            "angle": 0,
+            "content": "[60] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.758,
+                0.482,
+                0.839
+            ],
+            "angle": 0,
+            "content": "[61] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4606-4615, 2023. 1, 2, 3"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.093,
+                0.093,
+                0.482,
+                0.839
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22626"
+        }
+    ]
+]

2025/VTON-HandFit_ Virtual Try-on for Arbitrary Hand Pose Guided by Hand Priors Embedding/00462254-bb4b-4c24-9d0a-9a8f2054757a_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53f9d0e74a3c92be4664106578ce447a14be1065660c7007c9ce14b5ae087aff
+size 7436649

2025/VTON-HandFit_ Virtual Try-on for Arbitrary Hand Pose Guided by Hand Priors Embedding/full.md

@@ -0,0 +1,257 @@
+# VTON-HandFit: Virtual Try-on for Arbitrary Hand Pose Guided by Hand Priors Embedding
+
+Yujie Liang $^{1,2*}$ , Xiaobin Hu $^{3*}$ , Boyuan Jiang $^{3}$ , Donghao Luo $^{3\dagger}$ , Xu Peng $^{3}$ , Kai Wu $^{3}$ , Chengming Xu $^{3}$ , Wenhui Han $^{3}$ , Taisong Jin $^{1,2\dagger}$ , Chengjie Wang $^{3}$ , Rongrong Ji $^{1,2}$ $^{1}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, China.  
+ $^{2}$ School of Informatics, Xiamen University, China.  $^{3}$ Tencent Youtu Lab.
+
+liangyujie@stu.xmu.edu.cn,{xiaobinhu,byronjiang,michaelluo,ppxupeng, chengmingxu,jasoncjwang}@tencent.com,wukaiwork,outlook.com, whhan18@gmail.com,{jintaisong,rrji}@xmu.edu.cn
+
+# Abstract
+
+Although diffusion-based image virtual try-on has made considerable progress, emerging approaches still struggle to effectively address the issue of hand occlusion (i.e., clothing regions occluded by the hand part), leading to a notable degradation of the try-on performance. To tackle this issue widely existing in real-world scenarios, we propose VTON-HandFit, leveraging the power of hand priors to reconstruct the appearance and structure for hand occlusion cases. Firstly, we tailor a Hand-Pose Aggregation Net using the ControlNet-based structure explicitly and adaptively encoding the global hand and pose priors. Besides, to fully exploit the hand-related structure and appearance information, we propose Hand-feature Disentanglement Embedding module to disentangle the hand priors into the hand structure-parametric and visual-appearance features, and customize a masked cross attention for further decoupled feature embedding. Lastly, we customize a hand-canny constraint loss to better learn the structure edge knowledge from the hand template of model image. VTON-HandFit outperforms the baselines in qualitative and quantitative evaluations on the public dataset and our self-collected hand-occlusion Handfit-3K dataset particularly for the arbitrary hand pose occlusion cases in real-world scenarios. Our project page is at: https://vton-handfit.github.io.
+
+# 1. Introduction
+
+Image-based virtual try-on (VTON) [26] is a widely adopted and highly promising image synthesis technology in the e-commerce industry. Its primary goal is to enhance
+
+the shopping experience for consumers and minimize advertising expenses for clothing merchants. The VTON task involves generating an image of a human model wearing a specific garment. Over the past few years, countless researchers [9-11, 50] have dedicated significant efforts to achieve more realistic and precise virtual try-on results.
+
+Most recent studies on virtual try-on predominantly utilize generative adversarial networks (GANs) [13] or latent diffusion models (LDMs) [39] for image synthesis. Traditional GAN-based approaches [15-17] often struggle with accurately generating garment folds, realistic lighting and shadows, and lifelike human body representations. Consequently, more recent research has shifted towards LDM-based methods [61], which effectively improve the authenticity of clothed images. However, these researches focus on generating more realistic and natural try-on images while keeping the detailed features of garments. The hands of try-on models are usually strictly on the sides and garments are not occluded by the hands. Such a strict constraint hinders the virtual try-on paving the way toward real-world applications. For instance, as a highly flexible editing task with a wide range of degrees of freedom, digital humans usually meet the occasions that the garments are occluded by the hands, which severely deteriorates the performance of virtual try-on.
+
+Motivated by the above challenge, basically, there mainly exist two solutions to tackle this arbitrary hand problem: 1). perfect hand parsing model to accurately segment the tiny area among fingers. 2). involving the hand-reconstruction into the try-on processing. As shown in Fig. 1, CatVTON[8] use the parsing model to keep the hand part, leading to unavoidably the remaining artifacts and backgrounds from the model image. The existing parsing models fail to accurately segment the small
+
+![](images/ffb33d5ded6172475685713deedb9529d89316456ef43ff78a83006a4782f928.jpg)  
+Figure 1. Comparison of different models on virtual try-on with hand occlusion. The leftmost three images are reference images, showing the target clothing, the model, and a close-up of the model's hand. Our model excels in preserving hand details and achieving realistic clothing transfer, as highlighted in the red boxes. CatVTON utilizes a parsing model to retain the hand segment, resulting in the inevitable persistence of residual artifacts and backgrounds from the model image. IDM-VTON* indicates post-processing using HandRefiner, which fails to address missing hands and mistakenly alters correctly depicted ones.
+
+and fragmented finger areas, thus we embrace the latter strategy to involve the hand-reconstruction by disentangling the hand into the appearance and structure knowledge processing in the try-on stage. Then, we propose a novel diffusion-based virtual try-on method, termed as VTON-HandFit, to generate realistic and natural images for arbitrary hand position or hand-occlusion cases with the aid of hand 3D structure-parametric and visual-appearance priors. To embed the pose- and hand-related knowledge, we build a Hand-Pose Net to exploit the body and hand semantic and structure information by getting the body (e.g., DWPose, DensPose) and 3D hand HaMeR-based [61] reconstruction information. Secondly, considering the over-small character of hand that is easily ignored after down-sampling operation, we use the DINO [55] to dig the local appearance and texture knowledge of hand as the hand-visual embedding features. For the structure of hand, we adopt these structure-parametric priors (e.g., 3D MANO-related parameters, handtype and hand-box coordinates) as hand-structure embedding features. After getting structure-parametric and visual-appearance features, we aim to enhance the model's ability to respectively learn hand-related patterns. To achieve this, we inject the hand-structure embedding into the Hand-Pose Net with a primary focus on enhancing the latent features related to hand structure, and the hand-visual embedding into the mainstream U-Net specifically aiming to enhance texture-related features, respec
+
+tively. The disentanglement learning of structure and visual appearance enhances the network's ability to efficiently and respectively acquire hand-related knowledge. To prevent the interference of hand-related priors with garment features, we impose a constraint that restricts the hand-related features to operate solely within the hand mask area of the cross-attention mechanism. Finally, we tailor a handcanny constraint loss to more effectively acquire the structural edge knowledge from the hand template of the model image. Our contributions are summarized as follows:
+
+- We propose VTON-HandFit, a novel virtual try-on architecture considering to solve the Hand pose occlusion problem popularly existing in real-world scenarios.  
+- We propose a Hand-Pose Aggregation Net to encode the hand and pose structure priors and also adaptively to adjust the weight of each component. Meanwhile, hand-canny constraint loss is designed to learn the hand structural-edge knowledge of model images.  
+- We propose a novel hand-feature disentanglement embedding module disentangling the hand knowledge into the structure-parametric priors (e.g., MANO parameters) and visual-appearance priors (e.g., color). A mask cross-attention mechanism is proposed to constrain the hand-related features only working on the hand mask area to avoid the garment feature disturbance.  
+- We self-collect Handfit-3K test dataset from online retail
+
+![](images/74037f3537d914e6a0e9804d1cc30bb05894de5764f3a14727b1395c8c65a240.jpg)  
+Figure 2. An overview of our VTON-HandFit Network. The network consists of two main components: Hand-feature Disentanglement Embedding and Hand-Pose Aggregation Net. The Hand-feature Disentanglement Embedding module uses the HaMer model to extract hand priors, including hand type  $T_{h}$ , 3D vertices  $V_{h}$ , spatial joint locations  $J_{2d}$ , and joint rotation matrices  $\theta_{h}$ . These features are processed by the Hand-Struct processor to derive structural features  $c_{struc}$ . Simultaneously, hand images cropped using bounding boxes are processed by DINOv2 and the Hand-Appear processor to obtain visual features  $c_{appear}$ . The structural and visual features are integrated using mask cross attention. The Hand-Pose Aggregation Net module controls body and hand poses by aggregating DWpose, Densepose, and hand depth maps.
+
+sites to evaluate hand occlusion performance. Extensive qualitative and quantitative evaluations demonstrate our superiority over state-of-the-art virtual try-on models especially for hand pose occlusion cases, paving the road towards more complicated virtual try-on in real-world.
+
+# 2. Related Works
+
+Image-based virtual try-on. In the context of image-based virtual try-on, the objective is to generate a realistic representation of a target person wearing a specific garment, given a pair of corresponding images. A series of studies [26, 30, 47, 51] have been conducted using Generative Adversarial Networks (GANs) [13]. These studies typically involve the initial deformation of the garment to match the body shape of the person, followed by the utilization of a generator to superimpose the deformed garment onto the person image. Despite efforts made to minimize the discrepancy between the warped garment and the person [12, 21, 26], these approaches often lack the ability to generalize well to diverse person images, particularly those with complex backgrounds or intricate poses. There exists
+
+GAN-based try-on method [26] considers occluded problems and designs the warping operation to avoid the garment warping misalignment, which is totally different from diffusion-based algorithms as an inpainting problem.
+
+The remarkable generative capabilities of diffusion models have sparked interest in incorporating them into fashion synthesis, particularly in tasks like visual try-on [2, 4, 5, 22, 27, 43, 59]. TryOnDiffusion [61] employs ParallelUNets to preserve garment details and warp the garment for the try-on process. Subsequent studies have treated virtual try-on as an exemplar-based image inpainting problem [49]. These studies involved fine-tuning the inpainting diffusion models using virtual try-on datasets [24, 32] to generate virtual try-on images of superior quality. For instance, such as LADI-VTON [32] and DCI-VTON [14] have emerged, treating clothing as pseudo-words or employing warping networks to integrate garments into pre-trained diffusion models. IDM-VTON [7] incorporates sophisticated attention modules to encode the high-level semantics of the garment and extract low-level features, thereby preserving fine-grained details. While these diffusion methods address issues related to the nature and realism of synthesized images,
+
+they struggle to meet hard occasions of the hand-based occlusion which very widely exists in the real-world virtual try-on scenarios.
+
+Conditional control of diffusion models. In recent years, latent diffusion models [39] have demonstrated significant success in tasks such as text-to-image generation [19, 25, 37, 40] and image-to-image generation [23, 41]. To achieve more controllable generated results, several studies have introduced conditional control to diffusion models [53, 60]. ControlNet [57] adds spatial conditioning controls (e.g.,anny edges, Hough lines, user scribbles, human key points) to pretrained diffusion models via zero convolutions manner. T2I-Adapter [33] aligns the internal knowledge within T2I models with external control signals, allowing it to be trained based on different conditions. This approach enables the generation of results with rich control and editing effects, particularly in terms of color and structure. In our paper, we focus on the hard occasions of hand-based occlusion in image-based VTON task, and tailor a Hand-Pose Aggregation Net to control the body pose and hand structure. Furthermore, the disengagement knowledge of hand structure and appearance embedding is studied and then these embeddings are respectively injected into the attention mechanism, acting as a mask-based regional control.
+
+Reliable human hand generation. Previous studies [42] have acknowledged the problem of misshaped hands that are generated by existing text-to-image diffusion models. Controlling the hand is considerably more challenging compared to controlling the body pose, primarily due to the relative size of the main subject. Besides, diffusion models exhibit significantly improved capability in generating plausible whole-body poses. This difficulty arises from the intricate nature of learning the physical structure and pose of hands from training images, which encompasses extensive deformations and occlusions. Handrefiner [29] and RHanDS [44] are post-processing solutions adopting a hand mesh reconstruction model to rectify malformed hands. Hand-to-Diffusion [36] uses a two-stage approach, first generating hand effects and then producing the entire image. Hand1000 [56] employs a three-stage training process that integrates hand gesture recognition and optimized text embeddings. VTON-HandFit involves the hand corrections into the generation process not as a post-processing and disentangles the hand priors into the structure and appearance feature embedding for virtual try-on arbitrary hand pose in the real-world.
+
+# 3. Methodology
+
+Data collection of Handfit-3K. Considering that the hand of the current main virtual try-on dataset is separate from the body, there rarely exists a great amount of virtual try-on data for arbitrary hand pose and hard hand-occlusion cases. To meet this urgent need in real-world, we aim to
+
+construct a dataset of complex hand poses that occlude the garment. Specifically, we collect an additional dataset of 26,000 images sourced from online retail sites. These images underwent processing using a parsing model and a hand bounding box via HaMeR model [35] to identify instances of occlusion. We devote significant manual effort to inspecting the quality of the data, removing any erroneous or flawed entries, thereby ensuring the high quality of the virtual try-on dataset. As a result, we obtained a specialized hand occlusion test set comprising 3,620 images, termed as Handfit-3K. This dataset facilitates a targeted evaluation of our method's ability to effectively handle occlusions.
+
+# 3.1. Preliminary
+
+Latent Diffusion Models. Our VTON-HandFit is an extension of stable diffusion model, which is a probabilistic model specifically designed to estimate a data distribution  $\mathfrak{p}(\mathbf{x})$  by iteratively reducing noise in a normally distributed variable. Given an input image  $x_0$ , the noise estimation process is:
+
+$$
+\mathcal {L} _ {\text {n o i s e}} = \mathbb {E} _ {z \sim \mathcal {E} (x), C, \epsilon \sim \mathcal {N} (0, 1), t} \left[ | | \epsilon - \epsilon_ {\theta} (z _ {t}, t, C) | | _ {2} ^ {2} \right], \tag {1}
+$$
+
+where a Variational AutoEncoder  $\mathcal{E}$  compress input image  $x$  into low-dimension latent space  $z$ . A conditional U-Net denoiser  $\epsilon_{\theta}$  aims to estimate noise  $\theta$  in latent space with the aid of timestep  $t$ ,  $t$ -th noisy latent representation  $z_{t}$ , and other text-prompt conditions  $C$  extracted by text-encoder.
+
+# 3.2.VTON-HandFit
+
+Overview. An overview of our VTON-HandFit is illustrated in Fig. 2. Given a target person image  $x \in \mathbb{R}^{H \times W \times 3}$  and an input garment image  $g \in \mathbb{R}^{H \times W \times 3}$ , VTON-HandFit aims to generate a realistic outfitted image  $x_{g} \in \mathbb{R}^{H \times W \times 3}$ . The garment-agnostic person representation  $x_{m}$  is masked to erase the garment information by the combination of HumanParsing and HumanPose. Thus, this virtual try-on task is transformed into an example-based inpainting problem to recover the missing mask part. For the input of TryonNet, we simply concatenate the noise image  $(z_{t})$  and latent garment-agnostic features  $\mathcal{E}(x_{m})$  to avoid the other pose-relevant information disturbances. In Hand-Pose Aggregation Net, we also aggregate the pose-related knowledge including the DWPose, DensePose and 3D reconstruction depth map derived by the HaMeR mesh model. These pose-related priors are detached with TryonNet to explicitly be learned in Hand-Pose Net. DensePose provides more bodysemantic information, DWPose contains abundant skeleton structure, and 3D-hand-derived depth maps demonstrate the pixel-level global-aligned hand position and structure knowledge. These modules complement each other by mutually providing information, and working together to constrain the pose and hand generation. In hand-feature disentanglement embedding, we use the HaMeR model for 3D
+
+hand reconstruction and get the structure-parametric priors. These features are processed by the Hand-Struct processor to derive structural features  $c_{struc}$ . Simultaneously, cropped hand images are processed by DINOv2 and the Hand-Appear processor to obtain visual features  $c_{appear}$ . Besides, GarmentNet aims to extract the garment features into the TryonNet mainly by: 1). encoding the garment image into the latent space  $\mathcal{E}_g$ , 2). using the CLIP to get the image token features for injection.
+
+Hand-Pose Aggregation Net. To aggregate the hand and pose-relevant priors learning, we tailor a Hand-Pose Net to constrain the semantic and skeleton pose, and global pixel-level hand knowledge. Specifically, dwpose, densepose, and render blocks are designed based on the same zero-multilayer structure with different weights to respectively extract the corresponding features. Zero-multilayer structure consists of seven stacking of conv and SiLU activate function with the zero convolution layer at the last layer.
+
+$$
+\mathcal {F} _ {h p} = \mathcal {F} _ {d w} + \mathcal {F} _ {d p} + \mathcal {F} _ {r h} \times w _ {\text {h a n d}}, \tag {2}
+$$
+
+where  $\mathcal{F}_{dw}$ ,  $\mathcal{F}_{dp}$ ,  $\mathcal{F}_{rh}$  are the DWPose feature, Densepose feature, and hand feature after the Zero-multilayer structure, and then the summation is implemented on the pixel-level with the adjustable weight  $w_{\mathrm{hand}}$  of hand feature  $\mathcal{F}_{rh}$ . To mitigate the impact of detrimental noise on the hidden states of the neural network layers in the trainable Hand-Pose Net during the initial stages of training, we also adopt the zero convolution learning paradigm of ControlNet.
+
+Hand structural embedding. The HaMeR [35] model is utilized to reconstruct various MANO parameters and other hand-related information from 2D images, including hand mask  $M_h \in \mathbb{R}^{768 \times 1024}$ , hand types  $T_h \in \mathbb{R}^{n \times 1}$ , hand 3D vertices  $V_h \in \mathbb{R}^{n \times 778 \times 3}$ , spatial joint locations  $J_{2d} \in \mathbb{R}^{n \times 21 \times 2}$ , and joint rotation matrices  $\theta \in \mathbb{R}^{n \times 16 \times 3 \times 3}$  where  $n$  represents the number of hands. To accommodate variable numbers of hands as input, we pad the hand parameters to a fixed length  $n = 2$  by adding padding at the end. For hand type  $T_h$ , we use 0 to denote left hands, 1 for right hands, and -1 as a filler for images without hands. Similar padding is applied to other parameters to ensure alignment and correspondence across all dimensions. We apply Basis Point Set (BPS) [38] for dimensionality reduction of hand vertices and encode joint rotations as 6D vectors. And then, we import theses information as well as joint rotations and hand types into the hand-structure processor. The structure of hand-structure processor consists of two parallel blocks: three stacked layers of linear and ReLU activation  $(L_r)$  and two stacked layers of linear and layer normalization  $(L_n)$ . These streamlined hand vertices and joint locations are then fed into  $L_r$ . Concurrently, joint rotations and hand types are transformed via  $L_n$ . The outputs from these processes are aggregated to form the Hand Structural Embedding  $c_{struc}$ , encapsulating essential geometric and 3D information for
+
+precise hand representation.
+
+$$
+c _ {s t r u c} = \left[ L _ {r} \left(\boldsymbol {V} _ {\boldsymbol {h}}\right), L _ {r} \left(\boldsymbol {J} _ {\boldsymbol {2 d}}\right), L _ {n} \left(\boldsymbol {T} _ {\boldsymbol {h}}\right), L _ {n} \left(\boldsymbol {\theta} _ {\boldsymbol {h}}\right) \right]. \tag {3}
+$$
+
+Hand appearance embedding. To improve hand visual detail, we extract features  $\mathcal{F} \in \mathbb{R}^{n \times 1536}$  from the hand Bounding Box regions using DINOv2 [34]. The Hand-Appear processor (HA) uses a linear layer followed by Layer Normalization [1], culminating in the Hand Appearance Embedding  $c_{\text{appear}}$ . This process ensures the hand's visual fidelity and detail are meticulously preserved in our representation.
+
+$$
+c _ {\text {a p p e a r}} = \mathrm {H A} (\mathcal {F}). \tag {4}
+$$
+
+Mask-based regional control. Our method introduces a mask-based regional control to enhance hand detail generation and maintain quality in virtual try-on. Utilizing the hand mask  $M_h$ , we modulate Cross-Attention queries, seamlessly integrating Hand Structural Embedding within Hand-Pose Net's Cross-Attention module. Concurrently, Hand Appearance Embedding is adeptly fused into the Try-onNet architecture through mask cross-attention, ensuring precise detail preservation. The mask cross-attention can be represented as:
+
+$$
+\begin{array}{l} Z = \operatorname {A t t e n t i o n} (Q, K, V, M _ {h}) \\ = \operatorname {S o f t m a x} \left(\frac {\left(Q K ^ {\top}\right) \odot M _ {h}}{\sqrt {d}}\right) V, \tag {5} \\ \end{array}
+$$
+
+where  $Q, K$  and  $V$  are the query, key, and value matrices of the attention operation for hand appearance embedding cross-attention,  $M_h$  represents the hand mask.
+
+Hand canny shape constraint. During the training process,  $\mathcal{L}_{hand}$  is implemented to refine the clarity and precision of hand edges. Specifically, for a small  $R_{t} < T$ , we employ one-step denoising process prior to sampling  $x_0$  for the direct prediction of  $z_0$  from  $z_{t}$  and the model-predicted noise, using Variational Autoencoder for decoder.
+
+$$
+z _ {0} = \frac {z _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta}}{\sqrt {\bar {\alpha} _ {t}}}, \quad t \leq R _ {t}. \tag {6}
+$$
+
+Hand regions are extracted from the generated images using bounding box cropping and are processed with the Canny edge detection operator. These regions are then compared to their corresponding areas in the Ground Truth images using Mean Squared Error (MSE) loss. This loss function is specifically designed to sharpen the model's focus on improving edge details and clarity in the hand regions of the generated images.
+
+# 4. Experiment
+
+Baselines. To comprehensively evaluate our approach, we selected five state-of-the-art diffusion-based methods in the
+
+![](images/87c2702bedbfbc89cfbca50f4f4eb62cfa75b3a211cc8e5d581619b6cc4c3ee6.jpg)
+
+![](images/2331abd65af745254c356932b86a1bea485f1f1ac287398e9bf363e0f013e8f1.jpg)  
+(I) Qualitative results on DressCode  
+(II) Qualitative results on VITON-HD  
+Figure 3. Qualitative comparisons of VTON-HandFit with other methods on public datasets: (I) Dresscode and (II) VITON-HD.
+
+virtual try-on domain as baselines for comparison: LaDIVTON [32], StableVTON [24], OOTDiffusion [48], IDM-VTON [7], and CatVTON [54]. While these methods have made substantial contributions to the virtual try-on field, they exhibit certain limitations in handling hand occlusion issues, which motivated us to propose VTON-HandFit.
+
+Dataset. Our experiments utilize two publicly virtual try-on datasets: VITON-HD [6] and DressCode [31]. The VITON-HD dataset consists of 13,679 image pairs, divided into 11,647 for training and 2,032 for testing, featuring front-facing upper-body portraits paired with garments at a resolution of  $1024 \times 768$ . The DressCode dataset, larger and more varied, contains 48,392 training pairs and 5,400 testing pairs, including full-body portraits and a range of garments (upper-body, lower-body, and dresses) at the same resolution. To enhance our evaluation of the model's capability to address hand occlusion challenges, we build Handfit-3K test with 3,620 images sorted from 26,000 images collected in online retail sites.
+
+Evaluation metrics. Following previous works [24, 32], we adopt both paired and unpaired evaluation settings. In the paired setting, the quality of the generated images is assessed using Structural Similarity (SSIM) [45], Learned Perceptual Image Patch Similarity (LPIPS) [58], Frechet Inception Distance (FID) [18], and Kernel Inception Distance
+
+(KID) [3]. For the unpaired setting, we employ FID and KID for comparison. Hand generation quality is assessed using FID-H and KID-H for the hand region. Additionally, we use 2D Mean Per Joint Position Error (MPJPE) [20] to evaluate hand joint position accuracy.
+
+Implementation details. Our model is developed using the pre-trained StableDiffusion v1.5 [39]. Detailed body dynamics and structural information are captured using DWpose [52] for pose estimation, Detector2 [46] for Dense-sepose segmentation, and HaMeR for detailed hand depth maps and parameters. We use VITON-HD and DressCode datasets for training. Our training strategy is implemented in two phases. In the first phase, we train the Hand-Pose Aggregation Net and Hand-feature Disentanglement Embedding modules while freezing the other components. In the second phase, we finetune the TryonNet and GarmentNet. All models are trained uniformly with a batch size of 32, employing the AdamW [28] optimizer at a learning rate of 5e-5. Training is conducted on 8 NVIDIA V100 GPUs.
+
+# 4.1. Results on Public Dataset
+
+Qualitative results. Fig. 3 shows visual comparisons of the VITON-HD and DressCode datasets between VTON-HandFit and other methods. In previous methods, original hand regions are often reused, leading to artifacts such as
+
+<table><tr><td rowspan="3">Methods</td><td colspan="6">DressCode</td><td colspan="6">VITON-HD</td></tr><tr><td colspan="4">Paired</td><td colspan="2">Unpaired</td><td colspan="4">Paired</td><td colspan="2">Unpaired</td></tr><tr><td>SSIM ↑</td><td>LPIPS ↓</td><td>FID ↓</td><td>KID ↓</td><td>FID ↓</td><td>KID ↓</td><td>SSIM ↑</td><td>LPIPS ↓</td><td>FID ↓</td><td>KID ↓</td><td>FID ↓</td><td>KID ↓</td></tr><tr><td>LaDI-VTON (2023)</td><td>0.9025</td><td>0.0719</td><td>4.8636</td><td>1.5580</td><td>6.8421</td><td>2.3345</td><td>0.8763</td><td>0.0911</td><td>6.6044</td><td>1.0672</td><td>9.4095</td><td>1.6866</td></tr><tr><td>StableVTON (2024)</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>0.8665</td><td>0.0835</td><td>6.8581</td><td>1.2553</td><td>9.5868</td><td>1.4508</td></tr><tr><td>OOTDiffusion (2024)</td><td>0.8975</td><td>0.0725</td><td>3.9497</td><td>0.7198</td><td>6.7019</td><td>1.8630</td><td>0.8513</td><td>0.0964</td><td>6.5186</td><td>0.8961</td><td>9.6733</td><td>1.2061</td></tr><tr><td>IDM-VTON (2024)</td><td>0.9228</td><td>0.0478</td><td>3.8001</td><td>1.2012</td><td>5.6159</td><td>1.5536</td><td>0.8806</td><td>0.0789</td><td>6.3381</td><td>1.3224</td><td>9.6114</td><td>1.6387</td></tr><tr><td>IDM-VTON* (2024)</td><td>0.9007</td><td>0.0624</td><td>4.9170</td><td>2.0176</td><td>6.6867</td><td>2.5793</td><td>0.8316</td><td>0.1058</td><td>7.7491</td><td>2.3381</td><td>9.7712</td><td>1.9323</td></tr><tr><td>CatVTON (2024)</td><td>0.9011</td><td>0.0705</td><td>3.2755</td><td>0.6696</td><td>5.4220</td><td>1.5490</td><td>0.8694</td><td>0.0970</td><td>6.1394</td><td>0.9639</td><td>9.1434</td><td>1.2666</td></tr><tr><td>VTON-HandFit (Ours)</td><td>0.9404</td><td>0.0444</td><td>2.8666</td><td>0.8009</td><td>5.2314</td><td>1.0189</td><td>0.8850</td><td>0.0730</td><td>5.9564</td><td>0.9699</td><td>8.8178</td><td>0.9951</td></tr></table>
+
+Table 1. Quantitative results on VITON-HD and DressCode datasets. We compare the metrics under both paired (model's clothing is the same as the given cloth image) and unpaired settings (model's clothing differs) with other methods. IDM-VTON* indicates post-processing using HandRefiner.
+
+![](images/d9fa97f64b899d23b4070c22fbea7fb4d3ec7188994b33d825325e2d3ed392d4.jpg)  
+Figure 4. Qualitative comparisons of VTON-HandFit with other methods on our Handfit-3K dataset.
+
+blue outlines around the hands when garment colors significantly differ (third row). In contrast, our model effectively eliminates these residual backgrounds and preserves natural hand poses, enhancing visual fidelity and pose accuracy.
+
+Quantitative results. Tab. 1 presents the quantitative comparisons of VTON-HandFit against other state-of-the-art methods on VITON-HD and DressCode datasets. By enhancing pose accuracy and the fidelity of hand gestures, our approach achieves markedly improved synthesized image quality. Moreover, VTON-HandFit outperforms in FID and KID scores across both datasets in unpaired settings, underscoring the importance of precise hand pose representation for enhancing image quality in practical applications.
+
+Limitations of handrefiner. Despite the strides made in hand reconstruction through HandRefiner [29], as a post-processing operation, HandRefiner falls short in its inability to reconstruct the hands without complete shape and fails to depict the same ones with the original model accurately. These deficiencies not only undermine the model's practicality but also compromise the authenticity of its output, thereby substantially diminishing its performance metrics and overall applicability, as shown in Tab. 1.
+
+![](images/dfcf5b3138264a54321f89f0ea28afbeb99987b7953a88108154f40c5f9ae89e.jpg)  
+Figure 5. Qualitative comparisons in real-world scenarios.
+
+![](images/f6b17a074c0e9c0e6efaf9f8797e2ffd923fadad363685b5bcc2457c536c0594.jpg)  
+Figure 6. Qualitative comparisons of VTON-HandFit with variants on VITON-HD dataset.
+
+# 4.2. Results on Handfit-3K Dataset
+
+Qualitative results. Fig. 4 presents a qualitative comparison between VTON-HandFit and other diffusion-based methods. Our method outperforms in preserving detailed garment attributes and achieving natural hand poses. The
+
+<table><tr><td rowspan="2">Methods</td><td colspan="7">Paired</td><td colspan="5">Unpaired</td></tr><tr><td>SSIM ↑</td><td>LPIPS ↓</td><td>FID ↓</td><td>KID ↓</td><td>MPJPE ↓</td><td>FID-H ↓</td><td>KID-H ↓</td><td>FID ↓</td><td>KID ↓</td><td>MPJPE ↓</td><td>FID-H ↓</td><td>KID-H ↓</td></tr><tr><td>LaDI-VTON (2023)</td><td>0.8597</td><td>0.1480</td><td>12.8430</td><td>4.4112</td><td>107.6315</td><td>23.0199</td><td>17.4558</td><td>16.2506</td><td>7.0065</td><td>111.1547</td><td>21.1371</td><td>14.1960</td></tr><tr><td>StableVTON (2024)</td><td>0.8772</td><td>0.1022</td><td>11.3675</td><td>4.1360</td><td>54.7586</td><td>22.7470</td><td>16.699</td><td>14.0446</td><td>5.1966</td><td>57.3334</td><td>22.8099</td><td>16.3022</td></tr><tr><td>IDM-VTON (2024)</td><td>0.8816</td><td>0.0935</td><td>7.7159</td><td>2.1287</td><td>48.6378</td><td>11.5403</td><td>3.9727</td><td>11.2960</td><td>3.4638</td><td>57.3447</td><td>12.7721</td><td>4.4315</td></tr><tr><td>OOTDiffusion (2024)</td><td>0.8554</td><td>0.1184</td><td>8.7104</td><td>2.1363</td><td>83.9655</td><td>12.1984</td><td>3.6653</td><td>13.1969</td><td>4.4839</td><td>86.1678</td><td>13.8011</td><td>4.4479</td></tr><tr><td>CatVTON (2024)</td><td>0.8761</td><td>0.1123</td><td>10.6966</td><td>4.2581</td><td>60.6608</td><td>13.1264</td><td>4.8723</td><td>15.7081</td><td>6.9544</td><td>63.2245</td><td>15.5603</td><td>5.6134</td></tr><tr><td>VTON-HandFit (Ours)</td><td>0.8827</td><td>0.0819</td><td>5.6027</td><td>0.4932</td><td>25.1376</td><td>7.0574</td><td>1.2474</td><td>10.9786</td><td>3.3248</td><td>26.6892</td><td>9.0003</td><td>2.5362</td></tr></table>
+
+Table 2. Quantitative results on Handfit-3K dataset in the paired and unpaired settings.  
+
+<table><tr><td>Methods</td><td>SSIM ↑</td><td>LPIPS ↓</td><td>FID ↓</td><td>KID ↓</td></tr><tr><td>Baseline</td><td>0.8619</td><td>0.0923</td><td>6.9549</td><td>1.2651</td></tr><tr><td>Baseline+HPA(Densepose)</td><td>0.8660</td><td>0.0876</td><td>6.8924</td><td>1.1254</td></tr><tr><td>Baseline+HPA</td><td>0.8738</td><td>0.0812</td><td>6.3390</td><td>0.9502</td></tr><tr><td>Baseline+HPA+Cstruct</td><td>0.8784</td><td>0.0787</td><td>6.2225</td><td>1.1186</td></tr><tr><td>Baseline+HPA+Cstruct+Cappear</td><td>0.8787</td><td>0.0792</td><td>6.0515</td><td>1.0084</td></tr><tr><td>VTON-HandFit</td><td>0.8850</td><td>0.0730</td><td>5.9564</td><td>0.9699</td></tr></table>
+
+Table 3. Comparison of different variants on VITON-HD dataset.
+
+images generated by VTON-HandFit demonstrate superior consistency in texture, pattern, and fit relative to baseline methods. Furthermore, the hand poses in our images are more consistent with the original.
+
+Quantitative results. Tab. 2 consolidates the performance of VTON-HandFit against contemporary methods on the Handfit-3K dataset. VTON-HandFit excels in both paired and unpaired settings, showcasing marked advancements in MPJPE, FID-H, and KID-H metrics. This evidences its capability to generate visually compelling images with precise hand pose representation and detailed hand regions. The model's robustness is further affirmed in the unpaired setting, where it consistently leads across all metrics, reinforcing its practical applicability in virtual try-on scenarios with a focus on hand detail fidelity.
+
+Results of real-world scenarios. To validate the performance of our VTON-HandFit in real-world scenarios, we have collected several cases from natural environments, as illustrated in Fig. 5. It is evident that VTON-HandFit continues to perform robustly in real-world settings, with a noticeable enhancement in hand effects.
+
+# 4.3. Ablation Study
+
+Effectiveness of each component. We perform an ablation study to evaluate the contributions of the Hand-Pose Aggregation Net, Hand Structural Embedding  $C_{struct}$ , and Hand Appearance Embedding  $C_{appear}$ . To facilitate a clear comparison, we introduce several model variants to ascertain the effectiveness of each component. The 'Baseline' model excludes all specialized components. The 'Baseline+HPA (Densepose)' variant incorporates only Densepose for pose control. 'Baseline+HPA' utilizes Densepose, Depth, and DWpose for pose control. 'Baseline+HPA+C_{struct}' excludes the appearance embedding. Lastly, 'Baseline+HPA+C_{struct}+C_{appear}' incorpo
+
+rates all components. It's important to note that these variants do not include hand-canny constraint loss. Tab. 3 presents a quantitative comparison of our ablation experiments, highlighting the impact of adding different components on the image generation quality. Fig. 6 offers a qualitative analysis, showing the necessity of pose guidance for accurate arm pose replication in the model. By incorporating stronger hand structure and visual guidance, the generation of hands is significantly improved.
+
+# 5. Conclusion
+
+We propose VTON-HandFit, a novel model designed for addressing hand occlusion in virtual try-on scenario. Specifically, we have developed a Hand-Pose Aggregation Net by explicitly and adaptively encoding global hand and pose priors using the ControlNet-based structure. To fully exploit the hand-related structure and appearance information, we propose the Hand-feature Disentanglement Embedding module, which explicitly separates the hand priors into hand structure-parametric and visual-appearance features. Lastly, we have customized a hand-canny constraint loss to enhance the learning of structural edge knowledge from the hand template of the model image. Our approach has been extensively evaluated qualitatively and quantitatively, demonstrating its superiority over state-of-the-art virtual try-on models, particularly in cases involving hand pose occlusion. These results pave the way for more complex virtual try-on applications toward real-world scenarios.
+
+# Acknowledgments
+
+This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603) and National Natural Science Foundation of China (No. 62072386), supported by Yunnan Provincial Major S&T Special Plan Project (No. 202402AD080001), Henan Key R&D Project (No. 231111212000), Henan Center for Outstanding Overseas Scientists (No. GZS2022011), Open Foundation of Henan Key Lab of General Aviation Technology (No. ZHKF-230212), Key Lab of Oracle Information Processing of MOE (No. OIP2024E002). We also thank Digital Culture Lab of Tencent SSV for their invaluable contribution during our research.
+
+# References
+
+[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5  
+[2] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5968-5976, 2023. 3  
+[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6  
+[4] Shidong Cao, Wenhao Chai, Shengyu Hao, Yanting Zhang, Hangyue Chen, and Gaoang Wang. Diffashion: Reference-based fashion design with structure-aware transfer by diffusion models. IEEE Transactions on Multimedia, 2023. 3  
+[5] Mengting Chen, Xi Chen, Zhonghua Zhai, Chen Ju, Xuewen Hong, Jinsong Lan, and Shuai Xiao. Wear-any-way: Manipulable virtual try-on via sparse correspondence alignment. arXiv preprint arXiv:2403.12965, 2024. 3  
+[6] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 6  
+[7] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. ECCV, 2024. 3, 6  
+[8] Zheng Chong, Xiao Dong, Haoxiang Li, Shiyue Zhang, Wenqing Zhang, Xujie Zhang, Hanqing Zhao, and Xiaodan Liang. Catvton: Concatenation is all you need for virtual try-on with diffusion models. arXiv preprint arXiv:2407.15886, 2024. 1  
+[9] Morelli Davide, Fincato Matteo, Cornia Marcella, Landi Federico, Cesari Fabio, and Cucchiara Rita. Dress code: High-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1  
+[10] Haoye Dong, Xiaodan Liang, Xiaohui Shen, Bochao Wang, Hanjiang Lai, Jia Zhu, Zhiting Hu, and Jian Yin. Towards multi-pose guided virtual try-on network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9026-9035, 2019.  
+[11] Chongjian Ge, Yibing Song, Yuying Ge, Han Yang, Wei Liu, and Ping Luo. Disentangled cycle consistency for highly-realistic virtual try-on. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16928-16937, 2021. 1  
+[12] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8485-8493, 2021. 3  
+[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and
+
+Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 1, 3  
+[14] Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang. Taming the power of diffusion models for high-quality virtual try-on with appearance flow. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7599-7607, 2023. 3  
+[15] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 1  
+[16] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10471-10480, 2019.  
+[17] Sen He, Yi-Zhe Song, and Tao Xiang. Style-based global appearance flow for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3470-3479, 2022. 1  
+[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6  
+[19] Xiaobin Hu, Xu Peng, Donghao Luo, Xiaozhong Ji, Jinlong Peng, Zhengkai Jiang, Jiangning Zhang, Taisong Jin, Chengjie Wang, and Rongrong Ji. Diffumatting: Synthesizing arbitrary objects with matting-level annotation. arXiv preprint arXiv:2403.06168, 2024. 4  
+[20] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 6  
+[21] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzenes. Do not mask what you do not need to mask: a parser-free virtual try-on. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 619-635. Springer, 2020. 3  
+[22] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22623-22633. IEEE, 2023. 3  
+[23] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 4  
+[24] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8176-8185, 2024. 3, 6  
+[25] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF
+
+Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 4  
+[26] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In European Conference on Computer Vision, pages 204-219. Springer, 2022. 1, 3  
+[27] Yuhan Li, Hao Zhou, Wenxiang Shang, Ran Lin, Xuanhong Chen, and Bingbing Ni. Anyfit: Controllable virtual try-on for any combination of attire across any scenario. arXiv preprint arXiv:2405.18172, 2024. 3  
+[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6  
+[29] Wenquan Lu, Yufei Xu, Jing Zhang, Chaoyue Wang, and Dacheng Tao. Handrefiner: Refining malformed hands in generated images by diffusion-based conditional inpainting. In ACM Multimedia 2024, 2023. 4, 7  
+[30] Yifang Men, Yiming Mao, Yuning Jiang, Wei-Ying Ma, and Zhouhui Lian. Controllable person image synthesis with attribute-decomposed gan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5084-5093, 2020. 3  
+[31] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: High-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2231-2235, 2022. 6  
+[32] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. Ladi-vton: Latent diffusion textual-inversion enhanced virtual try-on. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8580-8589, 2023. 3, 6  
+[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 4  
+[34] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5  
+[35] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 4, 5  
+[36] Anton Pelykh, Ozge Mercanoglu Sincan, and Richard Bowden. Giving a hand to diffusion models: a two-stage approach to improving conditional human image generation. arXiv preprint arXiv:2403.10731, 2024. 4  
+[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 4
+
+[38] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4332-4341, 2019. 5  
+[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 4, 6  
+[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 4  
+[41] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 4  
+[42] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4695-4703, 2024. 4  
+[43] Ke Sun, Jian Cao, Qi Wang, Linrui Tian, Xindi Zhang, Lian Zhuo, Bang Zhang, Liefeng Bo, Wenbo Zhou, Weiming Zhang, et al. Outfitanyone: Ultra-high quality virtual try-on for any clothing and any person. arXiv preprint arXiv:2407.16224, 2024. 3  
+[44] Chengrui Wang, Pengfei Liu, Min Zhou, Ming Zeng, Xubin Li, Tiezheng Ge, et al. Rhand: Refining malformed hands for generated images with decoupled structure and style guidance. arXiv preprint arXiv:2404.13984, 2024. 4  
+[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6  
+[46] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2. https://github.com/facebookresearch/detectron2, 2019. 6  
+[47] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. Gqvton: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23550-23559, 2023. 3  
+[48] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. arXiv preprint arXiv:2403.01779, 2024.6  
+[49] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 3
+
+[50] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wangmeng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating-preserving image content. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7850-7859, 2020. 1  
+[51] Zhijing Yang, Junyang Chen, Yukai Shi, Hao Li, Tianshui Chen, and Liang Lin. Occlumix: Towards de-occlusion virtual try-on by semantically-guided mixup. IEEE Transactions on Multimedia, 25:1477–1488, 2023. 3  
+[52] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 6  
+[53] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023.4  
+[54] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and An-An Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8372-8382, 2024. 6  
+[55] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. ICLR, 2023. 2  
+[56] Haozhuo Zhang, Bin Zhu, Yu Cao, and Yanbin Hao. Hand1000: Generating realistic hands from text with only 1,000 images. arXiv preprint arXiv:2408.15461, 2024. 4  
+[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4  
+[58] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6  
+[59] Xujie Zhang, Ente Lin, Xiu Li, Yuxuan Luo, Michael Kampffmeyer, Xin Dong, and Xiaodan Liang. Mmtryon: Multi-modal multi-reference control for high-quality fashion generation. arXiv preprint arXiv:2405.00448, 2024. 3  
+[60] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 4  
+[61] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4606-4615, 2023. 1, 2, 3

2025/VTON-HandFit_ Virtual Try-on for Arbitrary Hand Pose Guided by Hand Priors Embedding/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8fa5d9afb09c8ad47a5ace8498c4ca6e8b78e45eaa1d97bf9754df9ce08c8855
+size 714006

2025/V^2Dial_ Unification of Video and Visual Dialog via Multimodal Experts/a2fbca87-c18a-41dc-864b-f66927e9f633_content_list.json

@@ -0,0 +1,1711 @@
+[
+    {
+        "type": "text",
+        "text": "$\\mathbf{V}^{2}$  Dial: Unification of Video and Visual Dialog via Multimodal Experts",
+        "text_level": 1,
+        "bbox": [
+            109,
+            125,
+            859,
+            154
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/eb6eb8aaa27e4b0f18b08931633881735cc23e776a2800be52f4fd8d37365f5c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            129,
+            186,
+            834,
+            252
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            286,
+            313,
+            301
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We present  $V^2$  Dial - a novel expert-based model specifically geared towards simultaneously handling image and video input data for multimodal conversational tasks. Current multimodal models primarily focus on simpler tasks (e.g., VQA, VideoQA, video-text retrieval) and often neglect the more challenging conversational counterparts, such as video and visual/image dialog. Moreover, works on both conversational tasks evolved separately from each other despite their apparent similarities, limiting their applicability potential. To this end, we propose to unify both tasks using a single model that for the first time jointly learns the spatial and temporal features of images and videos by routing them through dedicated experts and aligns them using matching and contrastive learning techniques. Furthermore, we systemically study the domain shift between the two tasks by investigating whether and to what extent these seemingly related tasks can mutually benefit from their respective training data. Extensive evaluations on the widely used video and visual dialog datasets of AVSD and VisDial show that our model achieves new state-of-the-art results across four benchmarks both in zero-shot and fine-tuning settings.",
+        "bbox": [
+            75,
+            316,
+            473,
+            636
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            662,
+            209,
+            679
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Enabled by the availability of large-scale training data [10, 14, 41] and advances in model design [12, 24, 37, 48, 52], the field of vision-and-language learning saw unprecedented success in recent years. However, current multimodal foundational models [7, 32, 37, 50, 54] still mainly focus on single-round tasks (e.g., VQA [8], VideoQA [58], video-text and text-video retrieval [59]). In contrast, the significantly more challenging conversational tasks, such as visual [1, 22] and video dialog [5], received considerably less attention. Furthermore, methods for these different tasks have advanced independently of each other despite the apparent structural similarities between them. They both operate on a visual input (i.e. an image or video), a short visual description (caption), and a dialog history composed of",
+        "bbox": [
+            75,
+            688,
+            470,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/269c233cda7f0815cbdd2c0db7d53c4ec50f90845f5b7802b8cf9e8a9cff3f6c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            284,
+            874,
+            429
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/8856e672e3d2aa85751f93c17eaf2ef9190b679e6e3170682ac2f180819d8ce0.jpg",
+        "image_caption": [
+            "Figure 1.  $\\mathbf{V}^2\\mathbf{D}$ ial uses multimodal experts and outperforms state-of-the-art methods on both video and visual dialog in zero-shot and fine-tuning evaluation settings."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            436,
+            872,
+            744
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "previous question-answer pairs. On the one hand, visual dialog models [4, 15, 16, 44, 56] have been primarily trained to rank a list of candidate answers using a Next Sentence Prediction (NSP) head similar to BERT [23] and negative sampling. Thus, they are benchmarked using retrieval met",
+        "bbox": [
+            496,
+            824,
+            892,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "8637",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "rics such as recall (R@k) and normalized discounted cumulative gain (NDCG). In contrast, video dialog models [2, 3, 6, 18, 20, 29, 46, 60] are trained to auto-regressively predict the next answer token using teacher forcing [57] and are evaluated using language generation metrics.",
+        "bbox": [
+            75,
+            90,
+            468,
+            167
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this work, we mitigate the shortcomings of current dialog systems by proposing  $\\mathbf{V}^2\\mathbf{Dial}$  - a novel multimodal expert-based model capable of unifying video and visual dialog tasks without any architectural changes. Specifically, we train dedicated multimodal expert layers that separately process the features of each input modality and learn how to align them using matching and contrastive learning techniques. A key novelty of our approach is that we use dedicated experts to jointly learn the spatial and temporal features of images and videos by routing them through the appropriate experts. Then, we couple these layers with a pretrained LLM to align their hidden states. Thanks to its modularity, our model can efficiently tackle image and video input data simultaneously and seamlessly train on both data types. In summary, our contributions are three-fold:",
+        "bbox": [
+            75,
+            167,
+            468,
+            391
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose  $\\mathbf{V}^2\\mathbf{Dial}$  - a multimodal expert-based model that unifies visual and video dialog by simultaneously learning from image and video data. As a core novelty, it employs two experts to separately learn the spatial and temporal features of images and videos.  $\\mathbf{V}^2\\mathbf{Dial}$  outperforms state-of-the-art models in both zero-shot and fine-tuning settings (see Figure 1).",
+            "- We are the first to systematically quantify the effect of domain shift between video and visual dialog tasks based on evaluations on the two widely used datasets of AVSD [5] and VisDial [22]. To this end, we propose an alternative ranking scheme that allows computing the VisDial retrieval metric for fully generative models and enables a fair comparison with previous works.",
+            "- We are the first to evaluate AVSD in a zero-shot setting, which provides a more solid generalization evaluation of video dialog models compared to the fine-tuned setting. For this, we establish the first benchmark comparison of recent state-of-the-art multimodal models."
+        ],
+        "bbox": [
+            76,
+            393,
+            468,
+            681
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            693,
+            218,
+            709
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Visual & Video Dialog. Modeled after human-human communication, visual and video dialog involve reasoning about a visual scene in the form of a video or an image through multiple question-answering rounds in natural language. In comparison to their single-round counterparts, VQA [8] and VideoQA [58], dialog models need to additionally reason about the previous dialog history together with the visual grounding and the current question to be able to answer it efficiently. The best performing visual dialog models [4, 42, 56, 61] leverage pre-trained VLMs and are trained using an NSP head, negative sampling, and binary classification loss. At test time, for each question, the can",
+        "bbox": [
+            75,
+            719,
+            467,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "didate answers are ranked based on their respective NSP scores to compute the retrieval metrics. Although some work [15, 16, 55] claim to train generative visual dialog models, they do so by providing a generative mask where each token can only attend to its left tokens. However, they are trained using the NSP head like the discriminative models. However, this training approach is limiting and suboptimal for a unifying model. Thus, we advocate for a fully generative training paradigm and adapt the ranking scheme of VisDial answers to cater to modern generative models.",
+        "bbox": [
+            496,
+            90,
+            890,
+            241
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In contrast, works on video dialog follow a purely generative training paradigm and achieved great success building on top of powerful pre-trained LLMs [30, 47]. For example, [28, 34] fine-tuned a LLM on AVSD and obtained performance boosts. More recent works [3, 29] combined LLMs with GNNs and pushed the state-of-the-art results even further. Others [60] introduced a regularization loss to mitigate hallucination. Although video dialog emerged as a natural extension to visual dialog with apparent data structure similarities, research on both tasks evolved separately. To this end, we propose a unifying model that can simultaneously learn both tasks without any architectural modifications and for the first time; systemically study the effect of domain shift between both tasks using the AVSD and VisDial v1.0 datasets.",
+        "bbox": [
+            496,
+            243,
+            890,
+            469
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Multimodal Expert-based Training. Enhancing models with expert-based training has shown promising potential in boosting performance while maintaining computational efficiency [25, 63, 64]. Some works [12, 54] explored using single modality specific experts within a multimodal transformer architecture. Specifically, they used one vision and one language specific FFN after a shared multi-head self-attention block. Other works [35, 43] explored using multiple sparse modality-agnostic experts and trained them using soft-routers. Our work is positioned at the middle ground of the previously mentioned research directions: We propose to use multiple hard-routed experts per modality to be able to capture more fine-grained features compared to a single expert or multiple modality agnostic experts. Specifically, to the best of our knowledge,  $\\mathbf{V}^2\\mathbf{Dial}$  is the first model that learns disentangled spatial and temporal features using two dedicated experts that jointly learn from image and video data. In addition, we propose to deploy two separate language experts (for caption and context) in order to tackle the unique challenges of multimodal conversational tasks.",
+        "bbox": [
+            496,
+            484,
+            890,
+            787
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.  $\\mathbf{V}^2$  Dial",
+        "text_level": 1,
+        "bbox": [
+            500,
+            801,
+            581,
+            820
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.1. Joint Problem Formulation",
+        "text_level": 1,
+        "bbox": [
+            500,
+            830,
+            746,
+            845
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We use a fully generative formulation to unify both video and visual dialog tasks. Specifically, given visual input V (video/image), a corresponding visual description (caption",
+        "bbox": [
+            500,
+            854,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "8638",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/76d158b2b310140434bcb6f983d2d8afde34438791dbe9bb380d981608e610be.jpg",
+        "image_caption": [
+            "Tr. Stage 1",
+            "Figure 2. Architectural overview of  $\\mathbf{V}^2$  Dial. We adopt a training strategy composed of three stages. First, we only train the multimodal expert layers using spatial-temporal and video/image text matching losses ( $\\mathcal{L}_{\\mathrm{stm}}$ ,  $\\mathcal{L}_{\\mathrm{vm}}$ ), spatial-temporal and video/image contrastive learning losses ( $\\mathcal{L}_{\\mathrm{stc}}$ ,  $\\mathcal{L}_{\\mathrm{vtc}}$ ), and masked language modeling loss ( $\\mathcal{L}_{\\mathrm{mlm}}$ ). Second, we couple the expert layers with a frozen pre-trained LLM end-to-end, using a generative loss  $\\mathcal{L}_{\\mathrm{gen}}$  to align their hidden representations. Finally, we additionally fine-tune the LLM weights on the downstream benchmarks. Each expert is a feed-forward network (FFN) composed of two fully connected layers."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            84,
+            104,
+            885,
+            330
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "C), a dialog history  $\\mathrm{H}_r = \\left\\{\\left(\\mathbb{Q}_1,\\mathbb{A}_1\\right),\\dots,\\left(\\mathbb{Q}_{r - 1},\\mathbb{A}_{r - 1}\\right)\\right\\}$  comprised of the previous question-answer pairs  $\\left\\{\\left(\\mathbb{Q}_i,\\mathbb{A}_i\\right)\\right\\}_{i = 1}^{r - 1}$  and the current question  $\\mathbb{Q}_r$  , a model is trained to auto-regressively predict a free-form answer  $\\mathbb{A}_r$  at round  $r$  Specifically, each answer token  $\\mathsf{a}_r^i$  satisfies",
+        "bbox": [
+            75,
+            438,
+            468,
+            516
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathsf {a} _ {r} ^ {i} = \\underset {\\mathsf {a} \\in \\mathcal {V}} {\\arg \\max } \\left[ \\mathbf {p} \\left(\\mathsf {a} | \\mathrm {V}, \\mathrm {C}, \\mathrm {H} _ {r} \\mathsf {Q} _ {r}, \\mathrm {A} _ {r} ^ {<   i}\\right) \\right], \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            138,
+            525,
+            468,
+            554
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbb{A}_r^{<i}$  denotes the previously predicted answer tokens and  $\\mathcal{V}$  the vocabulary. In the rest, we use \"context\" to refer to the concatenation of the history  $\\mathbb{H}_r$  and the question  $\\mathbb{Q}_r$ .",
+        "bbox": [
+            75,
+            566,
+            468,
+            613
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Architecture",
+        "text_level": 1,
+        "bbox": [
+            76,
+            622,
+            210,
+            636
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Overview. As can be seen from Figure 2, our model takes an image/video  $\\mathbf{v} \\in \\mathbb{R}^{F \\times 3 \\times H \\times W}$  as input, where  $F$  is the number of frames and is set to one for images, and  $(H, W)$  is the re-sized resolution. Then it processes every frame using a pre-trained EVA-CLIP [51] Image Encoder and concatenates every four spatially adjacent visual patches into a single one. Then, a linear layer maps each visual token into a lower dimensional vector  $\\mathbf{v}$  of dimension  $D$  to obtain",
+        "bbox": [
+            75,
+            645,
+            468,
+            767
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {V} = \\left[ \\begin{array}{c c c c} \\mathbf {v} _ {1} ^ {1} & \\mathbf {v} _ {1} ^ {2} & \\dots & \\mathbf {v} _ {1} ^ {F} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ \\mathbf {v} _ {P} ^ {1} & \\mathbf {v} _ {P} ^ {2} & \\dots & \\mathbf {v} _ {P} ^ {F} \\end{array} \\right] \\in \\mathbb {R} ^ {F \\times P \\times D}, \\qquad (2)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            137,
+            773,
+            468,
+            828
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $P = \\frac{1}{4} \\frac{H \\times W}{14^2}$  and  $D$  denote the visual input length and the joint hidden dimension, respectively. Thereafter, in stark contrast to previous works [13, 19] that performed spatial and temporal attention in series, our model separately",
+        "bbox": [
+            75,
+            838,
+            470,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "performs these operations using the masks  $\\mathbf{M}^{\\mathrm{spa}}$  and  $\\mathbf{M}^{\\mathrm{tmp}}$  as shown in Figure 2 on the visual features  $V$  to obtain",
+        "bbox": [
+            498,
+            438,
+            890,
+            469
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {V} ^ {\\mathrm {s p a}} = \\operatorname {S A} (\\mathbf {V}, \\mathbf {M} ^ {\\mathrm {s p a}}) \\in \\mathbb {R} ^ {(F P) \\times D} \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            570,
+            481,
+            890,
+            501
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {V} ^ {\\mathrm {t m p}} = \\operatorname {S A} (\\mathbf {V}, \\mathbf {M} ^ {\\mathrm {t m p}}) \\in \\mathbb {R} ^ {(F P) \\times D} \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            570,
+            503,
+            890,
+            523
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {M} _ {m, n} ^ {\\mathrm {s p a}} \\left(\\mathbf {v} _ {i} ^ {j}\\right) = \\delta_ {n j}, \\quad \\mathbf {M} _ {m, n} ^ {\\mathrm {t m p}} \\left(\\mathbf {v} _ {i} ^ {j}\\right) = \\delta_ {m i} \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            570,
+            527,
+            890,
+            547
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where SA and  $\\delta$  denote self-attention and Kronecker delta.",
+        "bbox": [
+            500,
+            560,
+            885,
+            574
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Subsequently, the textual input in the form of a caption and a context is processed by an embedding layer to obtain  $\\mathbf{T}^{\\mathrm{cap / ctx}}\\in \\mathbb{R}^{N_{\\mathrm{cap / ctx}}\\times D}$ , where  $N_{\\mathrm{cap}}$  and  $N_{\\mathrm{ctx}}$  are the respective lengths of the caption and context. These visual and textual features form the initial input to the multimodal expert layers which are pre-trained using a combination of matching, contrastive learning, and masked language modeling losses. Finally, they are coupled with a pre-trained LLM and are fine-tuned end-to-end using a generative loss.",
+        "bbox": [
+            496,
+            575,
+            890,
+            713
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Multimodal Expert Layers. These consist of  $N$  layers of stacked multi-head self-attention with layer normalization (SA), and several modality-specific and one modality-agnostic feed-forward networks that we refer to as experts. As shown in Figure 2, we propose to use a set of six experts denoted as  $\\{\\mathcal{E}_*\\}$ : three of which are vision-specific and two are language-specific and are activated in the first  $L$  layers. The remaining expert  $\\mathcal{E}_{\\mathrm{fus}}$  is the fusion expert and is only activated in the last  $(N - L)$  layers and operates on the concatenation of all available modalities (Equation 9). To the best of our knowledge, we propose for the first time to learn the spatial and temporal features using dedicated",
+        "bbox": [
+            496,
+            732,
+            890,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "8639",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/c58a252e17c236379fd3298fe8f998ed1bbe65c274a81671425a45ccb87e9054.jpg",
+        "image_caption": [
+            "Figure 3. Overview of the training and evaluation pipeline of  $\\mathbf{V}^2$ Dial. We show the different datasets used to train our model at each stage. Evaluations are conducted on the most popular video and visual dialog datasets of AVSD and VisDial, respectively. ( = video data, = image data, = closed / visual captioning data, = dialog data)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            101,
+            85,
+            869,
+            205
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "experts (i.e., the spatial  $\\mathcal{E}_{\\mathrm{spa}}$  and temporal  $\\mathcal{E}_{\\mathrm{tmp}}$  experts, respectively) as shown in Equation 11. This allows our model to unify video and visual dialog by jointly learning from image and video data. The visual expert  $\\mathcal{E}_{\\mathrm{vis}}$  operates on top of the concatenation of  $\\mathcal{E}_{\\mathrm{spa}}$  and  $\\mathcal{E}_{\\mathrm{tmp}}$  to learn a joint spatial-temporal video representation (Equation 10). Similarly, the textual experts  $\\mathcal{E}_{\\mathrm{cap}}$  and  $\\mathcal{E}_{\\mathrm{ctx}}$  operate on the caption and context embeddings  $\\mathbf{T}^{\\mathrm{cap}}$  and  $\\mathbf{T}^{\\mathrm{ctx}}$  (Equation 12). As seen in Table 1, the availability of the multimodal features depends on the visual input type (i.e., videos vs images) and the training stage. However, without the loss of generality, we can formulate the multimodal expert layers as follows:",
+        "bbox": [
+            75,
+            287,
+            473,
+            460
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {X} _ {0} = \\left[ \\mathbf {V} ^ {\\text {s p a}}, \\mathbf {V} ^ {\\text {t m p}}, \\mathbf {T} ^ {\\text {c a p}}, \\mathbf {T} ^ {\\text {c t x}} \\right], \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            102,
+            463,
+            468,
+            479
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\tilde {\\mathbf {X}} _ {l} = \\left[ \\tilde {\\mathbf {V}} _ {l} ^ {\\mathrm {s p a}}, \\tilde {\\mathbf {V}} _ {l} ^ {\\mathrm {t m p}}, \\tilde {\\mathbf {T}} _ {l} ^ {\\mathrm {c a p}}, \\tilde {\\mathbf {T}} _ {l} ^ {\\mathrm {e x t}} \\right] (7) \\\\ = \\operatorname {S A} \\left(\\mathbf {X} _ {l - 1}\\right) + \\mathbf {X} _ {l - 1} (8) \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            106,
+            483,
+            468,
+            517
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {X} _ {l} = \\left\\{ \\begin{array}{l l} \\left[ \\mathbf {V} _ {l} ^ {\\text {v i s}}, \\mathbf {T} _ {l} ^ {\\text {c a p}}, \\mathbf {T} _ {l} ^ {\\text {c t x}} \\right] & i f 1 \\leq l \\leq L \\\\ \\mathcal {E} _ {\\text {f u s}} \\left(\\tilde {\\mathbf {X}} _ {l}\\right) + \\tilde {\\mathbf {X}} _ {l} & i f L <   l \\leq N \\end{array} , \\right. \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            106,
+            522,
+            468,
+            556
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {V} _ {l} ^ {\\text {v i s}} = \\mathcal {E} _ {\\text {v i s}} \\left(\\tilde {\\mathbf {V}} _ {l} ^ {\\text {v i s}}\\right) + \\tilde {\\mathbf {V}} _ {l} ^ {\\text {v i s}}, \\quad \\tilde {\\mathbf {V}} _ {l} ^ {\\text {v i s}} := \\left[ \\begin{array}{l l} \\mathbf {V} _ {l} ^ {\\text {s p u}}, & \\mathbf {V} _ {l} ^ {\\text {t m p}} \\end{array} \\right], \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            99,
+            559,
+            468,
+            577
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {V} _ {l} ^ {\\mathrm {s p a}} = \\mathcal {E} _ {\\mathrm {s p a}} \\left(\\tilde {\\mathbf {V}} _ {l} ^ {\\mathrm {s p a}}\\right) + \\tilde {\\mathbf {V}} _ {l} ^ {\\mathrm {s p a}}, \\quad \\mathbf {V} _ {l} ^ {\\mathrm {t m p}} = \\mathcal {E} _ {\\mathrm {t m p}} \\left(\\tilde {\\mathbf {V}} _ {l} ^ {\\mathrm {t m p}}\\right) + \\tilde {\\mathbf {V}} _ {l} ^ {\\mathrm {t m p}}, \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            98,
+            579,
+            468,
+            597
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {T} _ {l} ^ {\\mathrm {c a p}} = \\mathcal {E} _ {\\mathrm {c a p}} \\left(\\tilde {\\mathbf {T}} _ {l} ^ {\\mathrm {c a p}}\\right) + \\tilde {\\mathbf {T}} _ {l} ^ {\\mathrm {c a p}}, \\quad \\mathbf {T} _ {l} ^ {\\mathrm {c t x}} = \\mathcal {E} _ {\\mathrm {c t x}} \\left(\\tilde {\\mathbf {T}} _ {l} ^ {\\mathrm {c t x}}\\right) + \\tilde {\\mathbf {T}} _ {l} ^ {\\mathrm {c t x}}. \\tag {12}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            98,
+            599,
+            468,
+            617
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "When dealing with images and non-dialog data, we drop  $\\mathbf{V}_l^{\\mathrm{tmp}}$  and  $\\mathbf{T}_l^{\\mathrm{cap}}$  from the previous equations and deactivate the respective expert.",
+        "bbox": [
+            76,
+            625,
+            468,
+            671
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Training",
+        "text_level": 1,
+        "bbox": [
+            76,
+            678,
+            181,
+            695
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3.1 Stage 1",
+        "text_level": 1,
+        "bbox": [
+            76,
+            702,
+            181,
+            717
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In the first stage, we only pre-train the multimodal expert layers, the vision encoder linear layer, and the spatial-temporal attention modules. Since we are the first to suggest learning the spatial and temporal features of videos and images using dedicated experts, we propose to train our model using spatial-temporal contrastive learning (STC) and spatial-temporal matching (STM). In addition, we use the established masked language modeling (MLM), vision-text $^1$  contrastive learning (VTC), and vision-text matching (VTM) similar to [19, 31, 32].",
+        "bbox": [
+            75,
+            726,
+            468,
+            878
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/be83844f2edadd414606c51cd18c8e7d7f557f52fe50f344a4961b2c2451c9d2.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Tr. Stage 1</td><td>Tr. Stage 2</td><td>Tr. Stage 3</td></tr><tr><td>Videos</td><td>Vspa, Vtmp, Tcap</td><td>Vspa, Vtmp, Tcap, Tctx</td><td>Vspa, Vtmp, Tcap, Tctx</td></tr><tr><td>Images</td><td>Vspa, Tcap</td><td>-</td><td>Vspa, Tcap, Tctx</td></tr></table>",
+        "bbox": [
+            503,
+            284,
+            893,
+            330
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Overview of the available features for each training stage and visual input type.",
+        "bbox": [
+            500,
+            340,
+            893,
+            369
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Spatial-Temporal Contrastive Learning aims to better align the spatial and temporal features of video data. To this end, we use output features of the last multi-modal exert layer<sup>2</sup> and learn a cosine similarity function",
+        "bbox": [
+            498,
+            396,
+            892,
+            458
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathrm {s} \\left(\\mathbf {V} ^ {\\text {s p a}}, \\mathbf {V} ^ {\\text {t m p}}\\right) = \\Theta_ {\\text {s p a}} \\left(\\mathbf {V} ^ {\\text {s p a}}\\right) ^ {\\top} \\Theta_ {\\text {t m p}} \\left(\\mathbf {V} ^ {\\text {t m p}}\\right), \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            539,
+            469,
+            890,
+            491
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "so that aligned spatial-temporal features result in higher similarity scores, where  $\\Theta_{*}$  are linear layers that map the features to a normalized lower dimensional vector space. Then, given spatial and temporal feature pairs, we compute the softmax normalized spatial-to-temporal and temporal-to-spatial similarities as",
+        "bbox": [
+            498,
+            500,
+            892,
+            584
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {i} ^ {\\mathrm {s 2 t}} \\left(\\mathbf {V} ^ {\\mathrm {s p a}}\\right) = \\frac {\\exp \\left(\\tilde {\\mathrm {s}} \\left(\\mathbf {V} ^ {\\mathrm {s p a}} , \\mathbf {V} _ {i} ^ {\\mathrm {t m p}}\\right) / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(\\tilde {\\mathrm {s}} \\left(\\mathbf {V} ^ {\\mathrm {s p a}} , \\mathbf {V} _ {k} ^ {\\mathrm {t m p}}\\right) / \\tau\\right)}, \\tag {14}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            566,
+            594,
+            890,
+            630
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {i} ^ {\\mathrm {t 2 s}} \\left(\\mathbf {V} ^ {\\mathrm {t m p}}\\right) = \\frac {\\exp \\left(\\tilde {\\mathrm {s}} \\left(\\mathbf {V} ^ {\\mathrm {t m p}} , \\mathbf {V} _ {i} ^ {\\mathrm {s p a}}\\right) / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(\\tilde {\\mathrm {s}} \\left(\\mathbf {V} ^ {\\mathrm {t m p}}, \\mathbf {V} _ {k} ^ {\\mathrm {s p a}}\\right) / \\tau\\right)}, \\tag {15}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            563,
+            632,
+            890,
+            667
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\tau$  is learnable temperature parameters, and  $\\tilde{\\mathbf{s}}$  is the maximum value of  $\\mathbf{s}$  as in [32]. Finally, we can compute the loss as the cross-entropy  $\\mathcal{H}$  between  $\\mathbf{p}$  and  $\\mathbf{y}$ :",
+        "bbox": [
+            500,
+            675,
+            890,
+            719
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s t c}} = \\frac {1}{2} \\mathbb {E} _ {\\left(\\mathbf {V} ^ {\\mathrm {s p a}}, \\mathbf {V} ^ {\\mathrm {t m p}}\\right)} \\left[ \\mathcal {H} \\left(\\mathbf {y} ^ {\\mathrm {s 2 t}}, \\mathbf {p} ^ {\\mathrm {s 2 t}}\\right) + \\mathcal {H} \\left(\\mathbf {y} ^ {\\mathrm {t 2 s}}, \\mathbf {p} ^ {\\mathrm {t 2 s}}\\right) \\right], \\tag {16}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            521,
+            728,
+            890,
+            753
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{y}^{s2t}$  and  $\\mathbf{y}^{t2s}$  are the golden one-hot similarities.",
+        "bbox": [
+            501,
+            763,
+            862,
+            780
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Spatial-Temporal Matching complements STC and teaches the model to distinguish between positive and negative spatial-temporal feature pairs. Specifically, a matched feature pair originates from the same video, whereas an unmatched pair is constructed using negative sampling from a",
+        "bbox": [
+            498,
+            799,
+            893,
+            876
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{1}$ Vision can either be video or image depending on the dataset.",
+        "bbox": [
+            94,
+            885,
+            426,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "2 Index dropped for clarity.",
+        "bbox": [
+            517,
+            885,
+            661,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "8640",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "different video. We use a classification token as a proxy of the joint spatial-temporal representations to learn a binary classification problem using the STM loss",
+        "bbox": [
+            75,
+            90,
+            472,
+            137
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s t m}} = \\mathbb {E} _ {\\left(\\mathbf {V} ^ {\\mathrm {s p a}}, \\mathbf {V} ^ {\\mathrm {t m p}}\\right)} \\left[ \\mathcal {H} \\left(\\mathbf {y} ^ {\\mathrm {s t m}}, \\mathbf {p} ^ {\\mathrm {s t m}}\\right) \\right], \\tag {17}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            151,
+            150,
+            468,
+            172
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{p}^{\\mathrm{stm}}$  and  $\\mathbf{y}^{\\mathrm{stm}}$  are the predicted and the ground-truth two-class probabilities, respectively.",
+        "bbox": [
+            76,
+            186,
+            468,
+            219
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We provide more details about the remaining established objectives (i.e., MLM, VTC, VTM) in the supplementary.",
+        "bbox": [
+            76,
+            220,
+            468,
+            252
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3.2 Stages 2 & 3",
+        "text_level": 1,
+        "bbox": [
+            76,
+            279,
+            218,
+            292
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the subsequent stages, we couple the multimodal expert layers with a pre-trained Flan-T5 $_{\\text{large}}$  [21] via a linear layer. Specifically, Stage 2 aims to align the hidden states of the proposed layers with those of the pre-trained LLM. To this end, we keep the LLM weights frozen and train the whole architecture end-to-end using the generative loss (i.e., next token prediction) on large scale video dialog data<sup>3</sup>, i.e.,",
+        "bbox": [
+            75,
+            306,
+            468,
+            428
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {g e n}} = \\mathbb {E} _ {\\mathbf {X} ^ {\\text {g e n}}} \\left[ \\mathcal {H} \\left(\\mathbf {y} _ {\\rightarrow} ^ {\\text {g e n}}, \\mathbf {p} ^ {\\text {g e n}}\\right)\\right], \\tag {18}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            145,
+            441,
+            468,
+            460
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {X} ^ {\\text {g e n}} = \\Theta_ {\\text {g e n}} \\left(\\mathrm {L L M} _ {\\mathrm {d e c}} \\left(\\left[ \\mathbf {X} ^ {\\text {e n c}}, \\mathbf {T} ^ {\\text {a n s}} \\right]\\right)\\right), \\tag {19}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            143,
+            462,
+            468,
+            482
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{X}^{\\mathrm{enc}}$ ,  $\\mathbf{T}^{\\mathrm{ans}}$  and  $\\Theta_{\\mathrm{gen}}$  are the LLM encoder output, the answer token embeddings, and a linear layer that maps the features to the vocabulary space, respectively.  $\\mathbf{y}_{\\rightarrow}^{\\mathrm{gen}}$  and  $\\mathbf{p}^{\\mathrm{gen}}$  denote the right-shifted ground-truth answer tokens and the predicted text token probabilities. Finally, in Stage 3, we unfreeze the LLM weights and fine-tune our model end-to-end on the downstream tasks of video and visual dialog using the same generative loss.",
+        "bbox": [
+            75,
+            494,
+            468,
+            617
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            636,
+            209,
+            652
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Datasets",
+        "text_level": 1,
+        "bbox": [
+            76,
+            662,
+            179,
+            676
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 3, we simultaneously use the video and image captioning datasets of WebVid-2M [10] and CC-3M [49] to pre-train the multimodal expert layers in Stage 1. Then in the second stage, we use  $25\\%$  of the recent large-scale video dialog dataset Champagne [26] before performing zero-shot evaluation on the widely used video and visual dialog datasets of AVSD [5] and VisDial [22], respectively. Finally, in the third stage, we perform a domain shift evaluation based on different combinations of AVSD and VisDial to quantify whether and to what extent these seemingly similar benchmarks benefit from each other in both zero-shot and fine-tuning evaluation settings.",
+        "bbox": [
+            75,
+            686,
+            468,
+            869
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/4d998f7030b1982834f959adcdc5fd2aefe6a045e5bde942b47c173b0e2ffc1f.jpg",
+        "image_caption": [
+            "Figure 4. Instead of training a dedicated NSP head, we propose a ranking scheme based on the cosine similarity of the candidate answers' embeddings with the respect to those of the generated ones. We used RoBERTa $_{\\text{large}}$  [38] and OpenAI Text Embedding-3 $_{\\text{large}}$  to generate these embeddings."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            87,
+            888,
+            148
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            500,
+            253,
+            683,
+            268
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We use the established official metrics for each dataset to fairly benchmark  $\\mathbf{V}^2\\mathbf{Dial}$  with previous works. Specifically, for all three AVSD datasets, we use BLEU ( $\\mathbf{B - n}$ ) [45], ROUGE-L ( $\\mathbf{R}$ ) [36], METEOR ( $\\mathbf{M}$ ) [11], and CIDEr ( $\\mathbf{C}$ ) [53]. Whereas for VisDial, we use the retrieval metrics of recall ( $\\mathbf{R}@\\mathbf{k}$ ), mean reciprocal rank (MRR), and normalized discounted cumulative gain (NDCG). However, since we are jointly tackling both tasks with a fully generative model, we propose to rank the VisDial candidate answers by means of cosine similarity with respect to the generated answer using the embeddings of a pre-trained sentence transformer (i.e. RoBERTa [38] and OpenAI Text Embedding-3). We posit that this approach is more natural, caters to the current advances in generative models, and appropriately captures the semantic similarities between the generated and the candidate answers. In addition, it allows for a seamless unification of AVSD and VisDial without any training or architectural modifications. As shown in Figure 4, our proposed adaptation does not alter the computation of the sparse metrics itself and only rethinks the ranking of the candidate answers allowing for a fair comparison with previous works.",
+        "bbox": [
+            498,
+            277,
+            892,
+            609
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3. Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            500,
+            618,
+            689,
+            633
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the first stage, we trained our model for a maximum of ten epochs and applied early stopping based on a validation split to select the best checkpoint. In the subsequent stages, we trained it for up to three and twelve epochs, respectively. In all stages, we used the AdamW [39] optimizer with the default parameters and a weight decay value of 0.01. Furthermore, we applied a linear learning rate schedule with warm-up and minimum and base values of  $5e - 5$  and  $1e - 4$ , respectively. We conducted our experiments on a cluster consisting of eight A100 GPUs.",
+        "bbox": [
+            498,
+            641,
+            890,
+            792
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.4. Zero-shot Evaluation",
+        "text_level": 1,
+        "bbox": [
+            500,
+            801,
+            699,
+            815
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "AVSD. We first assessed  $\\mathbf{V}^2$  Dial in a zero-shot setting on AVSD. This is in stark contrast to previous models that were exclusively evaluated in a fine-tuning setting. We instead advocate for complementing the fine-tuning evaluation setting with a zero-shot one, as it results in a more rigorous",
+        "bbox": [
+            498,
+            823,
+            890,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "3The weights of  $\\mathcal{E}_{\\mathrm{ctx}}$  are initialized with those of  $\\mathcal{E}_{\\mathrm{cap}}$  from Stage 1.",
+        "bbox": [
+            94,
+            885,
+            450,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "8641",
+        "bbox": [
+            482,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/3628d6fa70fcd7d764c33a498071bba3d30d2498147150211cda9877803913ad.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"7\">AVSD-DSTC10</td><td colspan=\"7\">AVSD-DSTC8</td><td colspan=\"7\">AVSD-DSTC7</td></tr><tr><td>B-1</td><td>B-2</td><td>B-3</td><td>B-4</td><td>M</td><td>R</td><td>C</td><td>B-1</td><td>B-2</td><td>B-3</td><td>B-4</td><td>M</td><td>R</td><td>C</td><td>B-1</td><td>B-2</td><td>B-3</td><td>B-4</td><td>M</td><td>R</td><td>C</td></tr><tr><td>\\( ^\\dagger \\)MoE-LLAVAarXiv&#x27;24 [35]</td><td>35.8</td><td>18.9</td><td>10.1</td><td>5.9</td><td>15.4</td><td>27.1</td><td>12.8</td><td>39.8</td><td>23.9</td><td>15.2</td><td>10.1</td><td>18.7</td><td>32.2</td><td>23.7</td><td>44.7</td><td>29.1</td><td>19.6</td><td>13.8</td><td>21.8</td><td>37.3</td><td>33.2</td></tr><tr><td>\\( ^\\dagger \\)MiniGPT4-video\\( _{CVPR^{\\prime}24} \\)[9]</td><td>37.9</td><td>19.9</td><td>11.3</td><td>6.8</td><td>16.2</td><td>28.7</td><td>17.7</td><td>34.8</td><td>17.6</td><td>9.7</td><td>5.8</td><td>15.8</td><td>26.3</td><td>13.3</td><td>37.8</td><td>21.2</td><td>12.7</td><td>8.2</td><td>18.4</td><td>30.2</td><td>17.7</td></tr><tr><td>\\( ^\\dagger \\)Video-Chat\\( _{AGL^{\\prime}24} \\)[40]</td><td>24.5</td><td>14.7</td><td>8.8</td><td>5.4</td><td>16.7</td><td>25.2</td><td>3.9</td><td>25.5</td><td>16.0</td><td>10.1</td><td>6.4</td><td>18.4</td><td>27.1</td><td>9.1</td><td>28.5</td><td>18.5</td><td>11.8</td><td>7.6</td><td>20.4</td><td>32.1</td><td>19.1</td></tr><tr><td>\\( ^\\dagger \\)MST-MIXER\\( _{ECCV^{\\prime}24} \\)[3]</td><td>0.1</td><td>0.0</td><td>0.0</td><td>0.0</td><td>3.1</td><td>6.8</td><td>3.0</td><td>0.2</td><td>0.1</td><td>0.1</td><td>0.0</td><td>3.3</td><td>7.1</td><td>4.3</td><td>0.2</td><td>0.1</td><td>0.0</td><td>0.0</td><td>3.4</td><td>6.9</td><td>4.6</td></tr><tr><td>\\( ^\\dagger \\)VideoChat\\( _{CVPR^{\\prime}24} \\)[33]</td><td>42.5</td><td>25.9</td><td>16.0</td><td>10.3</td><td>18.7</td><td>33.1</td><td>25.4</td><td>43.9</td><td>28.1</td><td>18.5</td><td>12.6</td><td>20.8</td><td>34.5</td><td>29.2</td><td>46.7</td><td>31.1</td><td>20.9</td><td>14.4</td><td>22.9</td><td>37.6</td><td>31.4</td></tr><tr><td>\\( {\\mathrm{V}}^{2}\\text{Dial } \\)</td><td>54.6</td><td>34.8</td><td>24.0</td><td>17.2</td><td>19.7</td><td>38.3</td><td>53.8</td><td>53.2</td><td>33.8</td><td>23.5</td><td>16.7</td><td>18.8</td><td>37.7</td><td>49.7</td><td>55.5</td><td>36.7</td><td>26.2</td><td>18.7</td><td>20.0</td><td>39.2</td><td>50.8</td></tr></table>",
+        "bbox": [
+            80,
+            88,
+            893,
+            189
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/2cbf97c4d37413a79a380a8d35af0794bf4976eb91ba250cfb854527e93a5854.jpg",
+        "table_caption": [
+            "Table 2. Zero-shot performance comparison on AVSD-DSTC10, AVSD-DSTC8 and AVSD-DSTC7. Best and second-best performances are in **bold** and **underlined**.  $\\spadesuit$  indicates that we evaluated the model. ( $\\mathbf{B} - \\mathbf{n} = \\mathrm{BLEU} - \\mathbf{n}$ ,  $\\mathbf{M} = \\mathrm{METEOR}$ ,  $\\mathbf{R} = \\mathrm{ROUGE-L}$ ,  $\\mathbf{C} = \\mathrm{CIDEr}$ )."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Sent. Trans.</td><td>R@1</td><td>R@5</td><td>R@10</td><td>MRR</td><td>NDCG</td></tr><tr><td>FROMAGecIMCL&#x27;23 [27]</td><td></td><td>17.6</td><td>20.1</td><td>25.1</td><td>22.0</td><td>16.5</td></tr><tr><td>ESPERCVPR&#x27;23 [62]</td><td>n/a</td><td>14.6</td><td>-</td><td>-</td><td>25.7</td><td>22.3</td></tr><tr><td>ChampagneICCV&#x27;23 [26]</td><td></td><td>-</td><td>-</td><td>-</td><td>-</td><td>25.5</td></tr><tr><td>MoE-LLAVAorXiv&#x27;24 [35]</td><td></td><td>10.6</td><td>25.4</td><td>36.4</td><td>19.6</td><td>26.7</td></tr><tr><td>MiniGPT-videoCVPR&#x27;24 [9]</td><td></td><td>7.4</td><td>17.4</td><td>26.5</td><td>14.6</td><td>23.2</td></tr><tr><td>Video-ChatGPTACL&#x27;24 [40]</td><td>RoBERTa</td><td>10.0</td><td>22.5</td><td>31.5</td><td>18.1</td><td>24.8</td></tr><tr><td>MST-MIXERECCV&#x27;24 [3]</td><td></td><td>18.2</td><td>22.1</td><td>25.7</td><td>21.9</td><td>24.6</td></tr><tr><td>VideoChat2CVPR&#x27;24 [33]</td><td></td><td>12.7</td><td>29.0</td><td>39.9</td><td>22.3</td><td>30.9</td></tr><tr><td rowspan=\"2\">V2 Dial</td><td>RoBERTa</td><td>20.0</td><td>30.2</td><td>39.3</td><td>26.9</td><td>33.3</td></tr><tr><td>OpenAI TE-3</td><td>22.1</td><td>41.2</td><td>48.1</td><td>32.7</td><td>32.0</td></tr></table>",
+        "bbox": [
+            80,
+            252,
+            467,
+            380
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Zero-shot performance comparison on the VisDial v1.0 val split. OpenAI TE-3 = OpenAI Text Embedding-3large.",
+        "bbox": [
+            75,
+            391,
+            468,
+            420
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "and challenging testbed for the proposed models. To this end, we establish; to the best of our knowledge; the first zero-shot benchmark comparison on AVSD comprised of recent capable multimodal models. As can be seen from Table 2, our model outperforms all baselines by a considerable margin across 6/7 metrics of AVSD-DSTC8 and AVSD-DSTC7. On the more recent and challenging version of the benchmark (i.e. AVSD-DSTC10),  $\\mathbf{V}^2$  Dial ranks first across all metrics. For instance, it more than doubles the CIDER score compared to VideoChat2 [33].",
+        "bbox": [
+            75,
+            446,
+            468,
+            599
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "VisDial. Additionally, we assessed the same model checkpoint on VisDial v1.0. As can be seen from Table 3,  $\\mathbf{V}^2\\mathbf{D}$ ial managed to outperform previous models such as FROMAGe [27] by a considerable margin across all metrics of the dataset. In addition, it outperformed Champagne [26] that was trained on x4 more dialog data by 7.8 absolute NDCG points. Furthermore, our model outperformed the more recent baselines of the previous section on 4/5 metrics, underlining its capability of dealing with both video and image input data types. Finally, replacing the sentence embeddings generated by RoBERTaLarge with those from OpenAI Text Emedding-3 improved the external ranking of the candidate answers and resulted in higher scores across all metrics, as can be seen in the last row of Table 3.",
+        "bbox": [
+            75,
+            622,
+            468,
+            834
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.5. Fine-tuning Evaluation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            845,
+            290,
+            862
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "AVSD. Similar to almost all previous works on AVSD, we assessed  $\\mathbf{V}^2$  Dial in a fine-tuning setting on all three bench",
+        "bbox": [
+            75,
+            869,
+            468,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "marks of the dataset. As can be seen from Table 4, our model managed to maintain it competitiveness ahead of recent models and outperformed them on the latest and most challenging AVSD-DSTC10 benchmarks across all evaluation metrics. For instance, it lifted CIDEr by over 6 absolute points compared to the second-best model. Furthermore, our model managed to maintain an on par performance with the state of the art on AVSD-DSTC8 and AVSD-DSTC7. As shown in Table 4,  $\\mathbf{V}^2$  Dial increased their respective CIDEr scores by over 2 and 3 absolute points compared to the second-best model.",
+        "bbox": [
+            496,
+            255,
+            890,
+            421
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "VisDial. Finally, we fine-tuned our model and MST-MIXER [3] that had the closest AVSD performance on Vissdial v1.0 using the same fully-generative approach. As can be seen from Table 5,  $\\mathbf{V}^2\\mathbf{D}\\mathbf{a}$ l managed to outperform all previous models on the strictest metric of the dataset by achieving a R@1 score of 44.2. However, when using OpenAI Text Embedding-3 our model managed to increase the R@1 and MRR scores to 44.9 and 52.4, respectively, thereby setting new state-of-the-art results. As expected and due to the more challenging aspect of a tackling VisDial as a fully generative task, our model performed slightly worse than the previous fine-tuned models on the remaining metrics of the dataset. However, when comparing our model with MST-MIXER that was trained using the same paradigm (i.e. the last two rows of Table 5), we can see that our model outperformed it across 4/5 metrics of the task and scored almost equally on NDCG.",
+        "bbox": [
+            496,
+            441,
+            890,
+            700
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.6. Domain Shift Evaluation",
+        "text_level": 1,
+        "bbox": [
+            500,
+            710,
+            725,
+            724
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Zero-shot setting. First, we fine-tuned our model's checkpoint from Stage 2 on AVSD and zero-shot evaluated it on VisDial. As can be seen from the second section of Table 6, our model's performance was lifted by a considerable margin across most metrics. Notably, the NDCG score improved by 9 absolute points compared to the results of Table 3. Then, we replicated the same experiment on AVSD after having fine-tuned the model on VisDial. Interestingly, our model's performance deteriorated across all metrics of the benchmark. This behavior could be explained by the nature of both datasets. Whereas AVSD encourages the model",
+        "bbox": [
+            496,
+            734,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "8642",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/b640c77073a30ac7c8a082061ffd9670ce0f17a234809a4e288efff57314f03a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            87,
+            472,
+            148
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A man sits at a computer desk and fixes the drawer. [...] He gets up and picks up laundry and leaves.",
+        "bbox": [
+            122,
+            151,
+            452,
+            174
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/456888d629e19c35696f469bc4e80f654893cea814dc4587115b0dd9fe7ea220.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            185,
+            472,
+            281
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/efb17e3b580c2a599c7115d699b6bd2b3be768fc461bd94d9bd143bfeb93e428.jpg",
+        "image_caption": [
+            "AVSD"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            287,
+            472,
+            349
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/1e2b881e39a4ebfd7779271d9e748408f29d4892dd1c171b0d85159e59dcabbd.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            614,
+            87,
+            735,
+            147
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f40284fce39604c17160a4dc0b946904053d6b11889a3054fa3801f57ee9e8ea.jpg",
+        "image_caption": [
+            "VisDial"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            503,
+            150,
+            851,
+            348
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/315a6e95474083edec48a87c8692012d0c30c3e5976708b6720446a7182bcc35.jpg",
+        "image_caption": [
+            "Figure 5. Zero-shot qualitative examples of  $\\mathbf{V}^2\\mathbf{Dial}$  before and after fine-tuning on VisDial and AVSD. The former teaches the model to answer question with brief responses whereas the latter teaches it to produce longer and more elaborate answers.",
+            "Table 4. Fine-tuning performance comparison on AVSD-DSTC10, AVSD-DSTC8 and AVSD-DSTC7. VideoLLAMA 2 [20] was trained on AVSD amongst other datasets. Additional model comparisons can be found in the supplementary material."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            415,
+            888,
+            508
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/1c3d2485ecadeb91e24f89b6565745bba2eb56cbfd9ca9465ebeb9c7c88f263d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Sent. Trans.</td><td>R@1</td><td>R@5</td><td>R@10</td><td>MRR</td><td>NDCG</td></tr><tr><td>\\( LTMI_{ECCV&#x27;20} \\) [44]</td><td></td><td>40.4</td><td>61.6</td><td>69.7</td><td>50.7</td><td>63.5</td></tr><tr><td>\\( LTMI-LGE_{EMNLP&#x27;21} \\) [17]</td><td></td><td>41.3</td><td>61.6</td><td>69.0</td><td>51.3</td><td>63.2</td></tr><tr><td>\\( GoG_{ACL&#x27;21} \\) [16]</td><td>n/a</td><td>41.2</td><td>61.8</td><td>69.4</td><td>51.3</td><td>62.6</td></tr><tr><td>\\( UTC_{CVPR&#x27;22} \\) [15]</td><td></td><td>41.3</td><td>59.8</td><td>66.3</td><td>50.6</td><td>61.0</td></tr><tr><td>Champagne \\( I_{CCV&#x27;23} \\) [26]</td><td></td><td>-</td><td>-</td><td>-</td><td>-</td><td>62.5</td></tr><tr><td>\\( MST-MIXER_{ECCV&#x27;24} \\) [3]</td><td>RoBERTa</td><td>42.2</td><td>51.6</td><td>57.8</td><td>47.7</td><td>52.5</td></tr><tr><td rowspan=\"2\">\\( V^2_Dial \\)</td><td>RoBERTa</td><td>45.4</td><td>54.7</td><td>61.1</td><td>50.9</td><td>54.0</td></tr><tr><td>OpenAI TE-3</td><td>46.1</td><td>59.3</td><td>65.7</td><td>53.2</td><td>53.1</td></tr></table>",
+        "bbox": [
+            81,
+            571,
+            470,
+            676
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 5. Fine-tuning performance comparison on the VisDial v1.0 val split.  $\\spadesuit$  indicates that we trained and evaluated the model.",
+        "bbox": [
+            76,
+            688,
+            468,
+            715
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "to produce long and elaborate responses, VisDial teaches it to produce brief answers instead, which diminishes its performance on the language generation metrics. The qualitative examples of Figure 5 clearly illustrate this phenomenon on both datasets.",
+        "bbox": [
+            75,
+            744,
+            468,
+            818
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Fine-tuning Setting. We first experimented with a curriculum learning strategy where we used one dataset for pretraining before finally fine-tuning on the other. As can be seen from the last section of Table 6, this training paradigm",
+        "bbox": [
+            75,
+            839,
+            470,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "resulted in performance drops on both datasets compared to Table 4 and Table 5 where the model was only trained on the data of the respective benchmark. This indicates that the converged model's weights on one dataset do not offer a good initialization for training on the remaining one. Allowed by our model design that can jointly handle video and image input data, we finally fine-tuned one single model on both datasets simultaneously. As seen from the last row of Table 6, this resulted in the best joint performance of our model across the two datasets. Although the results on AVSD slightly dropped compared to Table 4, our model lifted its performance on VisDial by a considerable margin. This could largely be attributed to the same previous observation, as training on VisDial incentivizes our model to shorten its responses on AVSD.",
+        "bbox": [
+            496,
+            574,
+            893,
+            800
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.7. Expert Swapping",
+        "text_level": 1,
+        "bbox": [
+            500,
+            815,
+            669,
+            832
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In order to validate the specialization of each expert, we conducted a swapping experiment where we routed some features through inadequate experts. We first swapped experts of the same modality (i.e., experts operating on vision",
+        "bbox": [
+            496,
+            839,
+            893,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "8643",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/3887422fc0b03ab50097ddbe52b1acf7b9b2897bfbf1900b351ea3abf063bf9d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Fine-tuning data</td><td colspan=\"4\">AVSD-DSTC10</td><td colspan=\"4\">AVSD-DSTC8</td><td colspan=\"4\">AVSD-DSTC7</td><td colspan=\"4\">VisDial</td><td></td></tr><tr><td>AVSD</td><td>VisDial</td><td>B-1</td><td>M</td><td>R</td><td>C</td><td>B-1</td><td>M</td><td>R</td><td>C</td><td>B-1</td><td>M</td><td>R</td><td>C</td><td>R@1</td><td>R@5</td><td>R@10</td><td>NDCG</td><td></td></tr><tr><td rowspan=\"2\">X</td><td rowspan=\"2\">X</td><td colspan=\"16\">Zero-shot (from Table 2 and Table 3)</td><td></td></tr><tr><td>54.6</td><td>19.7</td><td>38.3</td><td>53.8</td><td>53.2</td><td>18.8</td><td>37.7</td><td>49.7</td><td>55.5</td><td>20.0</td><td>39.2</td><td>50.8</td><td>20.0</td><td>30.2</td><td>39.3</td><td>33.3</td><td></td></tr><tr><td rowspan=\"2\">✓</td><td rowspan=\"2\">X</td><td colspan=\"12\">Fine-tuning (from Table 4)</td><td colspan=\"4\">Zero-shot</td><td></td></tr><tr><td>70.7</td><td>26.0</td><td>55.4</td><td>103.3</td><td>76.8</td><td>30.4</td><td>62.1</td><td>135.7</td><td>78.9</td><td>31.2</td><td>62.3</td><td>139.8</td><td>12.8</td><td>36.7</td><td>50.8</td><td>42.3</td><td></td></tr><tr><td rowspan=\"2\">X</td><td rowspan=\"2\">✓</td><td colspan=\"12\">Zero-shot</td><td colspan=\"4\">Fine-tuning (from Table 5)</td><td></td></tr><tr><td>11.5</td><td>6.8</td><td>20.1</td><td>14.6</td><td>11.5</td><td>7.3</td><td>20.7</td><td>20.9</td><td>7.9</td><td>6.2</td><td>17.4</td><td>18.2</td><td>44.2</td><td>53.3</td><td>59.5</td><td>52.3</td><td></td></tr><tr><td>✓</td><td>→</td><td>✓</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>42.2</td><td>50.1</td><td>56.3</td><td>51.3</td><td></td></tr><tr><td>✓</td><td>←</td><td>✓</td><td>69.6</td><td>25.7</td><td>55.0</td><td>100.5</td><td>75.9</td><td>29.8</td><td>61.4</td><td>132.1</td><td>77.6</td><td>30.4</td><td>61.5</td><td>134.5</td><td>-</td><td>-</td><td>-</td><td></td></tr><tr><td>✓</td><td>&amp;</td><td>✓</td><td>69.3</td><td>25.4</td><td>54.8</td><td>99.9</td><td>75.1</td><td>29.3</td><td>61.1</td><td>130.0</td><td>77.3</td><td>30.0</td><td>61.7</td><td>134.5</td><td>45.4</td><td>54.7</td><td>61.1</td><td>54.0</td></tr></table>",
+        "bbox": [
+            109,
+            88,
+            859,
+            233
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/ec05b3a151d0e78de3d34a68c3f41b07913f9304402af2b230ffaa6c55c14c24.jpg",
+        "table_caption": [
+            "Table 6. Domain shift evaluation between the respective most prominent video and visual dialog datasets of AVSD and VisDial.  $\\square \\rightarrow \\triangle$  means that the model was pre-trained on dataset  $\\square$  before fine-tuning on dataset  $\\triangle$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\" colspan=\"3\">Expert\nSwapping</td><td colspan=\"4\">AVSD-DSTC7</td><td colspan=\"2\">VisDial</td></tr><tr><td>B-1</td><td>M</td><td>R</td><td>C</td><td>R@1</td><td>NDCG</td></tr><tr><td colspan=\"3\">Original</td><td>78.9</td><td>31.2</td><td>62.3</td><td>139.8</td><td>44.2</td><td>52.3</td></tr><tr><td colspan=\"9\">Swapping experts of the same modality (vision / language)</td></tr><tr><td>Espa←</td><td>Etmp</td><td>77.0</td><td>29.5</td><td>61.2</td><td>133.7</td><td>-</td><td>-</td><td></td></tr><tr><td>Ecap←</td><td>Ectx</td><td>76.1</td><td>29.6</td><td>60.3</td><td>131.1</td><td>42.8</td><td>51.9</td><td></td></tr><tr><td colspan=\"9\">Swapping experts of different modalities</td></tr><tr><td>Espa←</td><td>Ecap</td><td>28.5</td><td>10.7</td><td>22.0</td><td>10.1</td><td>35.7</td><td>45.3</td><td></td></tr><tr><td>Etmp←</td><td>Ectx</td><td></td><td></td><td></td><td></td><td>-</td><td>-</td><td></td></tr><tr><td>Espa←</td><td>Ectx</td><td>34.4</td><td>12</td><td>25.4</td><td>11.7</td><td>32.4</td><td>42.9</td><td></td></tr><tr><td>Ecap←</td><td>Ecap</td><td></td><td></td><td></td><td></td><td>-</td><td>-</td><td></td></tr></table>",
+        "bbox": [
+            84,
+            294,
+            460,
+            449
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "or language data). As shown in Table 7, this resulted in performance drops across all metrics of both datasets, indicating that experts of the same modality are able to capture the semantic nuances of the data they specialize on. More interestingly, the performance of our model dropped more significantly when swapping experts of different modalities, as seen from the last section of Table 7. This showcases their ability to adjust to the nature of the data they process and to capture its modality specific features.",
+        "bbox": [
+            75,
+            530,
+            468,
+            667
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.8. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            76,
+            679,
+            230,
+            696
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of Pre-training Data. We trained two versions of our model, where one was only pre-trained on Stage 1 using WebVid-2M & CC-3M and the other only on Stage 2 with a subset of Champagne. As can be seen from the middle section of Table 8, our model witnessed a comparable drop in performance compared to the full model. This underlines the equal importance of these proposed training stages to the joint down-stream performance on AVSD and VisDial. We did not conduct ablations using either WebVid-2M or CC-3M in Stage 1 as this was sufficiently explored by other recent works [19] that showed the benefit of pre-training on both image and video data.",
+        "bbox": [
+            75,
+            703,
+            470,
+            885
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of Pre-training Objectives & Model Design. Then,",
+        "bbox": [
+            76,
+            885,
+            468,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/3f9988f2cc4d5aed314961503af32b5481b2636091810cc1e91526c00a3341f7.jpg",
+        "table_caption": [
+            "Table 7. Expert swapping results.  $\\mathcal{E}_{\\square} \\leftrightarrow \\mathcal{E}_{\\triangle}$  means that the  $\\square$  features are inadequately routed at test time through  $\\mathcal{E}_{\\triangle}$  and vice versa. The other experts remain unchanged."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model Ablations</td><td colspan=\"4\">AVSD-DSTC7</td><td colspan=\"2\">VisDial</td></tr><tr><td>B-1</td><td>M</td><td>R</td><td>C</td><td>R@1</td><td>NDCG</td></tr><tr><td>Full</td><td>78.9</td><td>31.2</td><td>62.3</td><td>139.8</td><td>44.2</td><td>52.3</td></tr><tr><td>w/o Tr. Stage 1</td><td>76.9</td><td>30.0</td><td>61.4</td><td>134.0</td><td>34.5</td><td>44.6</td></tr><tr><td>w/o Tr. Stage 2</td><td>77.8</td><td>30.7</td><td>61.7</td><td>134.7</td><td>32.6</td><td>44.1</td></tr><tr><td>w/o Lstc &amp; Lstm</td><td>77.2</td><td>29.9</td><td>61.1</td><td>133.2</td><td>33.1</td><td>44.6</td></tr><tr><td>w/o separate Espa &amp; Etmp</td><td>77.0</td><td>30.1</td><td>61.1</td><td>133.8</td><td>32.9</td><td>43.8</td></tr><tr><td>w/o experts {E*}</td><td>77.5</td><td>30.0</td><td>61.4</td><td>134.8</td><td>30.6</td><td>42.2</td></tr></table>",
+        "bbox": [
+            514,
+            294,
+            879,
+            395
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 8. Ablation results.",
+        "bbox": [
+            617,
+            405,
+            772,
+            419
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "we trained a version of our model without  $\\mathcal{L}_{\\mathrm{stc}}$  and  $\\mathcal{L}_{\\mathrm{stm}}$  in Stage 1 using the same schedule and training data as our full model. As shown in the fourth row of Table 8, this ablated version suffered a performance drop not only in AVSD but also in VisDial. This indicates that these losses improve not only the temporal capabilities of our model but also its spatial ones. Then, we trained a version that sequentially applies spatial and temporal attention, as in [13, 19]. Since this version does not have separate spatial-temporal experts, we also omitted the previous two objectives. As seen in the penultimate row of Table 8, this version underperformed our full model on both datasets, showcasing the effectiveness of our approach. Finally, we trained a version without all the expert layers. As shown in the last row, its performance dropped compared to our full model and performed the worst on VisDial.",
+        "bbox": [
+            496,
+            448,
+            893,
+            690
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            707,
+            619,
+            723
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We presented  $\\mathbf{V}^2\\mathbf{Dial}$  - a model that can jointly tackle video and visual conversational tasks using a multimodal expert-based approach that; for the first time, disentangles the learning of the spatial and temporal features of images and videos using two separate experts. Extensive evaluation on the respective widely used video and visual dialog datasets of AVSD and VisDial show that our model achieves new state-of-the-art zero-shot and fine-tuning performance. Finally, we conducted the first domain shift evaluation of AVSD and VisDial and provided insights on how to optimally leverage their respective training data.",
+        "bbox": [
+            496,
+            733,
+            893,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "8644",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgment",
+        "text_level": 1,
+        "bbox": [
+            78,
+            90,
+            228,
+            107
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work was partially funded by a LOEWE-Start-Professur (LOEWE/4b//519/05.01.002-(0006)/94), LOEWE-Spitzen-Professur (LOEWE/4a//519/05.00.002-(0010)/93), and an Alexander von Humboldt Professorship in Multimodal Reliable AI, sponsored by Germany's Federal Ministry for Education and Research.",
+        "bbox": [
+            76,
+            114,
+            468,
+            205
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            218,
+            173,
+            234
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Adnen Abdessaied, Mihai Bace, and Andreas Bulling. Neuro-Symbolic Visual Dialog. In Proceedings of the International Conference on Computational Linguistics (COLING), 2022. 1",
+            "[2] Adnen Abdessaied, Manuel Hochmeister, and Andreas Bulling. OLViT: Multi-Modal State Tracking via Attention-Based Embeddings for Video-Grounded Dialog. In Proceedings of the Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING), 2024. 2",
+            "[3] Adnen Abdessaied, Lei Shi, and Andreas Bulling. MultiModal Video Dialog State Tracking in the Wild. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 6, 7",
+            "[4] Adnen Abdessaied, Lei Shi, and Andreas Bulling. VD-GR: Boosting Visual Dialog With Cascaded Spatial-Temporal Multi-Modal Graphs. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024. 1, 2",
+            "[5] Huda Alamri, Vincent Cartillier, Abhishek Das, Jue Wang, Anoop Cherian, Irfan Essa, Dhruv Batra, Tim K Marks, Chiori Hori, Peter Anderson, et al. Audio visual scene-aware dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 5",
+            "[6] Huda Alamri, Anthony Bilic, Michael Hu, Apoorva Beedu, and Irfan Essa. End-to-end multimodal representation learning for video dialog. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2",
+            "[7] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1",
+            "[8] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: Visual Question Answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2015. 1, 2",
+            "[9] Kirolos Ataallah, Xiaogian Shen, Eslam Abdelrahman, Essam Sleiman, Deyao Zhu, Jian Ding, and Mohamed Elho"
+        ],
+        "bbox": [
+            86,
+            243,
+            470,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "seiny. MiniGPT4-video: Advancing multimodal llms for video understanding with interleaved visual-textual tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024. 6",
+            "[10] Max Bain, Arsha Nagrani, Güll Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 5",
+            "[11] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, 2005. 5",
+            "[12] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. VLMo: Unified vision-language pre-training with mixture-of-modality-experts. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2",
+            "[13] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is Space-Time Attention All You Need for Video Understanding? In Proceedings of the International Conference on Machine Learning (ICML), 2021. 3, 8",
+            "[14] Soravit Changpinyo, Piyush Sharma, and Nan Ding and Radu Soricut. Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1",
+            "[15] Cheng Chen, Yudong Zhu, Zhenshan Tan, Qingrong Cheng, Xin Jiang, Qun Liu, and Xiaodong Gu. UTC: A Unified Transformer with Inter-Task Contrastive Learning for Visual Dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 7",
+            "[16] Feilong Chen, Xiuyi Chen, Fandong Meng, Peng Li, and Jie Zhou. GoG: Relation-aware graph-over-graph network for visual dialog. In Findings of the Association for Computational Linguistics: ACL, 2021. 1, 2, 7",
+            "[17] Feilong Chen, Xiuyi Chen, Can Xu, and Daxin Jiang. Learning to ground visual objects for visual dialog. In Findings of the Association for Computational Linguistics: EMNLP, 2021. 7",
+            "[18] Zhe Chen, Hongcheng Liu, and Yu Wang. DialogMCF: Multimodal Context Flow for Audio Visual Scene-Aware Dialog. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2023. 2, 7",
+            "[19] Feng Cheng, Xizi Wang, Jie Lei, David Crandall, Mohit Bansal, and Gedas Bertasius. VindLU: A Recipe for Effective Video-and-Language Pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4, 8",
+            "[20] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs. arXiv preprint arXiv:2406.07476, 2024. 2, 7"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "8645",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[21] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research (JMLR), 2024. 5",
+            "[22] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José M.F. Moura, Devi Parikh, and Dhruv Batra. Visual Dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2, 5",
+            "[23] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2019. 1",
+            "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2020. 1",
+            "[25] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan First, et al. GLaM: Efficient scaling of language models with mixture-of-experts. In Proceedings of the International Conference on Machine Learning (ICML), 2022. 2",
+            "[26] Seungju Han, Jack Hessel, Nouha Dziri, Yejin Choi, and Youngjae Yu. Champagne: Learning real-world conversation from large-scale web videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 5, 6, 7",
+            "[27] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. In Proceedings of the International Conference on Machine Learning (ICML), 2023. 6",
+            "[28] Hung Le and Steven C.H. Hoi. Video-Grounded Dialogues with Pretrained Generation Language Models. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2020. 2",
+            "[29] Hung Le, Nancy F. Chen, and Steven Hoi. Learning reasoning paths over semantic graphs for video-grounded dialogues. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2, 7",
+            "[30] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2020. 2",
+            "[31] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 4",
+            "[32] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            902
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "frozen image encoders and large language models. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 1, 4",
+            "[33] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. MVBench: A Comprehensive Multi-modal Video Understanding Benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 6",
+            "[34] Zekang Li, Zongjia Li, Jinchao Zhang, Yang Feng, and Jie Zhou. Bridging text and video: A universal multimodal transformer for audio-visual scene-aware dialog. Transactions on Audio, Speech, and Language Processing, 2021. 2",
+            "[35] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Jina Huang, Junwu Zhang, Yatian Pang, Munan Ning, and Li Yuan. MoE-LLaVA: Mixture of Experts for Large Vision-Language Models. arXiv preprint arXiv:2401.15947, 2024. 2, 6",
+            "[36] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, 2004. 5",
+            "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1",
+            "[38] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. arXiv preprint arXiv 1907.11692, 2019.5",
+            "[39] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. In Proceedings of the International Conference on Learning Representations (ICLR), 2019. 5",
+            "[40] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-ChatGPT: Towards Detailed Video Understanding via Large Vision and Language Models. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2024. 6",
+            "[41] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1",
+            "[42] Vishvak Murahari, Dhruv Batra, Devi Parikh, and Abhishek Das. Large-scale pretraining for visual dialog: A simple state-of-the-art baseline. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2",
+            "[43] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal Contrastive Learning with LIMoE: the Language-Image Mixture of Experts. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2",
+            "[44] Van Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Efficient Attention Mechanism for Visual Dialog that Can Handle All the Interactions Between Multiple Inputs. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 7",
+            "[45] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            901
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "8646",
+        "bbox": [
+            482,
+            944,
+            516,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "translation. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2002. 5",
+            "[46] Hoang-Anh Pham, Thao Minh Le, Vuong Le, Tu Minh Phuong, and Truyen Tran. Video Dialog as Conversation about Objects Living in Space-Time. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2",
+            "[47] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 2",
+            "[48] Ankit P. Shah, Shijie Geng, Peng Gao, Anoop Cherian, Takaaki Hori, Tim K. Marks, Jonathan Le Roux, and Chiori Hori. Audio-Visual Scene-Aware Dialog and Reasoning using Audio-Visual Transformers with Joint Student-Teacher Learning. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022. 1",
+            "[49] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2018. 5",
+            "[50] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1",
+            "[51] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. EVA-CLIP: Improved Training Techniques for CLIP at Scale. arXiv preprint arXiv:2303.15389, 2023. 3",
+            "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 1",
+            "[53] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based Image Description Evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 5",
+            "[54] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2",
+            "[55] Yue Wang, Shafiq R. Joty, Michael R. Lyu, Irwin King, Caiming Xiong, and Steven C. H. Hoi. VD-BERT: A Unified Vision and Dialog Transformer with BERT. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 2",
+            "[56] Zihao Wang, Junli Wang, and Changjun Jiang. Unified multimodal model with unlikelihood training for visual dialog. In Proceedings of the ACM International Conference on Multimedia (ACM MM), 2022. 1, 2",
+            "[57] Ronald J Williams and David Zipser. A learning algorithm for continually running fully recurrent neural networks. Neural Comput., 1989. 2"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            900
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[58] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the ACM International Conference on Multimedia (ACM MM), 2017. 1, 2",
+            "[59] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1",
+            "[60] Sunjae Yoon, Eunseop Yoon, Hee Suk Yoon, Junyeong Kim, and Chang Yoo. Information-theoretic text hallucination reduction for video-grounded dialogue. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2022. 2, 7",
+            "[61] Xintong Yu, Hongming Zhang, Ruixin Hong, Yangqiu Song, and Changshui Zhang. VD-PCR: Improving visual dialog with pronoun coreference resolution. Pattern Recognition, 2022. 2",
+            "[62] Youngjae Yu, Jiwan Chung, Heeseung Yun, Jack Hessel, Jae Sung Park, Ximing Lu, Rowan Zellers, Prithviraj Ammanabrolu, Ronan Le Bras, Gunhee Kim, et al. Fusing pretrained language models with multimodal prompts through reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6",
+            "[63] Zexuan Zhong, Mengzhou Xia, Danqi Chen, and Mike Lewis. Lory: Fully Differentiable Mixture-of-Experts for Autoregressive Language Model Pre-training. In Proceedings of the Conference on Language Modeling (COLM), 2024. 2",
+            "[64] Tong Zhu, Xiaoye Qu, Daize Dong, Jiacheng Ruan, Jingqi Tong, Conghui He, and Yu Cheng. Llama-moe: Building mixture-of-experts from llama with continual pre-training. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2024. 2"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            583
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "8647",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2025/V^2Dial_ Unification of Video and Visual Dialog via Multimodal Experts/a2fbca87-c18a-41dc-864b-f66927e9f633_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d33567835efb1aff5dd563d9eae7d5da9daefd7655d906fe63c745ee430488ba
+size 2249550

2025/V^2Dial_ Unification of Video and Visual Dialog via Multimodal Experts/full.md

@@ -0,0 +1,358 @@
+# $\mathbf{V}^{2}$  Dial: Unification of Video and Visual Dialog via Multimodal Experts
+
+![](images/eb6eb8aaa27e4b0f18b08931633881735cc23e776a2800be52f4fd8d37365f5c.jpg)
+
+# Abstract
+
+We present  $V^2$  Dial - a novel expert-based model specifically geared towards simultaneously handling image and video input data for multimodal conversational tasks. Current multimodal models primarily focus on simpler tasks (e.g., VQA, VideoQA, video-text retrieval) and often neglect the more challenging conversational counterparts, such as video and visual/image dialog. Moreover, works on both conversational tasks evolved separately from each other despite their apparent similarities, limiting their applicability potential. To this end, we propose to unify both tasks using a single model that for the first time jointly learns the spatial and temporal features of images and videos by routing them through dedicated experts and aligns them using matching and contrastive learning techniques. Furthermore, we systemically study the domain shift between the two tasks by investigating whether and to what extent these seemingly related tasks can mutually benefit from their respective training data. Extensive evaluations on the widely used video and visual dialog datasets of AVSD and VisDial show that our model achieves new state-of-the-art results across four benchmarks both in zero-shot and fine-tuning settings.
+
+# 1. Introduction
+
+Enabled by the availability of large-scale training data [10, 14, 41] and advances in model design [12, 24, 37, 48, 52], the field of vision-and-language learning saw unprecedented success in recent years. However, current multimodal foundational models [7, 32, 37, 50, 54] still mainly focus on single-round tasks (e.g., VQA [8], VideoQA [58], video-text and text-video retrieval [59]). In contrast, the significantly more challenging conversational tasks, such as visual [1, 22] and video dialog [5], received considerably less attention. Furthermore, methods for these different tasks have advanced independently of each other despite the apparent structural similarities between them. They both operate on a visual input (i.e. an image or video), a short visual description (caption), and a dialog history composed of
+
+![](images/269c233cda7f0815cbdd2c0db7d53c4ec50f90845f5b7802b8cf9e8a9cff3f6c.jpg)
+
+![](images/8856e672e3d2aa85751f93c17eaf2ef9190b679e6e3170682ac2f180819d8ce0.jpg)  
+Figure 1.  $\mathbf{V}^2\mathbf{D}$ ial uses multimodal experts and outperforms state-of-the-art methods on both video and visual dialog in zero-shot and fine-tuning evaluation settings.
+
+previous question-answer pairs. On the one hand, visual dialog models [4, 15, 16, 44, 56] have been primarily trained to rank a list of candidate answers using a Next Sentence Prediction (NSP) head similar to BERT [23] and negative sampling. Thus, they are benchmarked using retrieval met
+
+rics such as recall (R@k) and normalized discounted cumulative gain (NDCG). In contrast, video dialog models [2, 3, 6, 18, 20, 29, 46, 60] are trained to auto-regressively predict the next answer token using teacher forcing [57] and are evaluated using language generation metrics.
+
+In this work, we mitigate the shortcomings of current dialog systems by proposing  $\mathbf{V}^2\mathbf{Dial}$  - a novel multimodal expert-based model capable of unifying video and visual dialog tasks without any architectural changes. Specifically, we train dedicated multimodal expert layers that separately process the features of each input modality and learn how to align them using matching and contrastive learning techniques. A key novelty of our approach is that we use dedicated experts to jointly learn the spatial and temporal features of images and videos by routing them through the appropriate experts. Then, we couple these layers with a pretrained LLM to align their hidden states. Thanks to its modularity, our model can efficiently tackle image and video input data simultaneously and seamlessly train on both data types. In summary, our contributions are three-fold:
+
+- We propose  $\mathbf{V}^2\mathbf{Dial}$  - a multimodal expert-based model that unifies visual and video dialog by simultaneously learning from image and video data. As a core novelty, it employs two experts to separately learn the spatial and temporal features of images and videos.  $\mathbf{V}^2\mathbf{Dial}$  outperforms state-of-the-art models in both zero-shot and fine-tuning settings (see Figure 1).  
+- We are the first to systematically quantify the effect of domain shift between video and visual dialog tasks based on evaluations on the two widely used datasets of AVSD [5] and VisDial [22]. To this end, we propose an alternative ranking scheme that allows computing the VisDial retrieval metric for fully generative models and enables a fair comparison with previous works.  
+- We are the first to evaluate AVSD in a zero-shot setting, which provides a more solid generalization evaluation of video dialog models compared to the fine-tuned setting. For this, we establish the first benchmark comparison of recent state-of-the-art multimodal models.
+
+# 2. Related Work
+
+Visual & Video Dialog. Modeled after human-human communication, visual and video dialog involve reasoning about a visual scene in the form of a video or an image through multiple question-answering rounds in natural language. In comparison to their single-round counterparts, VQA [8] and VideoQA [58], dialog models need to additionally reason about the previous dialog history together with the visual grounding and the current question to be able to answer it efficiently. The best performing visual dialog models [4, 42, 56, 61] leverage pre-trained VLMs and are trained using an NSP head, negative sampling, and binary classification loss. At test time, for each question, the can
+
+didate answers are ranked based on their respective NSP scores to compute the retrieval metrics. Although some work [15, 16, 55] claim to train generative visual dialog models, they do so by providing a generative mask where each token can only attend to its left tokens. However, they are trained using the NSP head like the discriminative models. However, this training approach is limiting and suboptimal for a unifying model. Thus, we advocate for a fully generative training paradigm and adapt the ranking scheme of VisDial answers to cater to modern generative models.
+
+In contrast, works on video dialog follow a purely generative training paradigm and achieved great success building on top of powerful pre-trained LLMs [30, 47]. For example, [28, 34] fine-tuned a LLM on AVSD and obtained performance boosts. More recent works [3, 29] combined LLMs with GNNs and pushed the state-of-the-art results even further. Others [60] introduced a regularization loss to mitigate hallucination. Although video dialog emerged as a natural extension to visual dialog with apparent data structure similarities, research on both tasks evolved separately. To this end, we propose a unifying model that can simultaneously learn both tasks without any architectural modifications and for the first time; systemically study the effect of domain shift between both tasks using the AVSD and VisDial v1.0 datasets.
+
+Multimodal Expert-based Training. Enhancing models with expert-based training has shown promising potential in boosting performance while maintaining computational efficiency [25, 63, 64]. Some works [12, 54] explored using single modality specific experts within a multimodal transformer architecture. Specifically, they used one vision and one language specific FFN after a shared multi-head self-attention block. Other works [35, 43] explored using multiple sparse modality-agnostic experts and trained them using soft-routers. Our work is positioned at the middle ground of the previously mentioned research directions: We propose to use multiple hard-routed experts per modality to be able to capture more fine-grained features compared to a single expert or multiple modality agnostic experts. Specifically, to the best of our knowledge,  $\mathbf{V}^2\mathbf{Dial}$  is the first model that learns disentangled spatial and temporal features using two dedicated experts that jointly learn from image and video data. In addition, we propose to deploy two separate language experts (for caption and context) in order to tackle the unique challenges of multimodal conversational tasks.
+
+# 3.  $\mathbf{V}^2$  Dial
+
+# 3.1. Joint Problem Formulation
+
+We use a fully generative formulation to unify both video and visual dialog tasks. Specifically, given visual input V (video/image), a corresponding visual description (caption
+
+![](images/76d158b2b310140434bcb6f983d2d8afde34438791dbe9bb380d981608e610be.jpg)  
+Tr. Stage 1  
+Figure 2. Architectural overview of  $\mathbf{V}^2$  Dial. We adopt a training strategy composed of three stages. First, we only train the multimodal expert layers using spatial-temporal and video/image text matching losses ( $\mathcal{L}_{\mathrm{stm}}$ ,  $\mathcal{L}_{\mathrm{vm}}$ ), spatial-temporal and video/image contrastive learning losses ( $\mathcal{L}_{\mathrm{stc}}$ ,  $\mathcal{L}_{\mathrm{vtc}}$ ), and masked language modeling loss ( $\mathcal{L}_{\mathrm{mlm}}$ ). Second, we couple the expert layers with a frozen pre-trained LLM end-to-end, using a generative loss  $\mathcal{L}_{\mathrm{gen}}$  to align their hidden representations. Finally, we additionally fine-tune the LLM weights on the downstream benchmarks. Each expert is a feed-forward network (FFN) composed of two fully connected layers.
+
+C), a dialog history  $\mathrm{H}_r = \left\{\left(\mathbb{Q}_1,\mathbb{A}_1\right),\dots,\left(\mathbb{Q}_{r - 1},\mathbb{A}_{r - 1}\right)\right\}$  comprised of the previous question-answer pairs  $\left\{\left(\mathbb{Q}_i,\mathbb{A}_i\right)\right\}_{i = 1}^{r - 1}$  and the current question  $\mathbb{Q}_r$  , a model is trained to auto-regressively predict a free-form answer  $\mathbb{A}_r$  at round  $r$  Specifically, each answer token  $\mathsf{a}_r^i$  satisfies
+
+$$
+\mathsf {a} _ {r} ^ {i} = \underset {\mathsf {a} \in \mathcal {V}} {\arg \max } \left[ \mathbf {p} \left(\mathsf {a} | \mathrm {V}, \mathrm {C}, \mathrm {H} _ {r} \mathsf {Q} _ {r}, \mathrm {A} _ {r} ^ {<   i}\right) \right], \tag {1}
+$$
+
+where  $\mathbb{A}_r^{<i}$  denotes the previously predicted answer tokens and  $\mathcal{V}$  the vocabulary. In the rest, we use "context" to refer to the concatenation of the history  $\mathbb{H}_r$  and the question  $\mathbb{Q}_r$ .
+
+# 3.2. Architecture
+
+Overview. As can be seen from Figure 2, our model takes an image/video  $\mathbf{v} \in \mathbb{R}^{F \times 3 \times H \times W}$  as input, where  $F$  is the number of frames and is set to one for images, and  $(H, W)$  is the re-sized resolution. Then it processes every frame using a pre-trained EVA-CLIP [51] Image Encoder and concatenates every four spatially adjacent visual patches into a single one. Then, a linear layer maps each visual token into a lower dimensional vector  $\mathbf{v}$  of dimension  $D$  to obtain
+
+$$
+\mathbf {V} = \left[ \begin{array}{c c c c} \mathbf {v} _ {1} ^ {1} & \mathbf {v} _ {1} ^ {2} & \dots & \mathbf {v} _ {1} ^ {F} \\ \vdots & \vdots & \ddots & \vdots \\ \mathbf {v} _ {P} ^ {1} & \mathbf {v} _ {P} ^ {2} & \dots & \mathbf {v} _ {P} ^ {F} \end{array} \right] \in \mathbb {R} ^ {F \times P \times D}, \qquad (2)
+$$
+
+where  $P = \frac{1}{4} \frac{H \times W}{14^2}$  and  $D$  denote the visual input length and the joint hidden dimension, respectively. Thereafter, in stark contrast to previous works [13, 19] that performed spatial and temporal attention in series, our model separately
+
+performs these operations using the masks  $\mathbf{M}^{\mathrm{spa}}$  and  $\mathbf{M}^{\mathrm{tmp}}$  as shown in Figure 2 on the visual features  $V$  to obtain
+
+$$
+\mathbf {V} ^ {\mathrm {s p a}} = \operatorname {S A} (\mathbf {V}, \mathbf {M} ^ {\mathrm {s p a}}) \in \mathbb {R} ^ {(F P) \times D} \tag {3}
+$$
+
+$$
+\mathbf {V} ^ {\mathrm {t m p}} = \operatorname {S A} (\mathbf {V}, \mathbf {M} ^ {\mathrm {t m p}}) \in \mathbb {R} ^ {(F P) \times D} \tag {4}
+$$
+
+$$
+\mathbf {M} _ {m, n} ^ {\mathrm {s p a}} \left(\mathbf {v} _ {i} ^ {j}\right) = \delta_ {n j}, \quad \mathbf {M} _ {m, n} ^ {\mathrm {t m p}} \left(\mathbf {v} _ {i} ^ {j}\right) = \delta_ {m i} \tag {5}
+$$
+
+where SA and  $\delta$  denote self-attention and Kronecker delta.
+
+Subsequently, the textual input in the form of a caption and a context is processed by an embedding layer to obtain  $\mathbf{T}^{\mathrm{cap / ctx}}\in \mathbb{R}^{N_{\mathrm{cap / ctx}}\times D}$ , where  $N_{\mathrm{cap}}$  and  $N_{\mathrm{ctx}}$  are the respective lengths of the caption and context. These visual and textual features form the initial input to the multimodal expert layers which are pre-trained using a combination of matching, contrastive learning, and masked language modeling losses. Finally, they are coupled with a pre-trained LLM and are fine-tuned end-to-end using a generative loss.
+
+Multimodal Expert Layers. These consist of  $N$  layers of stacked multi-head self-attention with layer normalization (SA), and several modality-specific and one modality-agnostic feed-forward networks that we refer to as experts. As shown in Figure 2, we propose to use a set of six experts denoted as  $\{\mathcal{E}_*\}$ : three of which are vision-specific and two are language-specific and are activated in the first  $L$  layers. The remaining expert  $\mathcal{E}_{\mathrm{fus}}$  is the fusion expert and is only activated in the last  $(N - L)$  layers and operates on the concatenation of all available modalities (Equation 9). To the best of our knowledge, we propose for the first time to learn the spatial and temporal features using dedicated
+
+![](images/c58a252e17c236379fd3298fe8f998ed1bbe65c274a81671425a45ccb87e9054.jpg)  
+Figure 3. Overview of the training and evaluation pipeline of  $\mathbf{V}^2$ Dial. We show the different datasets used to train our model at each stage. Evaluations are conducted on the most popular video and visual dialog datasets of AVSD and VisDial, respectively. ( = video data, = image data, = closed / visual captioning data, = dialog data).
+
+experts (i.e., the spatial  $\mathcal{E}_{\mathrm{spa}}$  and temporal  $\mathcal{E}_{\mathrm{tmp}}$  experts, respectively) as shown in Equation 11. This allows our model to unify video and visual dialog by jointly learning from image and video data. The visual expert  $\mathcal{E}_{\mathrm{vis}}$  operates on top of the concatenation of  $\mathcal{E}_{\mathrm{spa}}$  and  $\mathcal{E}_{\mathrm{tmp}}$  to learn a joint spatial-temporal video representation (Equation 10). Similarly, the textual experts  $\mathcal{E}_{\mathrm{cap}}$  and  $\mathcal{E}_{\mathrm{ctx}}$  operate on the caption and context embeddings  $\mathbf{T}^{\mathrm{cap}}$  and  $\mathbf{T}^{\mathrm{ctx}}$  (Equation 12). As seen in Table 1, the availability of the multimodal features depends on the visual input type (i.e., videos vs images) and the training stage. However, without the loss of generality, we can formulate the multimodal expert layers as follows:
+
+$$
+\mathbf {X} _ {0} = \left[ \mathbf {V} ^ {\text {s p a}}, \mathbf {V} ^ {\text {t m p}}, \mathbf {T} ^ {\text {c a p}}, \mathbf {T} ^ {\text {c t x}} \right], \tag {6}
+$$
+
+$$
+\begin{array}{l} \tilde {\mathbf {X}} _ {l} = \left[ \tilde {\mathbf {V}} _ {l} ^ {\mathrm {s p a}}, \tilde {\mathbf {V}} _ {l} ^ {\mathrm {t m p}}, \tilde {\mathbf {T}} _ {l} ^ {\mathrm {c a p}}, \tilde {\mathbf {T}} _ {l} ^ {\mathrm {e x t}} \right] (7) \\ = \operatorname {S A} \left(\mathbf {X} _ {l - 1}\right) + \mathbf {X} _ {l - 1} (8) \\ \end{array}
+$$
+
+$$
+\mathbf {X} _ {l} = \left\{ \begin{array}{l l} \left[ \mathbf {V} _ {l} ^ {\text {v i s}}, \mathbf {T} _ {l} ^ {\text {c a p}}, \mathbf {T} _ {l} ^ {\text {c t x}} \right] & i f 1 \leq l \leq L \\ \mathcal {E} _ {\text {f u s}} \left(\tilde {\mathbf {X}} _ {l}\right) + \tilde {\mathbf {X}} _ {l} & i f L <   l \leq N \end{array} , \right. \tag {9}
+$$
+
+$$
+\mathbf {V} _ {l} ^ {\text {v i s}} = \mathcal {E} _ {\text {v i s}} \left(\tilde {\mathbf {V}} _ {l} ^ {\text {v i s}}\right) + \tilde {\mathbf {V}} _ {l} ^ {\text {v i s}}, \quad \tilde {\mathbf {V}} _ {l} ^ {\text {v i s}} := \left[ \begin{array}{l l} \mathbf {V} _ {l} ^ {\text {s p u}}, & \mathbf {V} _ {l} ^ {\text {t m p}} \end{array} \right], \tag {10}
+$$
+
+$$
+\mathbf {V} _ {l} ^ {\mathrm {s p a}} = \mathcal {E} _ {\mathrm {s p a}} \left(\tilde {\mathbf {V}} _ {l} ^ {\mathrm {s p a}}\right) + \tilde {\mathbf {V}} _ {l} ^ {\mathrm {s p a}}, \quad \mathbf {V} _ {l} ^ {\mathrm {t m p}} = \mathcal {E} _ {\mathrm {t m p}} \left(\tilde {\mathbf {V}} _ {l} ^ {\mathrm {t m p}}\right) + \tilde {\mathbf {V}} _ {l} ^ {\mathrm {t m p}}, \tag {11}
+$$
+
+$$
+\mathbf {T} _ {l} ^ {\mathrm {c a p}} = \mathcal {E} _ {\mathrm {c a p}} \left(\tilde {\mathbf {T}} _ {l} ^ {\mathrm {c a p}}\right) + \tilde {\mathbf {T}} _ {l} ^ {\mathrm {c a p}}, \quad \mathbf {T} _ {l} ^ {\mathrm {c t x}} = \mathcal {E} _ {\mathrm {c t x}} \left(\tilde {\mathbf {T}} _ {l} ^ {\mathrm {c t x}}\right) + \tilde {\mathbf {T}} _ {l} ^ {\mathrm {c t x}}. \tag {12}
+$$
+
+When dealing with images and non-dialog data, we drop  $\mathbf{V}_l^{\mathrm{tmp}}$  and  $\mathbf{T}_l^{\mathrm{cap}}$  from the previous equations and deactivate the respective expert.
+
+# 3.3. Training
+
+# 3.3.1 Stage 1
+
+In the first stage, we only pre-train the multimodal expert layers, the vision encoder linear layer, and the spatial-temporal attention modules. Since we are the first to suggest learning the spatial and temporal features of videos and images using dedicated experts, we propose to train our model using spatial-temporal contrastive learning (STC) and spatial-temporal matching (STM). In addition, we use the established masked language modeling (MLM), vision-text $^1$  contrastive learning (VTC), and vision-text matching (VTM) similar to [19, 31, 32].
+
+<table><tr><td></td><td>Tr. Stage 1</td><td>Tr. Stage 2</td><td>Tr. Stage 3</td></tr><tr><td>Videos</td><td>Vspa, Vtmp, Tcap</td><td>Vspa, Vtmp, Tcap, Tctx</td><td>Vspa, Vtmp, Tcap, Tctx</td></tr><tr><td>Images</td><td>Vspa, Tcap</td><td>-</td><td>Vspa, Tcap, Tctx</td></tr></table>
+
+Table 1. Overview of the available features for each training stage and visual input type.
+
+Spatial-Temporal Contrastive Learning aims to better align the spatial and temporal features of video data. To this end, we use output features of the last multi-modal exert layer<sup>2</sup> and learn a cosine similarity function
+
+$$
+\mathrm {s} \left(\mathbf {V} ^ {\text {s p a}}, \mathbf {V} ^ {\text {t m p}}\right) = \Theta_ {\text {s p a}} \left(\mathbf {V} ^ {\text {s p a}}\right) ^ {\top} \Theta_ {\text {t m p}} \left(\mathbf {V} ^ {\text {t m p}}\right), \tag {13}
+$$
+
+so that aligned spatial-temporal features result in higher similarity scores, where  $\Theta_{*}$  are linear layers that map the features to a normalized lower dimensional vector space. Then, given spatial and temporal feature pairs, we compute the softmax normalized spatial-to-temporal and temporal-to-spatial similarities as
+
+$$
+p _ {i} ^ {\mathrm {s 2 t}} \left(\mathbf {V} ^ {\mathrm {s p a}}\right) = \frac {\exp \left(\tilde {\mathrm {s}} \left(\mathbf {V} ^ {\mathrm {s p a}} , \mathbf {V} _ {i} ^ {\mathrm {t m p}}\right) / \tau\right)}{\sum_ {k = 1} ^ {K} \exp \left(\tilde {\mathrm {s}} \left(\mathbf {V} ^ {\mathrm {s p a}} , \mathbf {V} _ {k} ^ {\mathrm {t m p}}\right) / \tau\right)}, \tag {14}
+$$
+
+$$
+p _ {i} ^ {\mathrm {t 2 s}} \left(\mathbf {V} ^ {\mathrm {t m p}}\right) = \frac {\exp \left(\tilde {\mathrm {s}} \left(\mathbf {V} ^ {\mathrm {t m p}} , \mathbf {V} _ {i} ^ {\mathrm {s p a}}\right) / \tau\right)}{\sum_ {k = 1} ^ {K} \exp \left(\tilde {\mathrm {s}} \left(\mathbf {V} ^ {\mathrm {t m p}}, \mathbf {V} _ {k} ^ {\mathrm {s p a}}\right) / \tau\right)}, \tag {15}
+$$
+
+where  $\tau$  is learnable temperature parameters, and  $\tilde{\mathbf{s}}$  is the maximum value of  $\mathbf{s}$  as in [32]. Finally, we can compute the loss as the cross-entropy  $\mathcal{H}$  between  $\mathbf{p}$  and  $\mathbf{y}$ :
+
+$$
+\mathcal {L} _ {\mathrm {s t c}} = \frac {1}{2} \mathbb {E} _ {\left(\mathbf {V} ^ {\mathrm {s p a}}, \mathbf {V} ^ {\mathrm {t m p}}\right)} \left[ \mathcal {H} \left(\mathbf {y} ^ {\mathrm {s 2 t}}, \mathbf {p} ^ {\mathrm {s 2 t}}\right) + \mathcal {H} \left(\mathbf {y} ^ {\mathrm {t 2 s}}, \mathbf {p} ^ {\mathrm {t 2 s}}\right) \right], \tag {16}
+$$
+
+where  $\mathbf{y}^{s2t}$  and  $\mathbf{y}^{t2s}$  are the golden one-hot similarities.
+
+Spatial-Temporal Matching complements STC and teaches the model to distinguish between positive and negative spatial-temporal feature pairs. Specifically, a matched feature pair originates from the same video, whereas an unmatched pair is constructed using negative sampling from a
+
+different video. We use a classification token as a proxy of the joint spatial-temporal representations to learn a binary classification problem using the STM loss
+
+$$
+\mathcal {L} _ {\mathrm {s t m}} = \mathbb {E} _ {\left(\mathbf {V} ^ {\mathrm {s p a}}, \mathbf {V} ^ {\mathrm {t m p}}\right)} \left[ \mathcal {H} \left(\mathbf {y} ^ {\mathrm {s t m}}, \mathbf {p} ^ {\mathrm {s t m}}\right) \right], \tag {17}
+$$
+
+where  $\mathbf{p}^{\mathrm{stm}}$  and  $\mathbf{y}^{\mathrm{stm}}$  are the predicted and the ground-truth two-class probabilities, respectively.
+
+We provide more details about the remaining established objectives (i.e., MLM, VTC, VTM) in the supplementary.
+
+# 3.3.2 Stages 2 & 3
+
+In the subsequent stages, we couple the multimodal expert layers with a pre-trained Flan-T5 $_{\text{large}}$  [21] via a linear layer. Specifically, Stage 2 aims to align the hidden states of the proposed layers with those of the pre-trained LLM. To this end, we keep the LLM weights frozen and train the whole architecture end-to-end using the generative loss (i.e., next token prediction) on large scale video dialog data<sup>3</sup>, i.e.,
+
+$$
+\mathcal {L} _ {\text {g e n}} = \mathbb {E} _ {\mathbf {X} ^ {\text {g e n}}} \left[ \mathcal {H} \left(\mathbf {y} _ {\rightarrow} ^ {\text {g e n}}, \mathbf {p} ^ {\text {g e n}}\right)\right], \tag {18}
+$$
+
+$$
+\mathbf {X} ^ {\text {g e n}} = \Theta_ {\text {g e n}} \left(\mathrm {L L M} _ {\mathrm {d e c}} \left(\left[ \mathbf {X} ^ {\text {e n c}}, \mathbf {T} ^ {\text {a n s}} \right]\right)\right), \tag {19}
+$$
+
+where  $\mathbf{X}^{\mathrm{enc}}$ ,  $\mathbf{T}^{\mathrm{ans}}$  and  $\Theta_{\mathrm{gen}}$  are the LLM encoder output, the answer token embeddings, and a linear layer that maps the features to the vocabulary space, respectively.  $\mathbf{y}_{\rightarrow}^{\mathrm{gen}}$  and  $\mathbf{p}^{\mathrm{gen}}$  denote the right-shifted ground-truth answer tokens and the predicted text token probabilities. Finally, in Stage 3, we unfreeze the LLM weights and fine-tune our model end-to-end on the downstream tasks of video and visual dialog using the same generative loss.
+
+# 4. Experiments
+
+# 4.1. Datasets
+
+As shown in Figure 3, we simultaneously use the video and image captioning datasets of WebVid-2M [10] and CC-3M [49] to pre-train the multimodal expert layers in Stage 1. Then in the second stage, we use  $25\%$  of the recent large-scale video dialog dataset Champagne [26] before performing zero-shot evaluation on the widely used video and visual dialog datasets of AVSD [5] and VisDial [22], respectively. Finally, in the third stage, we perform a domain shift evaluation based on different combinations of AVSD and VisDial to quantify whether and to what extent these seemingly similar benchmarks benefit from each other in both zero-shot and fine-tuning evaluation settings.
+
+![](images/4d998f7030b1982834f959adcdc5fd2aefe6a045e5bde942b47c173b0e2ffc1f.jpg)  
+Figure 4. Instead of training a dedicated NSP head, we propose a ranking scheme based on the cosine similarity of the candidate answers' embeddings with the respect to those of the generated ones. We used RoBERTa $_{\text{large}}$  [38] and OpenAI Text Embedding-3 $_{\text{large}}$  to generate these embeddings.
+
+# 4.2. Evaluation Metrics
+
+We use the established official metrics for each dataset to fairly benchmark  $\mathbf{V}^2\mathbf{Dial}$  with previous works. Specifically, for all three AVSD datasets, we use BLEU ( $\mathbf{B - n}$ ) [45], ROUGE-L ( $\mathbf{R}$ ) [36], METEOR ( $\mathbf{M}$ ) [11], and CIDEr ( $\mathbf{C}$ ) [53]. Whereas for VisDial, we use the retrieval metrics of recall ( $\mathbf{R}@\mathbf{k}$ ), mean reciprocal rank (MRR), and normalized discounted cumulative gain (NDCG). However, since we are jointly tackling both tasks with a fully generative model, we propose to rank the VisDial candidate answers by means of cosine similarity with respect to the generated answer using the embeddings of a pre-trained sentence transformer (i.e. RoBERTa [38] and OpenAI Text Embedding-3). We posit that this approach is more natural, caters to the current advances in generative models, and appropriately captures the semantic similarities between the generated and the candidate answers. In addition, it allows for a seamless unification of AVSD and VisDial without any training or architectural modifications. As shown in Figure 4, our proposed adaptation does not alter the computation of the sparse metrics itself and only rethinks the ranking of the candidate answers allowing for a fair comparison with previous works.
+
+# 4.3. Experimental Setup
+
+In the first stage, we trained our model for a maximum of ten epochs and applied early stopping based on a validation split to select the best checkpoint. In the subsequent stages, we trained it for up to three and twelve epochs, respectively. In all stages, we used the AdamW [39] optimizer with the default parameters and a weight decay value of 0.01. Furthermore, we applied a linear learning rate schedule with warm-up and minimum and base values of  $5e - 5$  and  $1e - 4$ , respectively. We conducted our experiments on a cluster consisting of eight A100 GPUs.
+
+# 4.4. Zero-shot Evaluation
+
+AVSD. We first assessed  $\mathbf{V}^2$  Dial in a zero-shot setting on AVSD. This is in stark contrast to previous models that were exclusively evaluated in a fine-tuning setting. We instead advocate for complementing the fine-tuning evaluation setting with a zero-shot one, as it results in a more rigorous
+
+<table><tr><td rowspan="2">Model</td><td colspan="7">AVSD-DSTC10</td><td colspan="7">AVSD-DSTC8</td><td colspan="7">AVSD-DSTC7</td></tr><tr><td>B-1</td><td>B-2</td><td>B-3</td><td>B-4</td><td>M</td><td>R</td><td>C</td><td>B-1</td><td>B-2</td><td>B-3</td><td>B-4</td><td>M</td><td>R</td><td>C</td><td>B-1</td><td>B-2</td><td>B-3</td><td>B-4</td><td>M</td><td>R</td><td>C</td></tr><tr><td>\( ^\dagger \)MoE-LLAVAarXiv&#x27;24 [35]</td><td>35.8</td><td>18.9</td><td>10.1</td><td>5.9</td><td>15.4</td><td>27.1</td><td>12.8</td><td>39.8</td><td>23.9</td><td>15.2</td><td>10.1</td><td>18.7</td><td>32.2</td><td>23.7</td><td>44.7</td><td>29.1</td><td>19.6</td><td>13.8</td><td>21.8</td><td>37.3</td><td>33.2</td></tr><tr><td>\( ^\dagger \)MiniGPT4-video\( _{CVPR^{\prime}24} \)[9]</td><td>37.9</td><td>19.9</td><td>11.3</td><td>6.8</td><td>16.2</td><td>28.7</td><td>17.7</td><td>34.8</td><td>17.6</td><td>9.7</td><td>5.8</td><td>15.8</td><td>26.3</td><td>13.3</td><td>37.8</td><td>21.2</td><td>12.7</td><td>8.2</td><td>18.4</td><td>30.2</td><td>17.7</td></tr><tr><td>\( ^\dagger \)Video-Chat\( _{AGL^{\prime}24} \)[40]</td><td>24.5</td><td>14.7</td><td>8.8</td><td>5.4</td><td>16.7</td><td>25.2</td><td>3.9</td><td>25.5</td><td>16.0</td><td>10.1</td><td>6.4</td><td>18.4</td><td>27.1</td><td>9.1</td><td>28.5</td><td>18.5</td><td>11.8</td><td>7.6</td><td>20.4</td><td>32.1</td><td>19.1</td></tr><tr><td>\( ^\dagger \)MST-MIXER\( _{ECCV^{\prime}24} \)[3]</td><td>0.1</td><td>0.0</td><td>0.0</td><td>0.0</td><td>3.1</td><td>6.8</td><td>3.0</td><td>0.2</td><td>0.1</td><td>0.1</td><td>0.0</td><td>3.3</td><td>7.1</td><td>4.3</td><td>0.2</td><td>0.1</td><td>0.0</td><td>0.0</td><td>3.4</td><td>6.9</td><td>4.6</td></tr><tr><td>\( ^\dagger \)VideoChat\( _{CVPR^{\prime}24} \)[33]</td><td>42.5</td><td>25.9</td><td>16.0</td><td>10.3</td><td>18.7</td><td>33.1</td><td>25.4</td><td>43.9</td><td>28.1</td><td>18.5</td><td>12.6</td><td>20.8</td><td>34.5</td><td>29.2</td><td>46.7</td><td>31.1</td><td>20.9</td><td>14.4</td><td>22.9</td><td>37.6</td><td>31.4</td></tr><tr><td>\( {\mathrm{V}}^{2}\text{Dial } \)</td><td>54.6</td><td>34.8</td><td>24.0</td><td>17.2</td><td>19.7</td><td>38.3</td><td>53.8</td><td>53.2</td><td>33.8</td><td>23.5</td><td>16.7</td><td>18.8</td><td>37.7</td><td>49.7</td><td>55.5</td><td>36.7</td><td>26.2</td><td>18.7</td><td>20.0</td><td>39.2</td><td>50.8</td></tr></table>
+
+Table 2. Zero-shot performance comparison on AVSD-DSTC10, AVSD-DSTC8 and AVSD-DSTC7. Best and second-best performances are in **bold** and **underlined**.  $\spadesuit$  indicates that we evaluated the model. ( $\mathbf{B} - \mathbf{n} = \mathrm{BLEU} - \mathbf{n}$ ,  $\mathbf{M} = \mathrm{METEOR}$ ,  $\mathbf{R} = \mathrm{ROUGE-L}$ ,  $\mathbf{C} = \mathrm{CIDEr}$ ).  
+
+<table><tr><td>Model</td><td>Sent. Trans.</td><td>R@1</td><td>R@5</td><td>R@10</td><td>MRR</td><td>NDCG</td></tr><tr><td>FROMAGecIMCL&#x27;23 [27]</td><td></td><td>17.6</td><td>20.1</td><td>25.1</td><td>22.0</td><td>16.5</td></tr><tr><td>ESPERCVPR&#x27;23 [62]</td><td>n/a</td><td>14.6</td><td>-</td><td>-</td><td>25.7</td><td>22.3</td></tr><tr><td>ChampagneICCV&#x27;23 [26]</td><td></td><td>-</td><td>-</td><td>-</td><td>-</td><td>25.5</td></tr><tr><td>MoE-LLAVAorXiv&#x27;24 [35]</td><td></td><td>10.6</td><td>25.4</td><td>36.4</td><td>19.6</td><td>26.7</td></tr><tr><td>MiniGPT-videoCVPR&#x27;24 [9]</td><td></td><td>7.4</td><td>17.4</td><td>26.5</td><td>14.6</td><td>23.2</td></tr><tr><td>Video-ChatGPTACL&#x27;24 [40]</td><td>RoBERTa</td><td>10.0</td><td>22.5</td><td>31.5</td><td>18.1</td><td>24.8</td></tr><tr><td>MST-MIXERECCV&#x27;24 [3]</td><td></td><td>18.2</td><td>22.1</td><td>25.7</td><td>21.9</td><td>24.6</td></tr><tr><td>VideoChat2CVPR&#x27;24 [33]</td><td></td><td>12.7</td><td>29.0</td><td>39.9</td><td>22.3</td><td>30.9</td></tr><tr><td rowspan="2">V2 Dial</td><td>RoBERTa</td><td>20.0</td><td>30.2</td><td>39.3</td><td>26.9</td><td>33.3</td></tr><tr><td>OpenAI TE-3</td><td>22.1</td><td>41.2</td><td>48.1</td><td>32.7</td><td>32.0</td></tr></table>
+
+Table 3. Zero-shot performance comparison on the VisDial v1.0 val split. OpenAI TE-3 = OpenAI Text Embedding-3large.
+
+and challenging testbed for the proposed models. To this end, we establish; to the best of our knowledge; the first zero-shot benchmark comparison on AVSD comprised of recent capable multimodal models. As can be seen from Table 2, our model outperforms all baselines by a considerable margin across 6/7 metrics of AVSD-DSTC8 and AVSD-DSTC7. On the more recent and challenging version of the benchmark (i.e. AVSD-DSTC10),  $\mathbf{V}^2$  Dial ranks first across all metrics. For instance, it more than doubles the CIDER score compared to VideoChat2 [33].
+
+VisDial. Additionally, we assessed the same model checkpoint on VisDial v1.0. As can be seen from Table 3,  $\mathbf{V}^2\mathbf{D}$ ial managed to outperform previous models such as FROMAGe [27] by a considerable margin across all metrics of the dataset. In addition, it outperformed Champagne [26] that was trained on x4 more dialog data by 7.8 absolute NDCG points. Furthermore, our model outperformed the more recent baselines of the previous section on 4/5 metrics, underlining its capability of dealing with both video and image input data types. Finally, replacing the sentence embeddings generated by RoBERTaLarge with those from OpenAI Text Emedding-3 improved the external ranking of the candidate answers and resulted in higher scores across all metrics, as can be seen in the last row of Table 3.
+
+# 4.5. Fine-tuning Evaluation
+
+AVSD. Similar to almost all previous works on AVSD, we assessed  $\mathbf{V}^2$  Dial in a fine-tuning setting on all three bench
+
+marks of the dataset. As can be seen from Table 4, our model managed to maintain it competitiveness ahead of recent models and outperformed them on the latest and most challenging AVSD-DSTC10 benchmarks across all evaluation metrics. For instance, it lifted CIDEr by over 6 absolute points compared to the second-best model. Furthermore, our model managed to maintain an on par performance with the state of the art on AVSD-DSTC8 and AVSD-DSTC7. As shown in Table 4,  $\mathbf{V}^2$  Dial increased their respective CIDEr scores by over 2 and 3 absolute points compared to the second-best model.
+
+VisDial. Finally, we fine-tuned our model and MST-MIXER [3] that had the closest AVSD performance on Vissdial v1.0 using the same fully-generative approach. As can be seen from Table 5,  $\mathbf{V}^2\mathbf{D}\mathbf{a}$ l managed to outperform all previous models on the strictest metric of the dataset by achieving a R@1 score of 44.2. However, when using OpenAI Text Embedding-3 our model managed to increase the R@1 and MRR scores to 44.9 and 52.4, respectively, thereby setting new state-of-the-art results. As expected and due to the more challenging aspect of a tackling VisDial as a fully generative task, our model performed slightly worse than the previous fine-tuned models on the remaining metrics of the dataset. However, when comparing our model with MST-MIXER that was trained using the same paradigm (i.e. the last two rows of Table 5), we can see that our model outperformed it across 4/5 metrics of the task and scored almost equally on NDCG.
+
+# 4.6. Domain Shift Evaluation
+
+Zero-shot setting. First, we fine-tuned our model's checkpoint from Stage 2 on AVSD and zero-shot evaluated it on VisDial. As can be seen from the second section of Table 6, our model's performance was lifted by a considerable margin across most metrics. Notably, the NDCG score improved by 9 absolute points compared to the results of Table 3. Then, we replicated the same experiment on AVSD after having fine-tuned the model on VisDial. Interestingly, our model's performance deteriorated across all metrics of the benchmark. This behavior could be explained by the nature of both datasets. Whereas AVSD encourages the model
+
+![](images/b640c77073a30ac7c8a082061ffd9670ce0f17a234809a4e288efff57314f03a.jpg)
+
+A man sits at a computer desk and fixes the drawer. [...] He gets up and picks up laundry and leaves.
+
+![](images/456888d629e19c35696f469bc4e80f654893cea814dc4587115b0dd9fe7ea220.jpg)
+
+![](images/efb17e3b580c2a599c7115d699b6bd2b3be768fc461bd94d9bd143bfeb93e428.jpg)  
+AVSD
+
+![](images/1e2b881e39a4ebfd7779271d9e748408f29d4892dd1c171b0d85159e59dcabbd.jpg)
+
+![](images/f40284fce39604c17160a4dc0b946904053d6b11889a3054fa3801f57ee9e8ea.jpg)  
+VisDial
+
+![](images/315a6e95474083edec48a87c8692012d0c30c3e5976708b6720446a7182bcc35.jpg)  
+Figure 5. Zero-shot qualitative examples of  $\mathbf{V}^2\mathbf{Dial}$  before and after fine-tuning on VisDial and AVSD. The former teaches the model to answer question with brief responses whereas the latter teaches it to produce longer and more elaborate answers.  
+Table 4. Fine-tuning performance comparison on AVSD-DSTC10, AVSD-DSTC8 and AVSD-DSTC7. VideoLLAMA 2 [20] was trained on AVSD amongst other datasets. Additional model comparisons can be found in the supplementary material.
+
+<table><tr><td>Model</td><td>Sent. Trans.</td><td>R@1</td><td>R@5</td><td>R@10</td><td>MRR</td><td>NDCG</td></tr><tr><td>\( LTMI_{ECCV&#x27;20} \) [44]</td><td></td><td>40.4</td><td>61.6</td><td>69.7</td><td>50.7</td><td>63.5</td></tr><tr><td>\( LTMI-LGE_{EMNLP&#x27;21} \) [17]</td><td></td><td>41.3</td><td>61.6</td><td>69.0</td><td>51.3</td><td>63.2</td></tr><tr><td>\( GoG_{ACL&#x27;21} \) [16]</td><td>n/a</td><td>41.2</td><td>61.8</td><td>69.4</td><td>51.3</td><td>62.6</td></tr><tr><td>\( UTC_{CVPR&#x27;22} \) [15]</td><td></td><td>41.3</td><td>59.8</td><td>66.3</td><td>50.6</td><td>61.0</td></tr><tr><td>Champagne \( I_{CCV&#x27;23} \) [26]</td><td></td><td>-</td><td>-</td><td>-</td><td>-</td><td>62.5</td></tr><tr><td>\( MST-MIXER_{ECCV&#x27;24} \) [3]</td><td>RoBERTa</td><td>42.2</td><td>51.6</td><td>57.8</td><td>47.7</td><td>52.5</td></tr><tr><td rowspan="2">\( V^2_Dial \)</td><td>RoBERTa</td><td>45.4</td><td>54.7</td><td>61.1</td><td>50.9</td><td>54.0</td></tr><tr><td>OpenAI TE-3</td><td>46.1</td><td>59.3</td><td>65.7</td><td>53.2</td><td>53.1</td></tr></table>
+
+Table 5. Fine-tuning performance comparison on the VisDial v1.0 val split.  $\spadesuit$  indicates that we trained and evaluated the model.
+
+to produce long and elaborate responses, VisDial teaches it to produce brief answers instead, which diminishes its performance on the language generation metrics. The qualitative examples of Figure 5 clearly illustrate this phenomenon on both datasets.
+
+Fine-tuning Setting. We first experimented with a curriculum learning strategy where we used one dataset for pretraining before finally fine-tuning on the other. As can be seen from the last section of Table 6, this training paradigm
+
+resulted in performance drops on both datasets compared to Table 4 and Table 5 where the model was only trained on the data of the respective benchmark. This indicates that the converged model's weights on one dataset do not offer a good initialization for training on the remaining one. Allowed by our model design that can jointly handle video and image input data, we finally fine-tuned one single model on both datasets simultaneously. As seen from the last row of Table 6, this resulted in the best joint performance of our model across the two datasets. Although the results on AVSD slightly dropped compared to Table 4, our model lifted its performance on VisDial by a considerable margin. This could largely be attributed to the same previous observation, as training on VisDial incentivizes our model to shorten its responses on AVSD.
+
+# 4.7. Expert Swapping
+
+In order to validate the specialization of each expert, we conducted a swapping experiment where we routed some features through inadequate experts. We first swapped experts of the same modality (i.e., experts operating on vision
+
+<table><tr><td colspan="2">Fine-tuning data</td><td colspan="4">AVSD-DSTC10</td><td colspan="4">AVSD-DSTC8</td><td colspan="4">AVSD-DSTC7</td><td colspan="4">VisDial</td><td></td></tr><tr><td>AVSD</td><td>VisDial</td><td>B-1</td><td>M</td><td>R</td><td>C</td><td>B-1</td><td>M</td><td>R</td><td>C</td><td>B-1</td><td>M</td><td>R</td><td>C</td><td>R@1</td><td>R@5</td><td>R@10</td><td>NDCG</td><td></td></tr><tr><td rowspan="2">X</td><td rowspan="2">X</td><td colspan="16">Zero-shot (from Table 2 and Table 3)</td><td></td></tr><tr><td>54.6</td><td>19.7</td><td>38.3</td><td>53.8</td><td>53.2</td><td>18.8</td><td>37.7</td><td>49.7</td><td>55.5</td><td>20.0</td><td>39.2</td><td>50.8</td><td>20.0</td><td>30.2</td><td>39.3</td><td>33.3</td><td></td></tr><tr><td rowspan="2">✓</td><td rowspan="2">X</td><td colspan="12">Fine-tuning (from Table 4)</td><td colspan="4">Zero-shot</td><td></td></tr><tr><td>70.7</td><td>26.0</td><td>55.4</td><td>103.3</td><td>76.8</td><td>30.4</td><td>62.1</td><td>135.7</td><td>78.9</td><td>31.2</td><td>62.3</td><td>139.8</td><td>12.8</td><td>36.7</td><td>50.8</td><td>42.3</td><td></td></tr><tr><td rowspan="2">X</td><td rowspan="2">✓</td><td colspan="12">Zero-shot</td><td colspan="4">Fine-tuning (from Table 5)</td><td></td></tr><tr><td>11.5</td><td>6.8</td><td>20.1</td><td>14.6</td><td>11.5</td><td>7.3</td><td>20.7</td><td>20.9</td><td>7.9</td><td>6.2</td><td>17.4</td><td>18.2</td><td>44.2</td><td>53.3</td><td>59.5</td><td>52.3</td><td></td></tr><tr><td>✓</td><td>→</td><td>✓</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>42.2</td><td>50.1</td><td>56.3</td><td>51.3</td><td></td></tr><tr><td>✓</td><td>←</td><td>✓</td><td>69.6</td><td>25.7</td><td>55.0</td><td>100.5</td><td>75.9</td><td>29.8</td><td>61.4</td><td>132.1</td><td>77.6</td><td>30.4</td><td>61.5</td><td>134.5</td><td>-</td><td>-</td><td>-</td><td></td></tr><tr><td>✓</td><td>&amp;</td><td>✓</td><td>69.3</td><td>25.4</td><td>54.8</td><td>99.9</td><td>75.1</td><td>29.3</td><td>61.1</td><td>130.0</td><td>77.3</td><td>30.0</td><td>61.7</td><td>134.5</td><td>45.4</td><td>54.7</td><td>61.1</td><td>54.0</td></tr></table>
+
+Table 6. Domain shift evaluation between the respective most prominent video and visual dialog datasets of AVSD and VisDial.  $\square \rightarrow \triangle$  means that the model was pre-trained on dataset  $\square$  before fine-tuning on dataset  $\triangle$ .  
+
+<table><tr><td rowspan="2" colspan="3">Expert
+Swapping</td><td colspan="4">AVSD-DSTC7</td><td colspan="2">VisDial</td></tr><tr><td>B-1</td><td>M</td><td>R</td><td>C</td><td>R@1</td><td>NDCG</td></tr><tr><td colspan="3">Original</td><td>78.9</td><td>31.2</td><td>62.3</td><td>139.8</td><td>44.2</td><td>52.3</td></tr><tr><td colspan="9">Swapping experts of the same modality (vision / language)</td></tr><tr><td>Espa←</td><td>Etmp</td><td>77.0</td><td>29.5</td><td>61.2</td><td>133.7</td><td>-</td><td>-</td><td></td></tr><tr><td>Ecap←</td><td>Ectx</td><td>76.1</td><td>29.6</td><td>60.3</td><td>131.1</td><td>42.8</td><td>51.9</td><td></td></tr><tr><td colspan="9">Swapping experts of different modalities</td></tr><tr><td>Espa←</td><td>Ecap</td><td>28.5</td><td>10.7</td><td>22.0</td><td>10.1</td><td>35.7</td><td>45.3</td><td></td></tr><tr><td>Etmp←</td><td>Ectx</td><td></td><td></td><td></td><td></td><td>-</td><td>-</td><td></td></tr><tr><td>Espa←</td><td>Ectx</td><td>34.4</td><td>12</td><td>25.4</td><td>11.7</td><td>32.4</td><td>42.9</td><td></td></tr><tr><td>Ecap←</td><td>Ecap</td><td></td><td></td><td></td><td></td><td>-</td><td>-</td><td></td></tr></table>
+
+or language data). As shown in Table 7, this resulted in performance drops across all metrics of both datasets, indicating that experts of the same modality are able to capture the semantic nuances of the data they specialize on. More interestingly, the performance of our model dropped more significantly when swapping experts of different modalities, as seen from the last section of Table 7. This showcases their ability to adjust to the nature of the data they process and to capture its modality specific features.
+
+# 4.8. Ablation Study
+
+Effect of Pre-training Data. We trained two versions of our model, where one was only pre-trained on Stage 1 using WebVid-2M & CC-3M and the other only on Stage 2 with a subset of Champagne. As can be seen from the middle section of Table 8, our model witnessed a comparable drop in performance compared to the full model. This underlines the equal importance of these proposed training stages to the joint down-stream performance on AVSD and VisDial. We did not conduct ablations using either WebVid-2M or CC-3M in Stage 1 as this was sufficiently explored by other recent works [19] that showed the benefit of pre-training on both image and video data.
+
+Effect of Pre-training Objectives & Model Design. Then,
+
+Table 7. Expert swapping results.  $\mathcal{E}_{\square} \leftrightarrow \mathcal{E}_{\triangle}$  means that the  $\square$  features are inadequately routed at test time through  $\mathcal{E}_{\triangle}$  and vice versa. The other experts remain unchanged.  
+
+<table><tr><td rowspan="2">Model Ablations</td><td colspan="4">AVSD-DSTC7</td><td colspan="2">VisDial</td></tr><tr><td>B-1</td><td>M</td><td>R</td><td>C</td><td>R@1</td><td>NDCG</td></tr><tr><td>Full</td><td>78.9</td><td>31.2</td><td>62.3</td><td>139.8</td><td>44.2</td><td>52.3</td></tr><tr><td>w/o Tr. Stage 1</td><td>76.9</td><td>30.0</td><td>61.4</td><td>134.0</td><td>34.5</td><td>44.6</td></tr><tr><td>w/o Tr. Stage 2</td><td>77.8</td><td>30.7</td><td>61.7</td><td>134.7</td><td>32.6</td><td>44.1</td></tr><tr><td>w/o Lstc &amp; Lstm</td><td>77.2</td><td>29.9</td><td>61.1</td><td>133.2</td><td>33.1</td><td>44.6</td></tr><tr><td>w/o separate Espa &amp; Etmp</td><td>77.0</td><td>30.1</td><td>61.1</td><td>133.8</td><td>32.9</td><td>43.8</td></tr><tr><td>w/o experts {E*}</td><td>77.5</td><td>30.0</td><td>61.4</td><td>134.8</td><td>30.6</td><td>42.2</td></tr></table>
+
+Table 8. Ablation results.
+
+we trained a version of our model without  $\mathcal{L}_{\mathrm{stc}}$  and  $\mathcal{L}_{\mathrm{stm}}$  in Stage 1 using the same schedule and training data as our full model. As shown in the fourth row of Table 8, this ablated version suffered a performance drop not only in AVSD but also in VisDial. This indicates that these losses improve not only the temporal capabilities of our model but also its spatial ones. Then, we trained a version that sequentially applies spatial and temporal attention, as in [13, 19]. Since this version does not have separate spatial-temporal experts, we also omitted the previous two objectives. As seen in the penultimate row of Table 8, this version underperformed our full model on both datasets, showcasing the effectiveness of our approach. Finally, we trained a version without all the expert layers. As shown in the last row, its performance dropped compared to our full model and performed the worst on VisDial.
+
+# 5. Conclusion
+
+We presented  $\mathbf{V}^2\mathbf{Dial}$  - a model that can jointly tackle video and visual conversational tasks using a multimodal expert-based approach that; for the first time, disentangles the learning of the spatial and temporal features of images and videos using two separate experts. Extensive evaluation on the respective widely used video and visual dialog datasets of AVSD and VisDial show that our model achieves new state-of-the-art zero-shot and fine-tuning performance. Finally, we conducted the first domain shift evaluation of AVSD and VisDial and provided insights on how to optimally leverage their respective training data.
+
+# Acknowledgment
+
+This work was partially funded by a LOEWE-Start-Professur (LOEWE/4b//519/05.01.002-(0006)/94), LOEWE-Spitzen-Professur (LOEWE/4a//519/05.00.002-(0010)/93), and an Alexander von Humboldt Professorship in Multimodal Reliable AI, sponsored by Germany's Federal Ministry for Education and Research.
+
+# References
+
+[1] Adnen Abdessaied, Mihai Bace, and Andreas Bulling. Neuro-Symbolic Visual Dialog. In Proceedings of the International Conference on Computational Linguistics (COLING), 2022. 1  
+[2] Adnen Abdessaied, Manuel Hochmeister, and Andreas Bulling. OLViT: Multi-Modal State Tracking via Attention-Based Embeddings for Video-Grounded Dialog. In Proceedings of the Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING), 2024. 2  
+[3] Adnen Abdessaied, Lei Shi, and Andreas Bulling. MultiModal Video Dialog State Tracking in the Wild. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 6, 7  
+[4] Adnen Abdessaied, Lei Shi, and Andreas Bulling. VD-GR: Boosting Visual Dialog With Cascaded Spatial-Temporal Multi-Modal Graphs. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024. 1, 2  
+[5] Huda Alamri, Vincent Cartillier, Abhishek Das, Jue Wang, Anoop Cherian, Irfan Essa, Dhruv Batra, Tim K Marks, Chiori Hori, Peter Anderson, et al. Audio visual scene-aware dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 5  
+[6] Huda Alamri, Anthony Bilic, Michael Hu, Apoorva Beedu, and Irfan Essa. End-to-end multimodal representation learning for video dialog. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2  
+[7] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1  
+[8] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: Visual Question Answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2015. 1, 2  
+[9] Kirolos Ataallah, Xiaogian Shen, Eslam Abdelrahman, Essam Sleiman, Deyao Zhu, Jian Ding, and Mohamed Elho
+
+seiny. MiniGPT4-video: Advancing multimodal llms for video understanding with interleaved visual-textual tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024. 6  
+[10] Max Bain, Arsha Nagrani, Güll Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 5  
+[11] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, 2005. 5  
+[12] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. VLMo: Unified vision-language pre-training with mixture-of-modality-experts. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2  
+[13] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is Space-Time Attention All You Need for Video Understanding? In Proceedings of the International Conference on Machine Learning (ICML), 2021. 3, 8  
+[14] Soravit Changpinyo, Piyush Sharma, and Nan Ding and Radu Soricut. Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1  
+[15] Cheng Chen, Yudong Zhu, Zhenshan Tan, Qingrong Cheng, Xin Jiang, Qun Liu, and Xiaodong Gu. UTC: A Unified Transformer with Inter-Task Contrastive Learning for Visual Dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 7  
+[16] Feilong Chen, Xiuyi Chen, Fandong Meng, Peng Li, and Jie Zhou. GoG: Relation-aware graph-over-graph network for visual dialog. In Findings of the Association for Computational Linguistics: ACL, 2021. 1, 2, 7  
+[17] Feilong Chen, Xiuyi Chen, Can Xu, and Daxin Jiang. Learning to ground visual objects for visual dialog. In Findings of the Association for Computational Linguistics: EMNLP, 2021. 7  
+[18] Zhe Chen, Hongcheng Liu, and Yu Wang. DialogMCF: Multimodal Context Flow for Audio Visual Scene-Aware Dialog. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2023. 2, 7  
+[19] Feng Cheng, Xizi Wang, Jie Lei, David Crandall, Mohit Bansal, and Gedas Bertasius. VindLU: A Recipe for Effective Video-and-Language Pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4, 8  
+[20] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs. arXiv preprint arXiv:2406.07476, 2024. 2, 7
+
+[21] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research (JMLR), 2024. 5  
+[22] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José M.F. Moura, Devi Parikh, and Dhruv Batra. Visual Dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2, 5  
+[23] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2019. 1  
+[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2020. 1  
+[25] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan First, et al. GLaM: Efficient scaling of language models with mixture-of-experts. In Proceedings of the International Conference on Machine Learning (ICML), 2022. 2  
+[26] Seungju Han, Jack Hessel, Nouha Dziri, Yejin Choi, and Youngjae Yu. Champagne: Learning real-world conversation from large-scale web videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 5, 6, 7  
+[27] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. In Proceedings of the International Conference on Machine Learning (ICML), 2023. 6  
+[28] Hung Le and Steven C.H. Hoi. Video-Grounded Dialogues with Pretrained Generation Language Models. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2020. 2  
+[29] Hung Le, Nancy F. Chen, and Steven Hoi. Learning reasoning paths over semantic graphs for video-grounded dialogues. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2, 7  
+[30] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2020. 2  
+[31] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 4  
+[32] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with
+
+frozen image encoders and large language models. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 1, 4  
+[33] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. MVBench: A Comprehensive Multi-modal Video Understanding Benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 6  
+[34] Zekang Li, Zongjia Li, Jinchao Zhang, Yang Feng, and Jie Zhou. Bridging text and video: A universal multimodal transformer for audio-visual scene-aware dialog. Transactions on Audio, Speech, and Language Processing, 2021. 2  
+[35] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Jina Huang, Junwu Zhang, Yatian Pang, Munan Ning, and Li Yuan. MoE-LLaVA: Mixture of Experts for Large Vision-Language Models. arXiv preprint arXiv:2401.15947, 2024. 2, 6  
+[36] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, 2004. 5  
+[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1  
+[38] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. arXiv preprint arXiv 1907.11692, 2019.5  
+[39] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. In Proceedings of the International Conference on Learning Representations (ICLR), 2019. 5  
+[40] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-ChatGPT: Towards Detailed Video Understanding via Large Vision and Language Models. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2024. 6  
+[41] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1  
+[42] Vishvak Murahari, Dhruv Batra, Devi Parikh, and Abhishek Das. Large-scale pretraining for visual dialog: A simple state-of-the-art baseline. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2  
+[43] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal Contrastive Learning with LIMoE: the Language-Image Mixture of Experts. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2  
+[44] Van Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Efficient Attention Mechanism for Visual Dialog that Can Handle All the Interactions Between Multiple Inputs. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 7  
+[45] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine
+
+translation. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2002. 5  
+[46] Hoang-Anh Pham, Thao Minh Le, Vuong Le, Tu Minh Phuong, and Truyen Tran. Video Dialog as Conversation about Objects Living in Space-Time. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2  
+[47] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 2  
+[48] Ankit P. Shah, Shijie Geng, Peng Gao, Anoop Cherian, Takaaki Hori, Tim K. Marks, Jonathan Le Roux, and Chiori Hori. Audio-Visual Scene-Aware Dialog and Reasoning using Audio-Visual Transformers with Joint Student-Teacher Learning. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022. 1  
+[49] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2018. 5  
+[50] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1  
+[51] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. EVA-CLIP: Improved Training Techniques for CLIP at Scale. arXiv preprint arXiv:2303.15389, 2023. 3  
+[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 1  
+[53] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based Image Description Evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 5  
+[54] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2  
+[55] Yue Wang, Shafiq R. Joty, Michael R. Lyu, Irwin King, Caiming Xiong, and Steven C. H. Hoi. VD-BERT: A Unified Vision and Dialog Transformer with BERT. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 2  
+[56] Zihao Wang, Junli Wang, and Changjun Jiang. Unified multimodal model with unlikelihood training for visual dialog. In Proceedings of the ACM International Conference on Multimedia (ACM MM), 2022. 1, 2  
+[57] Ronald J Williams and David Zipser. A learning algorithm for continually running fully recurrent neural networks. Neural Comput., 1989. 2
+
+[58] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the ACM International Conference on Multimedia (ACM MM), 2017. 1, 2  
+[59] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1  
+[60] Sunjae Yoon, Eunseop Yoon, Hee Suk Yoon, Junyeong Kim, and Chang Yoo. Information-theoretic text hallucination reduction for video-grounded dialogue. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2022. 2, 7  
+[61] Xintong Yu, Hongming Zhang, Ruixin Hong, Yangqiu Song, and Changshui Zhang. VD-PCR: Improving visual dialog with pronoun coreference resolution. Pattern Recognition, 2022. 2  
+[62] Youngjae Yu, Jiwan Chung, Heeseung Yun, Jack Hessel, Jae Sung Park, Ximing Lu, Rowan Zellers, Prithviraj Ammanabrolu, Ronan Le Bras, Gunhee Kim, et al. Fusing pretrained language models with multimodal prompts through reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6  
+[63] Zexuan Zhong, Mengzhou Xia, Danqi Chen, and Mike Lewis. Lory: Fully Differentiable Mixture-of-Experts for Autoregressive Language Model Pre-training. In Proceedings of the Conference on Language Modeling (COLM), 2024. 2  
+[64] Tong Zhu, Xiaoye Qu, Daize Dong, Jiacheng Ruan, Jingqi Tong, Conghui He, and Yu Cheng. Llama-moe: Building mixture-of-experts from llama with continual pre-training. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2024. 2

2025/V^2Dial_ Unification of Video and Visual Dialog via Multimodal Experts/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e7a0a0889400ae8440a8728e69e7a71aa46a65cde28fb56ddf069b16ecf9697
+size 780013

2025/Variance-Based Membership Inference Attacks Against Large-Scale Image Captioning Models/26bae7f8-8e37-4862-840e-6578a63d1d11_content_list.json

@@ -0,0 +1,1435 @@
+[
+    {
+        "type": "text",
+        "text": "Variance-Based Membership Inference Attacks Against Large-Scale Image Captioning Models",
+        "text_level": 1,
+        "bbox": [
+            122,
+            128,
+            875,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Daniel Samira Edan Habler Yuval Elovici Asaf Shabtai Ben-Gurion University of the Negev, Israel {samirada, habler}@post.bgu.ac.il, {elovici, shabaia}@bgu.ac.il",
+        "bbox": [
+            215,
+            203,
+            776,
+            256
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            291,
+            326,
+            306
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The proliferation of multi-modal generative models has introduced new privacy and security challenges, especially due to the risks of memorization and unintentional disclosure of sensitive information. This paper focuses on the vulnerability of multi-modal image captioning models to membership inference attacks (MIAs). These models, which synthesize textual descriptions from visual content, could inadvertently reveal personal or proprietary data embedded in their training datasets. We explore the feasibility of MIAs in the context of such models. Specifically, our approach leverages a variance-based strategy tailored for image captioning models, utilizing only image data without knowing the corresponding caption. We introduce the means-of-variance threshold attack (MVTA) and confidence-based weakly supervised attack (C-WSA) based on the metric, means-of-variance (MV), to assess variability among vector embeddings. Our experiments demonstrate that these models are susceptible to MIAs, indicating substantial privacy risks. The effectiveness of our methods is validated through rigorous evaluations on these real-world models, confirming the practical implications of our findings.",
+        "bbox": [
+            89,
+            324,
+            485,
+            643
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            91,
+            675,
+            220,
+            690
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The advancement from traditional machine learning (ML) to advanced generative models has progressed significantly in recent years [7]. An example of this development is seen in the field of multi-modal models, which are able to process information from various modalities [4]. These models have become increasingly sophisticated and are employed in challenging tasks involving various types of data (e.g., text, images, video). In this paper we focus on image captioning models [2, 3, 23, 24, 26, 34, 50], a multi-modal model used to generate textual descriptions for images.",
+        "bbox": [
+            89,
+            702,
+            482,
+            853
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "While multi-modal models have fostered the development of a variety of new multi-modal applications, they also present new privacy and security challenges [6, 28]. A crit",
+        "bbox": [
+            89,
+            854,
+            483,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ical concern is the potential of generative models, particularly those trained on large private datasets, to inadvertently memorize and expose sensitive information. Such risks can lead to issues related to intellectual property misuse and privacy breaches involving data from proprietary images or documents [5, 45]. Among the various threats are membership inference attacks (MIAs), which aim to determine whether specific data records were included in a model's training dataset [20]. The ability to infer membership can result in the exposure of personal health information, financial details, or undisclosed affiliations, thereby posing a threat to individual privacy and security.",
+        "bbox": [
+            511,
+            292,
+            906,
+            474
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Initially, studies on MIAs focused on classification models, revealing heightened privacy risks primarily due to model overfitting [9, 20, 47]. Subsequently, attention turned to generative models, where research highlighted their tendency to memorize training data, raising concerns about digital forgery and the loss of originality [10, 11, 39]. In MIAs, as defined by Shokri et al. [37], shadow models are used to mimic the target model's behavior. However, crafting effective shadow models requires extensive datasets that mirror the original training data, along with substantial computational resources. Additionally, in many MIAs [14, 19, 21, 30, 33, 33, 43, 44, 49], the shadow models are scaled to hundreds of samples, and must share similar architectures with the target model and use identical training algorithms. Their dependence on shadow model training limits their applicability for large-scale models [21, 43].",
+        "bbox": [
+            511,
+            474,
+            908,
+            717
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To the best of our knowledge, the study performed by Hu et al. [21] is the only research exploring MIAs on image captioning models. The authors introduced two attack methods, which shadow models are employed. Furthermore, they did not directly evaluate the methods on operational image captioning models, such as those from Hugging Face. Instead, they were assessed in a simulated environment, using artificial models trained on synthetic configurations, small datasets, and numerous training epochs.",
+        "bbox": [
+            511,
+            718,
+            908,
+            853
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent studies have proposed MIAs that do not rely on shadow models [22, 25]. Specifically, Ko et al. [22] introduced the cosine similarity (CS) attack (CSA) and",
+        "bbox": [
+            511,
+            854,
+            908,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "9210",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "the weakly supervised attack (WSA). These attacks target large-scale multi-modal contrastive language-image pretraining (CLIP) models [35] that utilize text-image data pairs. Both methods use a CS metric to measure the similarity between the embedding vectors of images and text from the target CLIP model. The second method, WSA, combines ground-truth data about non-members with a pseudomember set, which consists of samples flagged as likely members based on their CS values. A classifier is trained on the dataset to distinguish members from non-members.",
+        "bbox": [
+            89,
+            90,
+            480,
+            242
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we introduce two MIAs targeting image captioning models: the means-of-variance (MV) threshold attack (MVTA) and the confidence-based weakly supervised attack (C-WSA). In contrast to the techniques of Ko et al. [22], we tailor our methods specifically for image captioning models and enhance them with a new metric, the MV metric, which measures the variance between vectors. Our methods use only images as input, operating under the realistic assumption that an adversary lacks access to ground-truth reference text (i.e., captions) and instead uses captions generated by the target image captioning model. Additionally, we integrate a confidence threshold training method that leverages the pseudo-dataset's confidence to improve the classifier's performance.",
+        "bbox": [
+            89,
+            251,
+            482,
+            463
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We evaluate the proposed methods under varying levels of attacker knowledge (zero and partial) in a black-box scenario. Additionally, we conduct the evaluation using public models trained on large-scale datasets to simulate a realistic setting. The evaluation results show that our proposed methods identify membership information in multimodal image captioning models, using BLIP [23], ViT-GPT2 [13], and GIT [42], trained on large-scale datasets. Our methods enhance detection capabilities, achieving approximately  $20\\%$  to  $30\\%$  improvements over current methods. To summarize, our main contributions are as follows:",
+        "bbox": [
+            89,
+            470,
+            482,
+            638
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- New MIAs targeting image captioning models: We introduce the MVTA and C-WSA for the image captioning domain, two new attacks in which the input consists solely of images.",
+            "- The means-of-variance metric: We introduce a quantitative metric designed to assess the variance between vectors in a multi-dimensional space. While we used it for texts, this metric is general and can be used in various fields that utilize embedding vectors.",
+            "- Innovative training of the pseudo-member set: We introduce a new step in the process of training the attack model, a confidence-based training step that utilizes a pseudo-member set.",
+            "- Real-world application: By evaluating our methods on Hugging Face models, we demonstrate the practicality and effectiveness of our methods in real-world scenarios."
+        ],
+        "bbox": [
+            89,
+            657,
+            482,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            513,
+            89,
+            653,
+            104
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Previous studies on MIAs exploited the model's tendency to produce more confident predictions for data included in the training set compared to unseen data [9, 20, 37, 40, 47], which is also referred to as memorization, since they tend to memorize the training data. MIAs are more challenging in the context of generative models, where confidence scores or prediction labels are typically unavailable [10, 11]. In addition, it was shown that the time it takes to train a generative model and the size of the dataset used to train the model influence the level of memorization, where extended training epochs and smaller datasets typically lead to increased memorization, enhancing the success of attacks [39].",
+        "bbox": [
+            511,
+            114,
+            903,
+            296
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recent studies have demonstrated diffusion models' vulnerability to MIAs [12, 46]. In white-box settings, techniques involving statistical and neural network-based inference have shown high effectiveness, especially in models trained with small datasets and a large number of training epochs [14, 19, 30]. In black-box scenarios, methods leveraging semantic fidelity and novel classifier-based approaches have achieved high accuracy, though they often depend on access to non-training data and known training sets, which can limit their broader applicability [33, 44, 49].",
+        "bbox": [
+            511,
+            296,
+            903,
+            446
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Attacks on multi-modal models. The CLIP model [35], which is one of the many types of multi-model models, utilizes contrastive learning to align text captions with images, leveraging unsupervised learning to interpret and generate image content based on natural language descriptions. Traditionally, MIAs targeting models like CLIP have relied on computationally intensive shadow models [14, 19, 21, 30, 33, 33, 43, 44, 49]. However, recent research has proposed efficient, shadow-free approaches for MIAs against multimodal models. The textual unimodal detector [25] identifies the use of personally identifiable information (PII) in CLIP models using only textual data, reducing both computational costs and privacy risks without the need for images. Another method, the WSA [22], addresses practical MIAs against large-scale multi-modal CLIP models by combining ground-truth information about non-members with noisy member data from a weak attack to develop an attack model. The WSA starts by querying the target model for image and text features of each non-member and evaluates the CS between these features. The CS distribution, typically Gaussian, helps approximate the mean and variance and is used to flag samples with a CS significantly above the mean as likely members. This identified set, known as the pseudo-member set, is then used to construct an attack dataset for training an attack model that determines the membership status of any target sample based on its features. Notably, the attack model does not account for the possibility that labels marked as true during training might be mislabeled, which could potentially improve performance if some of these were excluded from training.",
+        "bbox": [
+            511,
+            448,
+            903,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "9211",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/dd7592c72a36cb89a4ead35101e56b5ae3fb2d00ab947bb7e8a5482afe54946e.jpg",
+        "table_caption": [
+            "Table 1. An overview of membership inference research on generation models. \"Reference Output\" indicates the attacker has access to training text references."
+        ],
+        "table_footnote": [
+            "\"✓\" - the knowledge is necessary for the attack, \"-\" the knowledge is unnecessary for the attack"
+        ],
+        "table_body": "<table><tr><td>Target Model Tasks</td><td>Attacks</td><td>Using Shadow Model</td><td>Final Model Prediction</td><td>Evaluated on Realistic Models</td><td>Other Information</td></tr><tr><td>Diffusion Models</td><td>[14, 19, 30, 33, 44, 49]</td><td>✓ or -</td><td>✓</td><td>✓ or -</td><td>Reference Output or -</td></tr><tr><td>LLMs</td><td>[8, 15, 16, 29, 31]</td><td>✓ or -</td><td>✓ or -</td><td>✓</td><td>-</td></tr><tr><td>Multi-Modal Models - CLIP</td><td>[22, 25]</td><td>-</td><td>✓</td><td>✓</td><td>-</td></tr><tr><td>Multi-Modal Models - Image Captioning</td><td>M4I [21]</td><td>✓</td><td>✓</td><td>-</td><td>Reference Output or -</td></tr><tr><td rowspan=\"2\">Ours</td><td>Means-of-Variance Threshold Attack</td><td>-</td><td>✓</td><td>✓</td><td>-</td></tr><tr><td>Confidence-Based Weakly Supervised Attack</td><td>-</td><td>✓</td><td>✓</td><td>-</td></tr></table>",
+        "bbox": [
+            91,
+            128,
+            903,
+            228
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Another type of multi-modal model, BLIP [23, 24], advances vision-language tasks by adapting to both understanding and generation challenges, including image-text retrieval, image captioning, and visual question answering. To the best of our knowledge, the work of Hu et al. [21] is the only one that explores MIAs on image captioning models. Their paper introduces two attack methods: metric-based and feature-based. The metric-based approach uses similarity metrics to infer membership status by comparing the similarity of model outputs with ground-truth data. The feature-based approach utilizes a pre-trained multi-modal feature extractor to analyze similarities between the input and output features. These attacks were not applied directly to operational image captioning models but rather in a controlled setting with synthetic configurations, in terms of small dataset size and a large number of training epochs, using an encoder-decoder framework to simulate the process. In addition, the authors also employed shadow models, which increases the complexity and computational costs.",
+        "bbox": [
+            88,
+            267,
+            485,
+            554
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Unlike previous works, we focuses on the realm of multimodal image captioning models, an area that has received little research attention in the context of MIAs. In this research, we avoid the conventional reliance on computationally intensive shadow models, which necessitate datasets that closely mirror the training data of the target model. Instead, we base our attack methodologies on metrics designed to detect discrepancies in data variance. By applying MIAs in scenarios that only involve image data without access to corresponding textual information, our proposed MIAs have a broader range of applications.",
+        "bbox": [
+            88,
+            555,
+            483,
+            720
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "An overview of membership inference research across various generation model tasks is provided in Table 1.",
+        "bbox": [
+            89,
+            722,
+            483,
+            753
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Threat Model",
+        "text_level": 1,
+        "bbox": [
+            89,
+            768,
+            232,
+            784
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The primary goal of the attacker is to determine whether a given sample is a member of the target image captioning model's training set or not. The attacker can query the model, submitting images and receiving the corresponding generated text captions in response, but does not have any access to the model itself (i.e., query-only). The attacker does not have any knowledge of the target model's archi",
+        "bbox": [
+            89,
+            794,
+            485,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "tectural, parameters, or the specific training algorithm used.",
+        "bbox": [
+            511,
+            267,
+            903,
+            282
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We emphasize that the attacker's dataset consists solely of images, allowing the attack to remain effective even in scenarios where access to the model's training captions or other textual references is limited or unavailable. We explore two different settings concerning the attacker's knowledge of the data used to train the target model:",
+        "bbox": [
+            511,
+            282,
+            903,
+            375
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Zero-knowledge setting: The attacker has no prior knowledge of the distribution of the data used to train the target model, making this approach challenging but realistic in restricted data access scenarios.",
+            "- Partial-knowledge setting: Following prior research [9, 21, 22, 33, 43], the attacker has access to a publicly available dataset of images,  $D_{all}$ , approximating the data distribution used to train the target model. The attacker also has access to a non-member dataset  $D_{no}$ , while the membership status of  $D_{all}$  is unknown. The non-member dataset can be collected from content published after the model's release or synthesized by image-generating AI models, such as publicly available diffusion models."
+        ],
+        "bbox": [
+            511,
+            377,
+            908,
+            574
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4. Methodology",
+        "text_level": 1,
+        "bbox": [
+            511,
+            588,
+            648,
+            606
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our methodology is designed to leverage our intuition regarding the use of variance in the outputs of a target image captioning model as a means of inferring infer membership information. The pipeline of our methodology consists of three main stages: text generation and multi-modal feature encoding, variance metric computation, and application of proposed attacks. First, the target model generates text captions from images. These captions and images are then encoded into feature embeddings, using pre-trained text and image encoders, respectively. The outputs are then analyzed to determine membership, aiming to ascertain whether a new image was part of the target model's training dataset. Next, we compute a variance-based metric, the MV metric, to quantify the variance among text embeddings generated for each image. This metric serves as the basis for identifying membership by capturing the relationship between variance and membership. Finally, we propose two attacks based on the MV metric: the MVTA and C-WSA. The MVTA operates in a zero-knowledge setting by exploit",
+        "bbox": [
+            511,
+            613,
+            908,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "9212",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ing the reduced variance of text embeddings for training set images. The C-WSA, designed for a partial-knowledge setting, combines non-member ground truth with probabilistic pseudo-member labels and employs a confidence threshold to train an attack classifier model. The stages of our methodology are described below.",
+        "bbox": [
+            89,
+            90,
+            482,
+            181
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1. Text Generation and Feature Encoding",
+        "text_level": 1,
+        "bbox": [
+            89,
+            188,
+            426,
+            205
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "This stage involves two steps: text generation and multimodal feature encoding, as shown in Figure 1. Typically, image captioning models use two text generation decoding strategies to generate concise descriptions: beam search [32] and top-  $p$  sampling [18]. Beam search is deterministic, consistently generating the same outputs for a given input by keeping the best candidates when choosing the next word. In contrast, top-  $p$  sampling, also known as nucleus sampling, is non-deterministic and introduces diversity by selecting from a probabilistic \"nucleus\" of options. Carlini et al. [8] showed that when trained on paired data of the prompt and image, an image generation model generates similar images from the same prompt for two different random seeds. Following this principle, we denote the caption generated for an image  $x$  using top-  $p$  sampling by the model Gen as  $\\mathrm{Gen}(x)$ . Then, if multiple captions generated for the same image  $x$  are highly similar under a distance measure  $d$  ( $\\mathrm{Gen}(x) \\approx_d \\mathrm{Gen}(x)$ ), it suggests that the model has memorized the image. Based on this, we assume that the variance among the  $n$  captions generated for an image will be lower if the target model has been trained on that image.",
+        "bbox": [
+            91,
+            212,
+            483,
+            542
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "During the feature encoding step, features are derived using a pretrained text encoder  $f_{text}$  to obtain text embeddings from the generated captions, and a pre-trained image encoder  $f_{img}$  to extract visual features from images. We note that the widespread availability of large-scale pre-trained models allows adversaries to easily obtain both encoders.",
+        "bbox": [
+            89,
+            544,
+            483,
+            635
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/11bc1b35e06c2bc2051cd015627ef0c9aa01a00665cd16de8d8ce9b00c79b17c.jpg",
+        "image_caption": [
+            "Figure 1. Overview of our method's text generation and multimodal feature encoding stage. This stage consists of two steps: text generation, in which captions are generated using the target model, and feature encoding, in which features from multiple modalities (text and image) are encoded using pre-trained models."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            102,
+            648,
+            473,
+            801
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2. Means-of-Variance Metric",
+        "text_level": 1,
+        "bbox": [
+            513,
+            90,
+            754,
+            104
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Based on our intuition, we introduce the new MV metric, which is designed to quantify the variance between embedding vectors. This metric constructs a variance vector of size  $m$ , where each index represents the variance of a specific feature calculated across  $n$  vector embeddings. Once the variance vector is constructed, the metric then computes the mean of the values within it. The resulting single value represents the overall variance among the vectors, providing a concise measure of dispersion. Figure 2 provides an illustration of the MV metric calculation process, demonstrating how the MV metric is computed.",
+        "bbox": [
+            511,
+            113,
+            906,
+            280
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/00ba8d3249090bf9edb1b2235a30344fcb60cfdc3b020da82095f1134bd91960.jpg",
+        "image_caption": [
+            "Figure 2. MV metric calculation. The feature matrix  $(f)$  consists of  $n$  vectors with  $m$  features. The variance of each feature across these vectors is computed to form the variance vector  $(v)$ , whose mean gives the MV score."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            542,
+            294,
+            880,
+            388
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.3. Proposed Attacks",
+        "text_level": 1,
+        "bbox": [
+            511,
+            479,
+            684,
+            494
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Means-of-variance threshold attack. In a zero-knowledge setting, we compute the MV metric for the set of the embedding texts generated from each image, quantifying the average variance among these captions. If the calculated  $MV$  metric for image  $x$  falls below a predetermined threshold  $\\tau$ ,  $x$  is classified as a member of the training set:",
+        "bbox": [
+            511,
+            503,
+            905,
+            594
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\text {P r e d i c t e d M e m b e r s h i p} = \\left\\{ \\begin{array}{l l} \\text {M e m b e r} & \\text {i f} M V <   \\tau \\\\ \\text {N o n - M e m b e r} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            531,
+            601,
+            905,
+            638
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The threshold  $\\tau$  is empirically determined by analyzing the distribution of MV scores for both members and non-members. In practice, MVTA serves as a baseline for C-WSA, with  $\\tau$  calculated using the mean and standard deviation of MV scores from non-member datasets  $(D_{no})$ .",
+        "bbox": [
+            511,
+            643,
+            906,
+            719
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "This approach adapts confidence or loss-based thresholds for data membership to multi-modal image captioning.",
+        "bbox": [
+            511,
+            719,
+            905,
+            750
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Confidence-based weakly supervised attack. Building on the WSA method [22], our second attack, the C-WSA, introduces an innovative adaptation in the partial-knowledge setting by utilizing the MV metric to enhance the model's ability to distinguish between member and non-member data. An overview of the attack is provided in Figure 3. Similar to a typical WSA, our attack integrates ground-truth information about non-members, along with probabilistic assessments of membership.",
+        "bbox": [
+            511,
+            763,
+            908,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "9213",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/37ae34c0425bbf0ebadb99588c10e2d8a4a8bbb6edc8803e44bab35fca2e006e.jpg",
+        "image_caption": [
+            "Figure 3. Overview of the C-WSA."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            106,
+            92,
+            472,
+            239
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In our modified attack pipeline, we begin by querying the target model with images from the known non-member dataset  $(D_{no})$  to generate  $n$  captions per image. Given the availability of public or synthetic datasets, an adversary can readily access or create them. Then, we calculate the MV metric for the dataset in order to establish a baseline distribution of variance scores among non-members, which we hypothesized to follow a Gaussian distribution. The MV scores for non-members are denoted as follows:",
+        "bbox": [
+            89,
+            292,
+            483,
+            429
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nM V _ {n o} = \\left\\{M V \\left(f _ {t e x t} (G e n (x))\\right) \\mid x \\in D _ {n o} \\right\\} \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            137,
+            439,
+            482,
+            455
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $f_{text}(Gen(x))$  refers to the text embeddings extracted by the text encoder  $f_{text}$  from the captions generated by the target model for each image  $x$ , and  $MV$  is the function that performs the metric on the vectors.",
+        "bbox": [
+            89,
+            465,
+            483,
+            525
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Using  $MV_{no}$ , we calculate the mean  $\\mu_{no}$  and standard deviation  $\\sigma_{no}$  of the MV scores:",
+        "bbox": [
+            89,
+            526,
+            483,
+            555
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mu_ {n o} = \\operatorname {m e a n} \\left(M V _ {n o}\\right), \\quad \\sigma_ {n o} = \\operatorname {s t d} \\left(M V _ {n o}\\right) \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            138,
+            566,
+            480,
+            583
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "These parameters  $(\\mu_{no}$  and  $\\sigma_{no})$  are crucial for identifying potential members by comparing their MV scores to the non-member distribution baseline mentioned above. Intuitively, if a sample's  $MV$  is substantially lower than  $\\mu_{no}$ , specifically by more than  $\\lambda \\sigma_{no}$  for a given constant  $\\lambda$ , it suggests that the sample is likely a member.",
+        "bbox": [
+            89,
+            593,
+            483,
+            683
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For the dataset  $D_{all}$ , which comprises potential members, we compute the MV scores using the text embeddings generated by the target model. Samples in  $D_{all}$  whose MV scores significantly deviate from this non-member baseline are identified as potential members. Specifically, if a sample's MV score exceeds a defined threshold calculated as  $\\mu_{no} - \\lambda \\sigma_{no}$ , it is classified as a pseudo-member. The corresponding subset of these identified samples ( $D_{p-mem}$ ), serves as our pseudo-member set. This set comprises a mix of likely members and potential false positives.",
+        "bbox": [
+            89,
+            684,
+            483,
+            834
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The formulas associated with this process are as follows:",
+        "bbox": [
+            109,
+            834,
+            482,
+            849
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\tau = \\mu_ {n o} - \\lambda \\sigma_ {n o} \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            228,
+            862,
+            482,
+            876
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nD _ {p - m e m} = \\{x \\mid M V (f _ {\\text {t e x t}} (\\operatorname {G e n} (x))) <   \\tau , x \\in D _ {\\text {a l l}} \\} \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            109,
+            885,
+            482,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "With the  $D_{p - mem}$  and  $D_{no}$  in place, we proceed to create the attack training set  $D_{attack}$ , comprising image features from  $D_{no} \\cup D_{p - mem}$  with their associated binary labels:",
+        "bbox": [
+            511,
+            90,
+            903,
+            137
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nD _ {\\text {a t t a c k}} = \\left\\{f _ {i m g} (x) \\mid x \\in D _ {n o} \\cup D _ {p - m e m}, b = 1 [ x \\in D _ {p - m e m} ] \\right\\} \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            535,
+            141,
+            903,
+            157
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $f_{img}(x)$  represents the image features extracted by the image encoder  $f_{img}$  for each image  $x$ , and  $b$  is a binary indicator with 1 representing a potential member and 0 representing a non-member. Finally, using  $D_{all}$ , we train the attack classifier model  $f_{attack}$  so that it is capable of determining the membership status of any target sample by evaluating the sample's image features.",
+        "bbox": [
+            511,
+            162,
+            906,
+            267
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In  $D_{all}$ , samples labeled with 0 are definitively non-members, whereas those labeled with 1 may not actually be members. This distinction guides our training strategy for the attack classifier model  $f_{attack}$ , which incorporates a confidence threshold. Initially, the attack model undergoes a \"warm-up\" phase, training without interruptions to establish baseline prediction tendencies. This phase lasts for a predefined ratio of epochs, referred to as \"warm-epoch.\" Then, before each training batch, we evaluate the membership confidence of the samples. For those identified as pseudo-members ( $b = 1$ ), training continues only if their membership confidence exceeds a predetermined threshold. This approach refines the training process, focusing primarily on samples most likely to be members, thus enhancing the accuracy and effectiveness of our attack classifier.",
+        "bbox": [
+            511,
+            268,
+            906,
+            494
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5. Evaluation",
+        "text_level": 1,
+        "bbox": [
+            513,
+            505,
+            630,
+            520
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1. Evaluation Setup",
+        "text_level": 1,
+        "bbox": [
+            513,
+            530,
+            681,
+            546
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Datasets. To evaluate our attacks, we utilize a composite of datasets including MSCOCO [27], FLICKR30k [48], Textcaps [38], IAPR TC-12 [17], and Nocaps [1] as  $D_{all}$ . These datasets contain approximately 113K, 30K 28K, 20K, and 15K images, respectively. Additionally, we generate artificial dataset images using Stable Diffusion v1.5 [36] based on the captions of FLICKR30k to simulate the creation of a new non-member dataset for the target model. In practice, the target models are trained on one dataset, and we sample from the combination of the other datasets to establish ground-truth non-members. We then further divide these non-members into two disjoint subsets: one provides the ground-truth non-member data as  $D_{no}$ , and another is utilized to assess the attack's performance.",
+        "bbox": [
+            511,
+            553,
+            906,
+            763
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the evaluation phase, both member and non-member data are split into training and testing sets. The training set is used as  $D_{all}$ , while the testing set is used to evaluate the attack model classifier. Our approach deviates from the traditional MIA evaluation setup, which typically involves splitting a single dataset into member and non-member subsets for small-scale classification models [21, 37]. Instead, similar to Ko et al. [22], we opted for a configuration that offers greater real-world relevance and practical impact.",
+        "bbox": [
+            511,
+            765,
+            908,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "9214",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Target models. We evaluate our attacks using three prominent public image captioning models from Hugging Face: BLIP Large [23], ViT-GPT2 [13], and GIT Base [42], all trained on large-scale datasets.",
+        "bbox": [
+            89,
+            90,
+            480,
+            151
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation metrics. We employ several key metrics to assess how effective our methods are: area under the curve (AUC), accuracy (ACC), and true positive rate (TPR) at a low false positive rate (FPR), which was used by [9].",
+        "bbox": [
+            89,
+            152,
+            482,
+            213
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2. Evaluation Results",
+        "text_level": 1,
+        "bbox": [
+            89,
+            224,
+            272,
+            239
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Attack performance. We evaluate the performance of our proposed attacks, MVTA and C-WSA. To perform a fair comparison with CSA and WSA [22], which utilize imagetext data pairs, we employ captions generated by the target models, producing a single caption per image using the top- $p$  sampling method, aligning with the approach taken in our methods to evaluate performance. We pair CSA with MVTA in a zero-knowledge setting and WSA with C-WSA in a partial-knowledge setting, enabling each method to be evaluated under equivalent data access conditions.",
+        "bbox": [
+            89,
+            250,
+            482,
+            398
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2 presents the evaluation results, showing the relative effectiveness of each attack method across the various metrics: AUC, ACC, and TPR at an FPR of  $1\\%$ . Overall, our proposed methods, MVTA and C-WSA, outperform the baseline CSA and WSA attacks across these key metrics, highlighting their robustness and precision in distinguishing member from non-member data.",
+        "bbox": [
+            89,
+            401,
+            482,
+            505
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "1. Performance in zero-knowledge setting: MVTA consistently outperforms CSA across all evaluated models. For instance, MVTA achieves AUC scores of 0.66 and 0.77 on BLIP and ViT-GPT2, respectively, compared to CSA's 0.54 and 0.72. This improvement demonstrates the effectiveness of MVTA's MV metric in capturing subtle distinctions in membership status even when data knowledge is minimal.",
+        "bbox": [
+            91,
+            507,
+            482,
+            626
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "2. Performance in partial-knowledge setting: C-WSA shows superior performance over WSA, especially on high-precision metrics such as TPR at an FPR of  $1\\%$ . C-WSA achieves an AUC of 0.91 on BLIP and 0.93 on ViT-GPT2, surpassing WSA's scores of 0.84 and 0.89, respectively. Additionally, C-WSA's TPR at an FPR of  $1\\%$  values are 0.38 for BLIP and 0.43 for ViT-GPT2, which indicates a substantial reduction in the number of false positives, highlighting C-WSA's accuracy in identifying true member instances.",
+        "bbox": [
+            89,
+            628,
+            482,
+            779
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "MVTA analysis. In this analysis, we focus on the MVTA and its use of the MV metric to distinguish between member and non-member data. Figure 4a presents the distribution of MV scores for both members and non-members. The histogram shows a separation between the two groups, with members tending to have lower MV scores while non-members exhibit higher values. We can observe that as the",
+        "bbox": [
+            89,
+            794,
+            482,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "threshold MV score decreases, the proportion of members increases, which is informative for determining an effective threshold for the attack and its consequences.",
+        "bbox": [
+            511,
+            90,
+            903,
+            136
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To illustrate how the captions from member and non-member images behave in the embedding space, Figure 4b presents features from the captions of one member and one non-member image in a two-dimensional space using t-SNE [41]. We find that the captions of member images are closer together than those of non-member images, indicating more consistent embeddings for members.",
+        "bbox": [
+            511,
+            136,
+            905,
+            242
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/f086fabc44e51918508d487d724aa6d1272fb2ae27e2fb2626d77985e9a51bbd.jpg",
+        "image_caption": [
+            "(a) Distribution of the MV metric values for members and non-members."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            563,
+            263,
+            831,
+            421
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/481cfb853e62f8b4fde4ea818bbd21edfde07ef2288845af7c2c03835267b50a.jpg",
+        "image_caption": [
+            "(b) Projection of member and non-member features from generated texts onto a two-dimensional embedding space using t-SNE [41].",
+            "Figure 4. Comparative visualization of the MV metric for members and non-members."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            578,
+            460,
+            831,
+            604
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "C-WSA analysis. In our evaluation of C-WSA, the parameter  $\\lambda$  is crucial for setting the threshold that determines whether a sample is classified as a member or non-member. Specifically, if a sample's MV score is lower than the mean score for non-members  $\\mu_{no}$  by more than  $\\lambda \\sigma_{no}$  (where  $\\sigma_{no}$  is the standard deviation of non-member MV scores), it is likely to be classified as a member. To identify the optimal value of  $\\lambda$ , we varied  $\\lambda$  from  $-2$  to 2 and established a baseline by randomly selecting samples from  $D_{all}$  as members for  $D_{p - mem}$ , independent of  $\\lambda$ . As shown in Table 3, increasing  $\\lambda$  raises the member ratio in  $D_{p - mem}$ , enhancing the accuracy of the labels for the attack model classifier. However, setting  $\\lambda$  too high may exclude relevant member",
+        "bbox": [
+            511,
+            704,
+            906,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "9215",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/a9bb6d567d6d4c566e2e68daf6977867bb4f43ee7f515c7d39263a2be3b31337.jpg",
+        "table_caption": [
+            "Table 2. Evaluation of attack performance on different public target models (BLIP, ViT-GPT2, and GIT) across the MSCOCO and TextCaps datasets. We bold the best results for each knowledge setting."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Dataset [Model]</td><td colspan=\"3\">MSCOCO [BLIP]</td><td colspan=\"3\">MSCOCO [ViT-GPT2]</td><td colspan=\"3\">MSCOCO [GIT]</td><td colspan=\"3\">Textcaps [GIT]</td></tr><tr><td>Method</td><td>AUC</td><td>ACC</td><td>TPR@FPR=1%</td><td>AUC</td><td>ACC</td><td>TPR@FPR=1%</td><td>AUC</td><td>ACC</td><td>TPR@FPR=1%</td><td>AUC</td><td>ACC</td><td>TPR@FPR=1%</td><td></td></tr><tr><td>CSA</td><td>0.5444</td><td>0.5358</td><td>0.0087</td><td>0.7207</td><td>0.6630</td><td>0.0466</td><td>0.5942</td><td>0.6265</td><td>0.0197</td><td>0.4484</td><td>0.5010</td><td>0.0110</td><td></td></tr><tr><td>MVTA</td><td>0.6601</td><td>0.6183</td><td>0.0224</td><td>0.7727</td><td>0.7026</td><td>0.0690</td><td>0.7630</td><td>0.7011</td><td>0.0441</td><td>0.5026</td><td>0.5124</td><td>0.0083</td><td></td></tr><tr><td>WSA</td><td>0.8398</td><td>0.7914</td><td>0.3144</td><td>0.8931</td><td>0.8415</td><td>0.3758</td><td>0.8626</td><td>0.8034</td><td>0.3328</td><td>0.8165</td><td>0.7945</td><td>0.3773</td><td></td></tr><tr><td>C-WSA</td><td>0.9131</td><td>0.8345</td><td>0.3826</td><td>0.9335</td><td>0.8619</td><td>0.4251</td><td>0.9255</td><td>0.8520</td><td>0.4075</td><td>0.8967</td><td>0.8479</td><td>0.4152</td><td></td></tr></table>",
+        "bbox": [
+            99,
+            127,
+            893,
+            196
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "data, potentially causing the attack model to miss broader patterns.",
+        "bbox": [
+            89,
+            220,
+            482,
+            250
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Our experiments indicate that  $\\lambda = 0.5$  achieves the highest average AUC values, with 0.91 for BLIP and 0.93 for ViT-GPT2. This choice not only maximizes the AUC but also achieves a favorable balance across the rest of the metrics, maintaining a TPR at a  $1\\%$  FPR.",
+        "bbox": [
+            89,
+            250,
+            482,
+            325
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We also evaluate the impact of adding a confidence threshold to the attack model's training. Figure 5 presents the average probability over the training epochs for C-WSA with a threshold (Figure 5a) and WSA without a threshold (Figure 5b). There are three possible groups of samples with pseudo-labels: members with label 1, non-members with label 1 (mislabeling), and non-members with label 0.",
+        "bbox": [
+            89,
+            325,
+            482,
+            431
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The inclusion of the confidence threshold enables the attack model to better manage the mislabeling group (non-members with label 1), as reflected in the reduced average probability for this group across the training epochs. When the threshold is applied, the attack model assigns lower probabilities to non-members, unlike the setting without the threshold, where the model continues to train on misclassified samples and assigns higher probabilities to them.",
+        "bbox": [
+            89,
+            431,
+            482,
+            551
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "However, this improvement comes at the cost of slightly lowering the average probability for true members with label 1. Despite this trade-off, the confidence threshold refines the training process by prioritizing reliable high-confidence samples. Further details on the selection of the confidence threshold and warm-up epoch parameters can be found in Appendix ??",
+        "bbox": [
+            89,
+            551,
+            482,
+            657
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Evaluation on fine-tuned public models. We also validate our approach on public pre-trained models that we fine-tuned. The results of this experiment (detailed in Appendix ??) show that our attacks outperform the baseline methods across most key metrics, further highlighting their effectiveness.",
+        "bbox": [
+            89,
+            672,
+            482,
+            763
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3. Sensitivity Analysis",
+        "text_level": 1,
+        "bbox": [
+            89,
+            771,
+            276,
+            787
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Impact of text generation decoding strategies. Figure 6 presents the performance of various attack methods in terms of the AUC metric as a function of the number of captions generated per image  $(n)$  and the text generation decoding strategy used (beam search or top-  $p$  sampling). The beam search results are placed at  $n = 0$ , while values from  $n = 1$  to  $n = 50$  correspond to captions generated using top-  $p$",
+        "bbox": [
+            89,
+            794,
+            482,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/c6b85213944d42dc6c21a8b65239f66b3cedeed368f0e3743a50bc473fec3c45.jpg",
+        "image_caption": [
+            "(a) C-WSA with threshold."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            516,
+            218,
+            709,
+            367
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4ed5791c254b1487139d2cb9aba117270d134965295cc000efa88ecb6d25df2d.jpg",
+        "image_caption": [
+            "(b) WSA without threshold.",
+            "Figure 5. Comparison of C-WSA with a confidence threshold and WSA without one, depicting the average probability of  $D_{p - mem}$  across the training epochs. M denotes membership status, and PL represents the pseudo-label assignment."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            709,
+            218,
+            903,
+            368
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "sampling. Notably, when there is only one caption generated, we perform our evaluation using the CS metric, because the MV metric, which calculates variance, cannot be applied.",
+        "bbox": [
+            511,
+            483,
+            906,
+            544
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We observe that the performance of the attacks improves when transitioning from beam search to top-  $p$  sampling. This suggests that the diversity introduced by top-  $p$  sampling better captures the nuances of the dataset, resulting in more informative features for the attack models.",
+        "bbox": [
+            511,
+            546,
+            906,
+            621
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Additionally, increasing the number of generated captions  $(n)$  positively impacts the performance of the attacks across all methods. This improvement is more pronounced for MVTA than C-WSA. A likely explanation for this difference is the relative simplicity of MVTA, which relies solely on the MV score calculated from the generated texts. In contrast, the greater complexity of C-WSA enables it to compensate for fewer generated captions, leveraging its advanced mechanisms to maintain competitive performance.",
+        "bbox": [
+            511,
+            623,
+            906,
+            760
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Based on the results, we set  $n = 40$  as a balance parameter between the number of captions generated and the resulting performance. This choice reflects a trade-off that achieves high AUC values while limiting the computational overhead associated with generating more captions.",
+        "bbox": [
+            511,
+            762,
+            906,
+            838
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Impact of member ratio in  $D_{all}$ . In this analysis, we evaluate the impact of the member ratio in  $D_{all}$  on the performance of the C-WSA attack. We vary the member ratio",
+        "bbox": [
+            511,
+            854,
+            908,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "9216",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/13889471387a619f2000e374106a7c598b2d8680d62ffc31da160e2632a45434.jpg",
+        "table_caption": [
+            "Table 3. Attack performance vs.  $\\lambda$  and the corresponding ratio of true members in pseudo-labeled data  $\\left( {\\lambda  = {0.5}}\\right.$  achieves optimal performance.)"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Dataset [Model]</td><td colspan=\"4\">MSCOCO [BLIP]</td><td colspan=\"4\">MSCOCO [ViT-GPT2]</td></tr><tr><td colspan=\"2\">Method</td><td colspan=\"4\">C-WSA</td><td colspan=\"4\">C-WSA</td></tr><tr><td>λ</td><td>Member Ratio in Dp-mem</td><td>AUC</td><td>TPR@FPR=1%</td><td>ACC</td><td>Member Ratio in Dp-mem</td><td>AUC</td><td>TPR@FPR=1%</td><td>ACC</td><td></td></tr><tr><td>Baseline</td><td>0.33</td><td>0.9041</td><td>0.3655</td><td>0.8359</td><td>0.33</td><td>0.6024</td><td>0.0755</td><td>0.5994</td><td></td></tr><tr><td>-2</td><td>0.34</td><td>0.8758</td><td>0.3408</td><td>0.8214</td><td>0.34</td><td>0.9152</td><td>0.3896</td><td>0.8428</td><td></td></tr><tr><td>-1.5</td><td>0.35</td><td>0.8867</td><td>0.3551</td><td>0.5997</td><td>0.35</td><td>0.9032</td><td>0.3557</td><td>0.8368</td><td></td></tr><tr><td>-1</td><td>0.36</td><td>0.8917</td><td>0.3622</td><td>0.8339</td><td>0.37</td><td>0.9083</td><td>0.4061</td><td>0.8541</td><td></td></tr><tr><td>-0.5</td><td>0.39</td><td>0.9236</td><td>0.4098</td><td>0.8541</td><td>0.41</td><td>0.9165</td><td>0.3665</td><td>0.8435</td><td></td></tr><tr><td>0</td><td>0.42</td><td>0.9166</td><td>0.3782</td><td>0.8484</td><td>0.46</td><td>0.9235</td><td>0.4142</td><td>0.8860</td><td></td></tr><tr><td>0.5</td><td>0.46</td><td>0.9131</td><td>0.3826</td><td>0.8345</td><td>0.53</td><td>0.9304</td><td>0.4106</td><td>0.8588</td><td></td></tr><tr><td>1.0</td><td>0.49</td><td>0.8809</td><td>0.2911</td><td>0.7991</td><td>0.60</td><td>0.9133</td><td>0.3671</td><td>0.8361</td><td></td></tr><tr><td>1.5</td><td>0.52</td><td>0.8040</td><td>0.1880</td><td>0.7335</td><td>0.69</td><td>0.8868</td><td>0.0755</td><td>0.8088</td><td></td></tr><tr><td>2</td><td>0.54</td><td>0.7236</td><td>0.1103</td><td>0.6685</td><td>0.74</td><td>0.8221</td><td>0.2080</td><td>0.7533</td><td></td></tr></table>",
+        "bbox": [
+            106,
+            127,
+            903,
+            287
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/7fff563a0d60d300b9d31a78fc002947c4146d192c8707e7db29638f075f3388.jpg",
+        "image_caption": [
+            "Figure 6. Comparison of AUC values for captions generated per image  $(n)$ . Beam search  $(bs)$  is used at  $n = 0$ , while top- $p$  sampling is used for  $n > 0$ . C refers to MSCOCO."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            99,
+            311,
+            473,
+            539
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "between 0.1 and 0.4 and assess the attack performance using two metrics: AUC and TPR at an FPR of  $1\\%$ , as shown in the Figure 7. The results show an improvement in attack performance as the member ratio increases. This is evident on both metrics, where higher member ratios lead to better AUC and TPR at a FPR of  $1\\%$  values. Intuitively, the performance gains indicate that a higher proportion of members in  $D_{all}$  provides the attack model with more representative samples, enhancing its ability to effectively distinguish members from non-members. Notably, even at a low member ratio of 0.1, the C-WSA attack achieves stable performance, with an AUC of 0.68 for BLIP and 0.84 for ViT-GPT2. For the TPR at a FPR of  $1\\%$ , the values are 0.19 for BLIP and 0.26 for ViT-GPT2. These results demonstrate that the attack remains effective even with a limited proportion of members in  $D_{all}$ .",
+        "bbox": [
+            88,
+            657,
+            482,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/e33c9072188bd395704203422f23d6468f29c3ed80b07553dd47f2fd519c49dd.jpg",
+        "image_caption": [
+            "Figure 7. Performance comparison in terms of AUC and TPR@FPR=1% across different member ratios in  $D_{all}$  for C-WSA configurations. C refers to MSCOCO."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            311,
+            702,
+            474
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/3137770a689ee9eceb6f1007d6d6a24aed209576c1eb42fddfc0ef3863ef3712.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            714,
+            314,
+            897,
+            473
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusions and Future Work",
+        "text_level": 1,
+        "bbox": [
+            511,
+            554,
+            790,
+            570
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We introduce two MIAs, MVTA and C-WSA, specifically designed for image captioning models in the zero-knowledge and partial-knowledge settings, respectively. These attacks leverage the MV metric, which serves as the foundation of our approach. The results obtained in our experiments highlight the vulnerability of image captioning models to MIAs and emphasize the privacy implications inherent in these models. Our evaluation showed that the effectiveness of our attacks is influenced by the text generation decoding strategies employed and the number of captions generated. Specifically, the diversity introduced by top- $p$  sampling and the variance in the number of generated captions provide a more informative set of features for MIAs. These factors affect our attacks' ability to distinguish between member and non-member data.",
+        "bbox": [
+            509,
+            580,
+            906,
+            805
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Future research could focus on: (1) developing and evaluating techniques to mitigate the effectiveness of the MVTA and C-WSA; (2) expanding the attacks' applicability to a wider range of models, such as diffusion models, by adapting the MV metric for use on different types of data (this would only require changes to the encoders).",
+        "bbox": [
+            511,
+            806,
+            908,
+            898
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "9217",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            91,
+            90,
+            187,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8948-8957, 2019. 5",
+            "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 1",
+            "[3] Jyoti Aneja, Aditya Deshpande, and Alexander G Schwing. Convolutional image captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5561-5570, 2018. 1",
+            "[4] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE transactions on pattern analysis and machine intelligence, 41(2):423-443, 2018. 1",
+            "[5] Siddharth Bhatia, Arjit Jain, and Bryan Hooi. Exgan: Adversarial generation of extreme samples. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6750-6758, 2021. 1",
+            "[6] Miles Brundage, Shahar Avin, Jack Clark, Helen Toner, Peter Eckersley, Ben Garfinkel, Allan Dafoe, Paul Scharre, Thomas Zeitzoff, Bobby Filar, et al. The malicious use of artificial intelligence: Forecasting, prevention, and mitigation. arXiv preprint arXiv:1802.07228, 2018. 1",
+            "[7] Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yutong Dai, Philip S Yu, and Lichao Sun. A comprehensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt. arXiv preprint arXiv:2303.04226, 2023. 1",
+            "[8] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21), pages 2633-2650, 2021. 3, 4",
+            "[9] Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and Florian Tramer. Membership inference attacks from first principles. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1897-1914. IEEE, 2022. 1, 2, 3, 6",
+            "[10] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. arXiv preprint arXiv:2202.07646, 2022. 1, 2",
+            "[11] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2",
+            "[12] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2"
+        ],
+        "bbox": [
+            93,
+            114,
+            483,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 6",
+            "[14] Jinhao Duan, Fei Kong, Shiqi Wang, Xiaoshuang Shi, and Kaidi Xu. Are diffusion models vulnerable to membership inference attacks? arXiv preprint arXiv:2302.01316, 2023. 1, 2, 3",
+            "[15] Michael Duan, Anshuman Suri, Niloofar Mireshghallah, Sewon Min, Weijia Shi, Luke Zettlemoyer, Yulia Tsvetkov, Yejin Choi, David Evans, and Hannaneh Hajishirzi. Do membership inference attacks work on large language models? arXiv preprint arXiv:2402.07841, 2024. 3",
+            "[16] Wenjie Fu, Huandong Wang, Chen Gao, Guanghua Liu, Yong Li, and Tao Jiang. Mia-tuner: Adapting large language models as pre-training text detector. arXiv preprint arXiv:2408.08661, 2024.3",
+            "[17] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International workshop ontoImage, 2006. 5",
+            "[18] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 4",
+            "[19] Hailong Hu and Jun Pang. Membership inference of diffusion models. arXiv preprint arXiv:2301.09956, 2023. 1, 2, 3",
+            "[20] Hongsheng Hu, Zoran Salcic, Lichao Sun, Gillian Dobbie, Philip S Yu, and Xuyun Zhang. Membership inference attacks on machine learning: A survey. ACM Computing Surveys (CSUR), 54(11s):1-37, 2022. 1, 2",
+            "[21] Pingyi Hu, Zihan Wang, Ruoxi Sun, Hu Wang, and Minhui Xue. M'4 i: Multi-modal models membership inference. Advances in Neural Information Processing Systems, 35:1867-1882, 2022. 1, 2, 3, 5",
+            "[22] Myeongseob Ko, Ming Jin, Chenguang Wang, and Ruoxi Jia. Practical membership inference attacks against large-scale multi-modal models: A pilot study. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4871-4881, 2023. 1, 2, 3, 4, 5, 6",
+            "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 1, 2, 3, 6",
+            "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 3",
+            "[25] Songze Li, Ruoxi Cheng, and Xiaojun Jia. Identity inference from clip models using only textual data. arXiv preprint arXiv:2405.14517, 2024. 1, 2, 3",
+            "[26] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. 1"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "9218",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5",
+            "[28] Bo Liu, Ming Ding, Sina Shaham, Wenny Rahayu, Farhad Farokhi, and Zihuai Lin. When machine learning meets privacy: A survey and outlook. ACM Computing Surveys (CSUR), 54(2):1-36, 2021. 1",
+            "[29] Pratyush Maini, Hengrui Jia, Nicolas Papernot, and Adam Dziedzic. Lm dataset inference: Did you train on my dataset? arXiv preprint arXiv:2406.06443, 2024. 3",
+            "[30] Tomoya Matsumoto, Takayuki Miura, and Naoto Yanai. Membership inference attacks against diffusion models. arXiv preprint arXiv:2302.03262, 2023. 1, 2, 3",
+            "[31] Justus Mattern, Fatemehsadat Mireshghallah, Zhijing Jin, Bernhard Scholkopf, Mrinmaya Sachan, and Taylor Berg-Kirkpatrick. Membership inference attacks against language models via neighbourhood comparison. arXiv preprint arXiv:2305.18462, 2023. 3",
+            "[32] Franz Josef Och and Hermann Ney. The alignment template approach to statistical machine translation. Computational linguistics, 30(4):417-449, 2004. 4",
+            "[33] Yan Pang and Tianhao Wang. Black-box membership inference attacks against fine-tuned diffusion models. arXiv preprint arXiv:2312.08207, 2023. 1, 2, 3",
+            "[34] Marco Pedersoli, Thomas Lucas, Cordelia Schmid, and Jakob Verbeek. Areas of attention for image captioning. In Proceedings of the IEEE international conference on computer vision, pages 1242-1250, 2017. 1",
+            "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2",
+            "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5",
+            "[37] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP), pages 3-18. IEEE, 2017. 1, 2, 5",
+            "[38] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 742–758. Springer, 2020. 5",
+            "[39] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1, 2"
+        ],
+        "bbox": [
+            91,
+            90,
+            480,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[40] Stacey Truex, Ling Liu, Mehmet Emre Gursoy, Lei Yu, and Wenqi Wei. Demystifying membership inference attacks in machine learning as a service. IEEE transactions on services computing, 14(6):2073-2089, 2019. 2",
+            "[41] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 6",
+            "[42] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2, 6",
+            "[43] Yuxin Wen, Arpit Bansal, Hamid Kazemi, Eitan Borgnia, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Canary in a coalmine: Better membership inference with ensembled adversarial queries. arXiv preprint arXiv:2210.10750, 2022. 1, 2, 3",
+            "[44] Yixin Wu, Ning Yu, Zheng Li, Michael Backes, and Yang Zhang. Membership inference attacks against text-to-image generation models. arXiv preprint arXiv:2210.00968, 2022. 1, 2, 3",
+            "[45] Chugui Xu, Ju Ren, Deyu Zhang, Yaoxue Zhang, Zhan Qin, and Kui Ren. Ganobfuscator: Mitigating information leakage under gan via differential privacy. IEEE Transactions on Information Forensics and Security, 14(9):2358-2371, 2019. 1",
+            "[46] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 2",
+            "[47] Samuel Yeom, Irene Giacomelli, Matt Fredrikson, and Somesh Jha. Privacy risk in machine learning: Analyzing the connection to overfitting. In 2018 IEEE 31st computer security foundations symposium (CSF), pages 268-282. IEEE, 2018. 1, 2",
+            "[48] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5",
+            "[49] Minxing Zhang, Ning Yu, Rui Wen, Michael Backes, and Yang Zhang. Generated distributions are all you need for membership inference attacks against generative models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4839-4849, 2024. 1, 2, 3",
+            "[50] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Houdong Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1"
+        ],
+        "bbox": [
+            514,
+            90,
+            905,
+            810
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "9219",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2025/Variance-Based Membership Inference Attacks Against Large-Scale Image Captioning Models/26bae7f8-8e37-4862-840e-6578a63d1d11_model.json

@@ -0,0 +1,2090 @@
+[
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.107,
+                0.003,
+                0.182,
+                0.043
+            ],
+            "angle": 0,
+            "content": "CVF"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.237,
+                0.001,
+                0.812,
+                0.047
+            ],
+            "angle": 0,
+            "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.123,
+                0.13,
+                0.877,
+                0.177
+            ],
+            "angle": 0,
+            "content": "Variance-Based Membership Inference Attacks Against Large-Scale Image Captioning Models"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.216,
+                0.204,
+                0.777,
+                0.257
+            ],
+            "angle": 0,
+            "content": "Daniel Samira Edan Habler Yuval Elovici Asaf Shabtai Ben-Gurion University of the Negev, Israel {samirada, habler}@post.bgu.ac.il, {elovici, shabaia}@bgu.ac.il"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.248,
+                0.292,
+                0.327,
+                0.308
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.325,
+                0.486,
+                0.644
+            ],
+            "angle": 0,
+            "content": "The proliferation of multi-modal generative models has introduced new privacy and security challenges, especially due to the risks of memorization and unintentional disclosure of sensitive information. This paper focuses on the vulnerability of multi-modal image captioning models to membership inference attacks (MIAs). These models, which synthesize textual descriptions from visual content, could inadvertently reveal personal or proprietary data embedded in their training datasets. We explore the feasibility of MIAs in the context of such models. Specifically, our approach leverages a variance-based strategy tailored for image captioning models, utilizing only image data without knowing the corresponding caption. We introduce the means-of-variance threshold attack (MVTA) and confidence-based weakly supervised attack (C-WSA) based on the metric, means-of-variance (MV), to assess variability among vector embeddings. Our experiments demonstrate that these models are susceptible to MIAs, indicating substantial privacy risks. The effectiveness of our methods is validated through rigorous evaluations on these real-world models, confirming the practical implications of our findings."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.092,
+                0.676,
+                0.222,
+                0.691
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.703,
+                0.483,
+                0.854
+            ],
+            "angle": 0,
+            "content": "The advancement from traditional machine learning (ML) to advanced generative models has progressed significantly in recent years [7]. An example of this development is seen in the field of multi-modal models, which are able to process information from various modalities [4]. These models have become increasingly sophisticated and are employed in challenging tasks involving various types of data (e.g., text, images, video). In this paper we focus on image captioning models [2, 3, 23, 24, 26, 34, 50], a multi-modal model used to generate textual descriptions for images."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.856,
+                0.484,
+                0.901
+            ],
+            "angle": 0,
+            "content": "While multi-modal models have fostered the development of a variety of new multi-modal applications, they also present new privacy and security challenges [6, 28]. A crit"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.293,
+                0.908,
+                0.475
+            ],
+            "angle": 0,
+            "content": "ical concern is the potential of generative models, particularly those trained on large private datasets, to inadvertently memorize and expose sensitive information. Such risks can lead to issues related to intellectual property misuse and privacy breaches involving data from proprietary images or documents [5, 45]. Among the various threats are membership inference attacks (MIAs), which aim to determine whether specific data records were included in a model's training dataset [20]. The ability to infer membership can result in the exposure of personal health information, financial details, or undisclosed affiliations, thereby posing a threat to individual privacy and security."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.476,
+                0.909,
+                0.718
+            ],
+            "angle": 0,
+            "content": "Initially, studies on MIAs focused on classification models, revealing heightened privacy risks primarily due to model overfitting [9, 20, 47]. Subsequently, attention turned to generative models, where research highlighted their tendency to memorize training data, raising concerns about digital forgery and the loss of originality [10, 11, 39]. In MIAs, as defined by Shokri et al. [37], shadow models are used to mimic the target model's behavior. However, crafting effective shadow models requires extensive datasets that mirror the original training data, along with substantial computational resources. Additionally, in many MIAs [14, 19, 21, 30, 33, 33, 43, 44, 49], the shadow models are scaled to hundreds of samples, and must share similar architectures with the target model and use identical training algorithms. Their dependence on shadow model training limits their applicability for large-scale models [21, 43]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.719,
+                0.909,
+                0.854
+            ],
+            "angle": 0,
+            "content": "To the best of our knowledge, the study performed by Hu et al. [21] is the only research exploring MIAs on image captioning models. The authors introduced two attack methods, which shadow models are employed. Furthermore, they did not directly evaluate the methods on operational image captioning models, such as those from Hugging Face. Instead, they were assessed in a simulated environment, using artificial models trained on synthetic configurations, small datasets, and numerous training epochs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.856,
+                0.909,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Recent studies have proposed MIAs that do not rely on shadow models [22, 25]. Specifically, Ko et al. [22] introduced the cosine similarity (CS) attack (CSA) and"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.482,
+                0.945,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "9210"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.092,
+                0.482,
+                0.243
+            ],
+            "angle": 0,
+            "content": "the weakly supervised attack (WSA). These attacks target large-scale multi-modal contrastive language-image pretraining (CLIP) models [35] that utilize text-image data pairs. Both methods use a CS metric to measure the similarity between the embedding vectors of images and text from the target CLIP model. The second method, WSA, combines ground-truth data about non-members with a pseudomember set, which consists of samples flagged as likely members based on their CS values. A classifier is trained on the dataset to distinguish members from non-members."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.252,
+                0.483,
+                0.464
+            ],
+            "angle": 0,
+            "content": "In this paper, we introduce two MIAs targeting image captioning models: the means-of-variance (MV) threshold attack (MVTA) and the confidence-based weakly supervised attack (C-WSA). In contrast to the techniques of Ko et al. [22], we tailor our methods specifically for image captioning models and enhance them with a new metric, the MV metric, which measures the variance between vectors. Our methods use only images as input, operating under the realistic assumption that an adversary lacks access to ground-truth reference text (i.e., captions) and instead uses captions generated by the target image captioning model. Additionally, we integrate a confidence threshold training method that leverages the pseudo-dataset's confidence to improve the classifier's performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.472,
+                0.483,
+                0.639
+            ],
+            "angle": 0,
+            "content": "We evaluate the proposed methods under varying levels of attacker knowledge (zero and partial) in a black-box scenario. Additionally, we conduct the evaluation using public models trained on large-scale datasets to simulate a realistic setting. The evaluation results show that our proposed methods identify membership information in multimodal image captioning models, using BLIP [23], ViT-GPT2 [13], and GIT [42], trained on large-scale datasets. Our methods enhance detection capabilities, achieving approximately \\(20\\%\\) to \\(30\\%\\) improvements over current methods. To summarize, our main contributions are as follows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.091,
+                0.659,
+                0.483,
+                0.72
+            ],
+            "angle": 0,
+            "content": "- New MIAs targeting image captioning models: We introduce the MVTA and C-WSA for the image captioning domain, two new attacks in which the input consists solely of images."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.092,
+                0.72,
+                0.483,
+                0.796
+            ],
+            "angle": 0,
+            "content": "- The means-of-variance metric: We introduce a quantitative metric designed to assess the variance between vectors in a multi-dimensional space. While we used it for texts, this metric is general and can be used in various fields that utilize embedding vectors."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.092,
+                0.797,
+                0.483,
+                0.855
+            ],
+            "angle": 0,
+            "content": "- Innovative training of the pseudo-member set: We introduce a new step in the process of training the attack model, a confidence-based training step that utilizes a pseudo-member set."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.092,
+                0.857,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "- Real-world application: By evaluating our methods on Hugging Face models, we demonstrate the practicality and effectiveness of our methods in real-world scenarios."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.091,
+                0.659,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.09,
+                0.655,
+                0.105
+            ],
+            "angle": 0,
+            "content": "2. Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.116,
+                0.905,
+                0.297
+            ],
+            "angle": 0,
+            "content": "Previous studies on MIAs exploited the model's tendency to produce more confident predictions for data included in the training set compared to unseen data [9, 20, 37, 40, 47], which is also referred to as memorization, since they tend to memorize the training data. MIAs are more challenging in the context of generative models, where confidence scores or prediction labels are typically unavailable [10, 11]. In addition, it was shown that the time it takes to train a generative model and the size of the dataset used to train the model influence the level of memorization, where extended training epochs and smaller datasets typically lead to increased memorization, enhancing the success of attacks [39]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.297,
+                0.905,
+                0.448
+            ],
+            "angle": 0,
+            "content": "Recent studies have demonstrated diffusion models' vulnerability to MIAs [12, 46]. In white-box settings, techniques involving statistical and neural network-based inference have shown high effectiveness, especially in models trained with small datasets and a large number of training epochs [14, 19, 30]. In black-box scenarios, methods leveraging semantic fidelity and novel classifier-based approaches have achieved high accuracy, though they often depend on access to non-training data and known training sets, which can limit their broader applicability [33, 44, 49]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.449,
+                0.905,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Attacks on multi-modal models. The CLIP model [35], which is one of the many types of multi-model models, utilizes contrastive learning to align text captions with images, leveraging unsupervised learning to interpret and generate image content based on natural language descriptions. Traditionally, MIAs targeting models like CLIP have relied on computationally intensive shadow models [14, 19, 21, 30, 33, 33, 43, 44, 49]. However, recent research has proposed efficient, shadow-free approaches for MIAs against multimodal models. The textual unimodal detector [25] identifies the use of personally identifiable information (PII) in CLIP models using only textual data, reducing both computational costs and privacy risks without the need for images. Another method, the WSA [22], addresses practical MIAs against large-scale multi-modal CLIP models by combining ground-truth information about non-members with noisy member data from a weak attack to develop an attack model. The WSA starts by querying the target model for image and text features of each non-member and evaluates the CS between these features. The CS distribution, typically Gaussian, helps approximate the mean and variance and is used to flag samples with a CS significantly above the mean as likely members. This identified set, known as the pseudo-member set, is then used to construct an attack dataset for training an attack model that determines the membership status of any target sample based on its features. Notably, the attack model does not account for the possibility that labels marked as true during training might be mislabeled, which could potentially improve performance if some of these were excluded from training."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.482,
+                0.945,
+                0.515,
+                0.957
+            ],
+            "angle": 0,
+            "content": "9211"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.09,
+                0.09,
+                0.908,
+                0.119
+            ],
+            "angle": 0,
+            "content": "Table 1. An overview of membership inference research on generation models. \"Reference Output\" indicates the attacker has access to training text references."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.093,
+                0.129,
+                0.905,
+                0.229
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Target Model Tasks</td><td>Attacks</td><td>Using Shadow Model</td><td>Final Model Prediction</td><td>Evaluated on Realistic Models</td><td>Other Information</td></tr><tr><td>Diffusion Models</td><td>[14, 19, 30, 33, 44, 49]</td><td>✓ or -</td><td>✓</td><td>✓ or -</td><td>Reference Output or -</td></tr><tr><td>LLMs</td><td>[8, 15, 16, 29, 31]</td><td>✓ or -</td><td>✓ or -</td><td>✓</td><td>-</td></tr><tr><td>Multi-Modal Models - CLIP</td><td>[22, 25]</td><td>-</td><td>✓</td><td>✓</td><td>-</td></tr><tr><td>Multi-Modal Models - Image Captioning</td><td>M4I [21]</td><td>✓</td><td>✓</td><td>-</td><td>Reference Output or -</td></tr><tr><td rowspan=\"2\">Ours</td><td>Means-of-Variance Threshold Attack</td><td>-</td><td>✓</td><td>✓</td><td>-</td></tr><tr><td>Confidence-Based Weakly Supervised Attack</td><td>-</td><td>✓</td><td>✓</td><td>-</td></tr></table>"
+        },
+        {
+            "type": "table_footnote",
+            "bbox": [
+                0.111,
+                0.23,
+                0.566,
+                0.242
+            ],
+            "angle": 0,
+            "content": "\"✓\" - the knowledge is necessary for the attack, \"-\" the knowledge is unnecessary for the attack"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.268,
+                0.486,
+                0.555
+            ],
+            "angle": 0,
+            "content": "Another type of multi-modal model, BLIP [23, 24], advances vision-language tasks by adapting to both understanding and generation challenges, including image-text retrieval, image captioning, and visual question answering. To the best of our knowledge, the work of Hu et al. [21] is the only one that explores MIAs on image captioning models. Their paper introduces two attack methods: metric-based and feature-based. The metric-based approach uses similarity metrics to infer membership status by comparing the similarity of model outputs with ground-truth data. The feature-based approach utilizes a pre-trained multi-modal feature extractor to analyze similarities between the input and output features. These attacks were not applied directly to operational image captioning models but rather in a controlled setting with synthetic configurations, in terms of small dataset size and a large number of training epochs, using an encoder-decoder framework to simulate the process. In addition, the authors also employed shadow models, which increases the complexity and computational costs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.556,
+                0.485,
+                0.722
+            ],
+            "angle": 0,
+            "content": "Unlike previous works, we focuses on the realm of multimodal image captioning models, an area that has received little research attention in the context of MIAs. In this research, we avoid the conventional reliance on computationally intensive shadow models, which necessitate datasets that closely mirror the training data of the target model. Instead, we base our attack methodologies on metrics designed to detect discrepancies in data variance. By applying MIAs in scenarios that only involve image data without access to corresponding textual information, our proposed MIAs have a broader range of applications."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.091,
+                0.723,
+                0.485,
+                0.754
+            ],
+            "angle": 0,
+            "content": "An overview of membership inference research across various generation model tasks is provided in Table 1."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.769,
+                0.233,
+                0.785
+            ],
+            "angle": 0,
+            "content": "3. Threat Model"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.795,
+                0.486,
+                0.903
+            ],
+            "angle": 0,
+            "content": "The primary goal of the attacker is to determine whether a given sample is a member of the target image captioning model's training set or not. The attacker can query the model, submitting images and receiving the corresponding generated text captions in response, but does not have any access to the model itself (i.e., query-only). The attacker does not have any knowledge of the target model's archi"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.268,
+                0.905,
+                0.284
+            ],
+            "angle": 0,
+            "content": "tectural, parameters, or the specific training algorithm used."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.284,
+                0.905,
+                0.375
+            ],
+            "angle": 0,
+            "content": "We emphasize that the attacker's dataset consists solely of images, allowing the attack to remain effective even in scenarios where access to the model's training captions or other textual references is limited or unavailable. We explore two different settings concerning the attacker's knowledge of the data used to train the target model:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.378,
+                0.907,
+                0.437
+            ],
+            "angle": 0,
+            "content": "- Zero-knowledge setting: The attacker has no prior knowledge of the distribution of the data used to train the target model, making this approach challenging but realistic in restricted data access scenarios."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.438,
+                0.909,
+                0.575
+            ],
+            "angle": 0,
+            "content": "- Partial-knowledge setting: Following prior research [9, 21, 22, 33, 43], the attacker has access to a publicly available dataset of images, \\( D_{all} \\), approximating the data distribution used to train the target model. The attacker also has access to a non-member dataset \\( D_{no} \\), while the membership status of \\( D_{all} \\) is unknown. The non-member dataset can be collected from content published after the model's release or synthesized by image-generating AI models, such as publicly available diffusion models."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.513,
+                0.378,
+                0.909,
+                0.575
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.589,
+                0.65,
+                0.607
+            ],
+            "angle": 0,
+            "content": "4. Methodology"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.614,
+                0.909,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Our methodology is designed to leverage our intuition regarding the use of variance in the outputs of a target image captioning model as a means of inferring infer membership information. The pipeline of our methodology consists of three main stages: text generation and multi-modal feature encoding, variance metric computation, and application of proposed attacks. First, the target model generates text captions from images. These captions and images are then encoded into feature embeddings, using pre-trained text and image encoders, respectively. The outputs are then analyzed to determine membership, aiming to ascertain whether a new image was part of the target model's training dataset. Next, we compute a variance-based metric, the MV metric, to quantify the variance among text embeddings generated for each image. This metric serves as the basis for identifying membership by capturing the relationship between variance and membership. Finally, we propose two attacks based on the MV metric: the MVTA and C-WSA. The MVTA operates in a zero-knowledge setting by exploit"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.482,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "9212"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.092,
+                0.483,
+                0.182
+            ],
+            "angle": 0,
+            "content": "ing the reduced variance of text embeddings for training set images. The C-WSA, designed for a partial-knowledge setting, combines non-member ground truth with probabilistic pseudo-member labels and employs a confidence threshold to train an attack classifier model. The stages of our methodology are described below."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.09,
+                0.189,
+                0.427,
+                0.206
+            ],
+            "angle": 0,
+            "content": "4.1. Text Generation and Feature Encoding"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.093,
+                0.213,
+                0.485,
+                0.544
+            ],
+            "angle": 0,
+            "content": "This stage involves two steps: text generation and multimodal feature encoding, as shown in Figure 1. Typically, image captioning models use two text generation decoding strategies to generate concise descriptions: beam search [32] and top- \\( p \\) sampling [18]. Beam search is deterministic, consistently generating the same outputs for a given input by keeping the best candidates when choosing the next word. In contrast, top- \\( p \\) sampling, also known as nucleus sampling, is non-deterministic and introduces diversity by selecting from a probabilistic \"nucleus\" of options. Carlini et al. [8] showed that when trained on paired data of the prompt and image, an image generation model generates similar images from the same prompt for two different random seeds. Following this principle, we denote the caption generated for an image \\( x \\) using top- \\( p \\) sampling by the model Gen as \\( \\mathrm{Gen}(x) \\). Then, if multiple captions generated for the same image \\( x \\) are highly similar under a distance measure \\( d \\) (\\( \\mathrm{Gen}(x) \\approx_d \\mathrm{Gen}(x) \\)), it suggests that the model has memorized the image. Based on this, we assume that the variance among the \\( n \\) captions generated for an image will be lower if the target model has been trained on that image."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.545,
+                0.484,
+                0.636
+            ],
+            "angle": 0,
+            "content": "During the feature encoding step, features are derived using a pretrained text encoder \\( f_{text} \\) to obtain text embeddings from the generated captions, and a pre-trained image encoder \\( f_{img} \\) to extract visual features from images. We note that the widespread availability of large-scale pre-trained models allows adversaries to easily obtain both encoders."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.104,
+                0.649,
+                0.475,
+                0.803
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.09,
+                0.814,
+                0.483,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Figure 1. Overview of our method's text generation and multimodal feature encoding stage. This stage consists of two steps: text generation, in which captions are generated using the target model, and feature encoding, in which features from multiple modalities (text and image) are encoded using pre-trained models."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.091,
+                0.755,
+                0.106
+            ],
+            "angle": 0,
+            "content": "4.2. Means-of-Variance Metric"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.114,
+                0.907,
+                0.281
+            ],
+            "angle": 0,
+            "content": "Based on our intuition, we introduce the new MV metric, which is designed to quantify the variance between embedding vectors. This metric constructs a variance vector of size \\( m \\), where each index represents the variance of a specific feature calculated across \\( n \\) vector embeddings. Once the variance vector is constructed, the metric then computes the mean of the values within it. The resulting single value represents the overall variance among the vectors, providing a concise measure of dispersion. Figure 2 provides an illustration of the MV metric calculation process, demonstrating how the MV metric is computed."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.543,
+                0.295,
+                0.882,
+                0.39
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.513,
+                0.401,
+                0.907,
+                0.457
+            ],
+            "angle": 0,
+            "content": "Figure 2. MV metric calculation. The feature matrix \\((f)\\) consists of \\(n\\) vectors with \\(m\\) features. The variance of each feature across these vectors is computed to form the variance vector \\((v)\\), whose mean gives the MV score."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.481,
+                0.686,
+                0.496
+            ],
+            "angle": 0,
+            "content": "4.3. Proposed Attacks"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.504,
+                0.906,
+                0.595
+            ],
+            "angle": 0,
+            "content": "Means-of-variance threshold attack. In a zero-knowledge setting, we compute the MV metric for the set of the embedding texts generated from each image, quantifying the average variance among these captions. If the calculated \\(MV\\) metric for image \\(x\\) falls below a predetermined threshold \\(\\tau\\), \\(x\\) is classified as a member of the training set:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.532,
+                0.602,
+                0.906,
+                0.639
+            ],
+            "angle": 0,
+            "content": "\\[\n\\text {P r e d i c t e d M e m b e r s h i p} = \\left\\{ \\begin{array}{l l} \\text {M e m b e r} & \\text {i f} M V <   \\tau \\\\ \\text {N o n - M e m b e r} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.644,
+                0.907,
+                0.72
+            ],
+            "angle": 0,
+            "content": "The threshold \\(\\tau\\) is empirically determined by analyzing the distribution of MV scores for both members and non-members. In practice, MVTA serves as a baseline for C-WSA, with \\(\\tau\\) calculated using the mean and standard deviation of MV scores from non-member datasets \\((D_{no})\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.72,
+                0.906,
+                0.751
+            ],
+            "angle": 0,
+            "content": "This approach adapts confidence or loss-based thresholds for data membership to multi-modal image captioning."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.765,
+                0.909,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Confidence-based weakly supervised attack. Building on the WSA method [22], our second attack, the C-WSA, introduces an innovative adaptation in the partial-knowledge setting by utilizing the MV metric to enhance the model's ability to distinguish between member and non-member data. An overview of the attack is provided in Figure 3. Similar to a typical WSA, our attack integrates ground-truth information about non-members, along with probabilistic assessments of membership."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.482,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "9213"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.107,
+                0.093,
+                0.473,
+                0.24
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.18,
+                0.254,
+                0.393,
+                0.267
+            ],
+            "angle": 0,
+            "content": "Figure 3. Overview of the C-WSA."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.293,
+                0.484,
+                0.43
+            ],
+            "angle": 0,
+            "content": "In our modified attack pipeline, we begin by querying the target model with images from the known non-member dataset \\((D_{no})\\) to generate \\(n\\) captions per image. Given the availability of public or synthetic datasets, an adversary can readily access or create them. Then, we calculate the MV metric for the dataset in order to establish a baseline distribution of variance scores among non-members, which we hypothesized to follow a Gaussian distribution. The MV scores for non-members are denoted as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.138,
+                0.44,
+                0.483,
+                0.457
+            ],
+            "angle": 0,
+            "content": "\\[\nM V _ {n o} = \\left\\{M V \\left(f _ {t e x t} (G e n (x))\\right) \\mid x \\in D _ {n o} \\right\\} \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.466,
+                0.484,
+                0.526
+            ],
+            "angle": 0,
+            "content": "where \\( f_{text}(Gen(x)) \\) refers to the text embeddings extracted by the text encoder \\( f_{text} \\) from the captions generated by the target model for each image \\( x \\), and \\( MV \\) is the function that performs the metric on the vectors."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.527,
+                0.484,
+                0.556
+            ],
+            "angle": 0,
+            "content": "Using \\(MV_{no}\\), we calculate the mean \\(\\mu_{no}\\) and standard deviation \\(\\sigma_{no}\\) of the MV scores:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.14,
+                0.568,
+                0.482,
+                0.584
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mu_ {n o} = \\operatorname {m e a n} \\left(M V _ {n o}\\right), \\quad \\sigma_ {n o} = \\operatorname {s t d} \\left(M V _ {n o}\\right) \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.594,
+                0.484,
+                0.684
+            ],
+            "angle": 0,
+            "content": "These parameters \\((\\mu_{no}\\) and \\(\\sigma_{no})\\) are crucial for identifying potential members by comparing their MV scores to the non-member distribution baseline mentioned above. Intuitively, if a sample's \\(MV\\) is substantially lower than \\(\\mu_{no}\\), specifically by more than \\(\\lambda \\sigma_{no}\\) for a given constant \\(\\lambda\\), it suggests that the sample is likely a member."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.685,
+                0.484,
+                0.835
+            ],
+            "angle": 0,
+            "content": "For the dataset \\(D_{all}\\), which comprises potential members, we compute the MV scores using the text embeddings generated by the target model. Samples in \\(D_{all}\\) whose MV scores significantly deviate from this non-member baseline are identified as potential members. Specifically, if a sample's MV score exceeds a defined threshold calculated as \\(\\mu_{no} - \\lambda \\sigma_{no}\\), it is classified as a pseudo-member. The corresponding subset of these identified samples (\\(D_{p-mem}\\)), serves as our pseudo-member set. This set comprises a mix of likely members and potential false positives."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.11,
+                0.835,
+                0.483,
+                0.85
+            ],
+            "angle": 0,
+            "content": "The formulas associated with this process are as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.23,
+                0.863,
+                0.483,
+                0.877
+            ],
+            "angle": 0,
+            "content": "\\[\n\\tau = \\mu_ {n o} - \\lambda \\sigma_ {n o} \\tag {4}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.11,
+                0.886,
+                0.483,
+                0.902
+            ],
+            "angle": 0,
+            "content": "\\[\nD _ {p - m e m} = \\{x \\mid M V (f _ {\\text {t e x t}} (\\operatorname {G e n} (x))) <   \\tau , x \\in D _ {\\text {a l l}} \\} \\tag {5}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.092,
+                0.905,
+                0.138
+            ],
+            "angle": 0,
+            "content": "With the \\(D_{p - mem}\\) and \\(D_{no}\\) in place, we proceed to create the attack training set \\(D_{attack}\\), comprising image features from \\(D_{no} \\cup D_{p - mem}\\) with their associated binary labels:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.537,
+                0.142,
+                0.905,
+                0.158
+            ],
+            "angle": 0,
+            "content": "\\[\nD _ {\\text {a t t a c k}} = \\left\\{f _ {i m g} (x) \\mid x \\in D _ {n o} \\cup D _ {p - m e m}, b = 1 [ x \\in D _ {p - m e m} ] \\right\\} \\tag {6}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.163,
+                0.907,
+                0.268
+            ],
+            "angle": 0,
+            "content": "where \\( f_{img}(x) \\) represents the image features extracted by the image encoder \\( f_{img} \\) for each image \\( x \\), and \\( b \\) is a binary indicator with 1 representing a potential member and 0 representing a non-member. Finally, using \\( D_{all} \\), we train the attack classifier model \\( f_{attack} \\) so that it is capable of determining the membership status of any target sample by evaluating the sample's image features."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.269,
+                0.908,
+                0.495
+            ],
+            "angle": 0,
+            "content": "In \\( D_{all} \\), samples labeled with 0 are definitively non-members, whereas those labeled with 1 may not actually be members. This distinction guides our training strategy for the attack classifier model \\( f_{attack} \\), which incorporates a confidence threshold. Initially, the attack model undergoes a \"warm-up\" phase, training without interruptions to establish baseline prediction tendencies. This phase lasts for a predefined ratio of epochs, referred to as \"warm-epoch.\" Then, before each training batch, we evaluate the membership confidence of the samples. For those identified as pseudo-members (\\( b = 1 \\)), training continues only if their membership confidence exceeds a predetermined threshold. This approach refines the training process, focusing primarily on samples most likely to be members, thus enhancing the accuracy and effectiveness of our attack classifier."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.506,
+                0.631,
+                0.521
+            ],
+            "angle": 0,
+            "content": "5. Evaluation"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.531,
+                0.683,
+                0.547
+            ],
+            "angle": 0,
+            "content": "5.1. Evaluation Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.554,
+                0.907,
+                0.765
+            ],
+            "angle": 0,
+            "content": "Datasets. To evaluate our attacks, we utilize a composite of datasets including MSCOCO [27], FLICKR30k [48], Textcaps [38], IAPR TC-12 [17], and Nocaps [1] as \\( D_{all} \\). These datasets contain approximately 113K, 30K 28K, 20K, and 15K images, respectively. Additionally, we generate artificial dataset images using Stable Diffusion v1.5 [36] based on the captions of FLICKR30k to simulate the creation of a new non-member dataset for the target model. In practice, the target models are trained on one dataset, and we sample from the combination of the other datasets to establish ground-truth non-members. We then further divide these non-members into two disjoint subsets: one provides the ground-truth non-member data as \\( D_{no} \\), and another is utilized to assess the attack's performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.766,
+                0.909,
+                0.902
+            ],
+            "angle": 0,
+            "content": "In the evaluation phase, both member and non-member data are split into training and testing sets. The training set is used as \\( D_{all} \\), while the testing set is used to evaluate the attack model classifier. Our approach deviates from the traditional MIA evaluation setup, which typically involves splitting a single dataset into member and non-member subsets for small-scale classification models [21, 37]. Instead, similar to Ko et al. [22], we opted for a configuration that offers greater real-world relevance and practical impact."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.482,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "9214"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.092,
+                0.482,
+                0.152
+            ],
+            "angle": 0,
+            "content": "Target models. We evaluate our attacks using three prominent public image captioning models from Hugging Face: BLIP Large [23], ViT-GPT2 [13], and GIT Base [42], all trained on large-scale datasets."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.153,
+                0.483,
+                0.214
+            ],
+            "angle": 0,
+            "content": "Evaluation metrics. We employ several key metrics to assess how effective our methods are: area under the curve (AUC), accuracy (ACC), and true positive rate (TPR) at a low false positive rate (FPR), which was used by [9]."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.225,
+                0.273,
+                0.24
+            ],
+            "angle": 0,
+            "content": "5.2. Evaluation Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.25,
+                0.483,
+                0.4
+            ],
+            "angle": 0,
+            "content": "Attack performance. We evaluate the performance of our proposed attacks, MVTA and C-WSA. To perform a fair comparison with CSA and WSA [22], which utilize imagetext data pairs, we employ captions generated by the target models, producing a single caption per image using the top-\\(p\\) sampling method, aligning with the approach taken in our methods to evaluate performance. We pair CSA with MVTA in a zero-knowledge setting and WSA with C-WSA in a partial-knowledge setting, enabling each method to be evaluated under equivalent data access conditions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.402,
+                0.483,
+                0.506
+            ],
+            "angle": 0,
+            "content": "Table 2 presents the evaluation results, showing the relative effectiveness of each attack method across the various metrics: AUC, ACC, and TPR at an FPR of \\(1\\%\\). Overall, our proposed methods, MVTA and C-WSA, outperform the baseline CSA and WSA attacks across these key metrics, highlighting their robustness and precision in distinguishing member from non-member data."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.092,
+                0.508,
+                0.483,
+                0.627
+            ],
+            "angle": 0,
+            "content": "1. Performance in zero-knowledge setting: MVTA consistently outperforms CSA across all evaluated models. For instance, MVTA achieves AUC scores of 0.66 and 0.77 on BLIP and ViT-GPT2, respectively, compared to CSA's 0.54 and 0.72. This improvement demonstrates the effectiveness of MVTA's MV metric in capturing subtle distinctions in membership status even when data knowledge is minimal."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.629,
+                0.483,
+                0.78
+            ],
+            "angle": 0,
+            "content": "2. Performance in partial-knowledge setting: C-WSA shows superior performance over WSA, especially on high-precision metrics such as TPR at an FPR of \\(1\\%\\). C-WSA achieves an AUC of 0.91 on BLIP and 0.93 on ViT-GPT2, surpassing WSA's scores of 0.84 and 0.89, respectively. Additionally, C-WSA's TPR at an FPR of \\(1\\%\\) values are 0.38 for BLIP and 0.43 for ViT-GPT2, which indicates a substantial reduction in the number of false positives, highlighting C-WSA's accuracy in identifying true member instances."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.795,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "MVTA analysis. In this analysis, we focus on the MVTA and its use of the MV metric to distinguish between member and non-member data. Figure 4a presents the distribution of MV scores for both members and non-members. The histogram shows a separation between the two groups, with members tending to have lower MV scores while non-members exhibit higher values. We can observe that as the"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.092,
+                0.905,
+                0.137
+            ],
+            "angle": 0,
+            "content": "threshold MV score decreases, the proportion of members increases, which is informative for determining an effective threshold for the attack and its consequences."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.137,
+                0.906,
+                0.243
+            ],
+            "angle": 0,
+            "content": "To illustrate how the captions from member and non-member images behave in the embedding space, Figure 4b presents features from the captions of one member and one non-member image in a two-dimensional space using t-SNE [41]. We find that the captions of member images are closer together than those of non-member images, indicating more consistent embeddings for members."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.565,
+                0.265,
+                0.832,
+                0.422
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.563,
+                0.426,
+                0.857,
+                0.451
+            ],
+            "angle": 0,
+            "content": "(a) Distribution of the MV metric values for members and non-members."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.58,
+                0.462,
+                0.832,
+                0.605
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.562,
+                0.612,
+                0.858,
+                0.649
+            ],
+            "angle": 0,
+            "content": "(b) Projection of member and non-member features from generated texts onto a two-dimensional embedding space using t-SNE [41]."
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.512,
+                0.661,
+                0.905,
+                0.688
+            ],
+            "angle": 0,
+            "content": "Figure 4. Comparative visualization of the MV metric for members and non-members."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.705,
+                0.907,
+                0.901
+            ],
+            "angle": 0,
+            "content": "C-WSA analysis. In our evaluation of C-WSA, the parameter \\(\\lambda\\) is crucial for setting the threshold that determines whether a sample is classified as a member or non-member. Specifically, if a sample's MV score is lower than the mean score for non-members \\(\\mu_{no}\\) by more than \\(\\lambda \\sigma_{no}\\) (where \\(\\sigma_{no}\\) is the standard deviation of non-member MV scores), it is likely to be classified as a member. To identify the optimal value of \\(\\lambda\\), we varied \\(\\lambda\\) from \\(-2\\) to 2 and established a baseline by randomly selecting samples from \\(D_{all}\\) as members for \\(D_{p - mem}\\), independent of \\(\\lambda\\). As shown in Table 3, increasing \\(\\lambda\\) raises the member ratio in \\(D_{p - mem}\\), enhancing the accuracy of the labels for the attack model classifier. However, setting \\(\\lambda\\) too high may exclude relevant member"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.482,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "9215"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.09,
+                0.089,
+                0.908,
+                0.119
+            ],
+            "angle": 0,
+            "content": "Table 2. Evaluation of attack performance on different public target models (BLIP, ViT-GPT2, and GIT) across the MSCOCO and TextCaps datasets. We bold the best results for each knowledge setting."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.101,
+                0.128,
+                0.895,
+                0.198
+            ],
+            "angle": 0,
+            "content": "<table><tr><td colspan=\"2\">Dataset [Model]</td><td colspan=\"3\">MSCOCO [BLIP]</td><td colspan=\"3\">MSCOCO [ViT-GPT2]</td><td colspan=\"3\">MSCOCO [GIT]</td><td colspan=\"3\">Textcaps [GIT]</td></tr><tr><td>Method</td><td>AUC</td><td>ACC</td><td>TPR@FPR=1%</td><td>AUC</td><td>ACC</td><td>TPR@FPR=1%</td><td>AUC</td><td>ACC</td><td>TPR@FPR=1%</td><td>AUC</td><td>ACC</td><td>TPR@FPR=1%</td><td></td></tr><tr><td>CSA</td><td>0.5444</td><td>0.5358</td><td>0.0087</td><td>0.7207</td><td>0.6630</td><td>0.0466</td><td>0.5942</td><td>0.6265</td><td>0.0197</td><td>0.4484</td><td>0.5010</td><td>0.0110</td><td></td></tr><tr><td>MVTA</td><td>0.6601</td><td>0.6183</td><td>0.0224</td><td>0.7727</td><td>0.7026</td><td>0.0690</td><td>0.7630</td><td>0.7011</td><td>0.0441</td><td>0.5026</td><td>0.5124</td><td>0.0083</td><td></td></tr><tr><td>WSA</td><td>0.8398</td><td>0.7914</td><td>0.3144</td><td>0.8931</td><td>0.8415</td><td>0.3758</td><td>0.8626</td><td>0.8034</td><td>0.3328</td><td>0.8165</td><td>0.7945</td><td>0.3773</td><td></td></tr><tr><td>C-WSA</td><td>0.9131</td><td>0.8345</td><td>0.3826</td><td>0.9335</td><td>0.8619</td><td>0.4251</td><td>0.9255</td><td>0.8520</td><td>0.4075</td><td>0.8967</td><td>0.8479</td><td>0.4152</td><td></td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.222,
+                0.483,
+                0.251
+            ],
+            "angle": 0,
+            "content": "data, potentially causing the attack model to miss broader patterns."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.251,
+                0.483,
+                0.327
+            ],
+            "angle": 0,
+            "content": "Our experiments indicate that \\(\\lambda = 0.5\\) achieves the highest average AUC values, with 0.91 for BLIP and 0.93 for ViT-GPT2. This choice not only maximizes the AUC but also achieves a favorable balance across the rest of the metrics, maintaining a TPR at a \\(1\\%\\) FPR."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.327,
+                0.483,
+                0.432
+            ],
+            "angle": 0,
+            "content": "We also evaluate the impact of adding a confidence threshold to the attack model's training. Figure 5 presents the average probability over the training epochs for C-WSA with a threshold (Figure 5a) and WSA without a threshold (Figure 5b). There are three possible groups of samples with pseudo-labels: members with label 1, non-members with label 1 (mislabeling), and non-members with label 0."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.432,
+                0.483,
+                0.552
+            ],
+            "angle": 0,
+            "content": "The inclusion of the confidence threshold enables the attack model to better manage the mislabeling group (non-members with label 1), as reflected in the reduced average probability for this group across the training epochs. When the threshold is applied, the attack model assigns lower probabilities to non-members, unlike the setting without the threshold, where the model continues to train on misclassified samples and assigns higher probabilities to them."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.553,
+                0.483,
+                0.659
+            ],
+            "angle": 0,
+            "content": "However, this improvement comes at the cost of slightly lowering the average probability for true members with label 1. Despite this trade-off, the confidence threshold refines the training process by prioritizing reliable high-confidence samples. Further details on the selection of the confidence threshold and warm-up epoch parameters can be found in Appendix ??"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.674,
+                0.483,
+                0.764
+            ],
+            "angle": 0,
+            "content": "Evaluation on fine-tuned public models. We also validate our approach on public pre-trained models that we fine-tuned. The results of this experiment (detailed in Appendix ??) show that our attacks outperform the baseline methods across most key metrics, further highlighting their effectiveness."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.772,
+                0.277,
+                0.789
+            ],
+            "angle": 0,
+            "content": "5.3. Sensitivity Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.795,
+                0.483,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Impact of text generation decoding strategies. Figure 6 presents the performance of various attack methods in terms of the AUC metric as a function of the number of captions generated per image \\((n)\\) and the text generation decoding strategy used (beam search or top- \\(p\\) sampling). The beam search results are placed at \\(n = 0\\), while values from \\(n = 1\\) to \\(n = 50\\) correspond to captions generated using top- \\(p\\)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.517,
+                0.219,
+                0.71,
+                0.368
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.542,
+                0.372,
+                0.686,
+                0.384
+            ],
+            "angle": 0,
+            "content": "(a) C-WSA with threshold."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.71,
+                0.219,
+                0.905,
+                0.369
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.734,
+                0.372,
+                0.881,
+                0.384
+            ],
+            "angle": 0,
+            "content": "(b) WSA without threshold."
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.513,
+                0.397,
+                0.908,
+                0.454
+            ],
+            "angle": 0,
+            "content": "Figure 5. Comparison of C-WSA with a confidence threshold and WSA without one, depicting the average probability of \\( D_{p - mem} \\) across the training epochs. M denotes membership status, and PL represents the pseudo-label assignment."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.484,
+                0.907,
+                0.545
+            ],
+            "angle": 0,
+            "content": "sampling. Notably, when there is only one caption generated, we perform our evaluation using the CS metric, because the MV metric, which calculates variance, cannot be applied."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.547,
+                0.907,
+                0.622
+            ],
+            "angle": 0,
+            "content": "We observe that the performance of the attacks improves when transitioning from beam search to top- \\( p \\) sampling. This suggests that the diversity introduced by top- \\( p \\) sampling better captures the nuances of the dataset, resulting in more informative features for the attack models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.624,
+                0.907,
+                0.761
+            ],
+            "angle": 0,
+            "content": "Additionally, increasing the number of generated captions \\((n)\\) positively impacts the performance of the attacks across all methods. This improvement is more pronounced for MVTA than C-WSA. A likely explanation for this difference is the relative simplicity of MVTA, which relies solely on the MV score calculated from the generated texts. In contrast, the greater complexity of C-WSA enables it to compensate for fewer generated captions, leveraging its advanced mechanisms to maintain competitive performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.763,
+                0.907,
+                0.839
+            ],
+            "angle": 0,
+            "content": "Based on the results, we set \\( n = 40 \\) as a balance parameter between the number of captions generated and the resulting performance. This choice reflects a trade-off that achieves high AUC values while limiting the computational overhead associated with generating more captions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.856,
+                0.909,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Impact of member ratio in \\( D_{all} \\). In this analysis, we evaluate the impact of the member ratio in \\( D_{all} \\) on the performance of the C-WSA attack. We vary the member ratio"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.482,
+                0.945,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "9216"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.09,
+                0.089,
+                0.908,
+                0.119
+            ],
+            "angle": 0,
+            "content": "Table 3. Attack performance vs. \\( \\lambda \\) and the corresponding ratio of true members in pseudo-labeled data \\( \\left( {\\lambda  = {0.5}}\\right. \\) achieves optimal performance.)"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.107,
+                0.128,
+                0.905,
+                0.288
+            ],
+            "angle": 0,
+            "content": "<table><tr><td colspan=\"2\">Dataset [Model]</td><td colspan=\"4\">MSCOCO [BLIP]</td><td colspan=\"4\">MSCOCO [ViT-GPT2]</td></tr><tr><td colspan=\"2\">Method</td><td colspan=\"4\">C-WSA</td><td colspan=\"4\">C-WSA</td></tr><tr><td>λ</td><td>Member Ratio in Dp-mem</td><td>AUC</td><td>TPR@FPR=1%</td><td>ACC</td><td>Member Ratio in Dp-mem</td><td>AUC</td><td>TPR@FPR=1%</td><td>ACC</td><td></td></tr><tr><td>Baseline</td><td>0.33</td><td>0.9041</td><td>0.3655</td><td>0.8359</td><td>0.33</td><td>0.6024</td><td>0.0755</td><td>0.5994</td><td></td></tr><tr><td>-2</td><td>0.34</td><td>0.8758</td><td>0.3408</td><td>0.8214</td><td>0.34</td><td>0.9152</td><td>0.3896</td><td>0.8428</td><td></td></tr><tr><td>-1.5</td><td>0.35</td><td>0.8867</td><td>0.3551</td><td>0.5997</td><td>0.35</td><td>0.9032</td><td>0.3557</td><td>0.8368</td><td></td></tr><tr><td>-1</td><td>0.36</td><td>0.8917</td><td>0.3622</td><td>0.8339</td><td>0.37</td><td>0.9083</td><td>0.4061</td><td>0.8541</td><td></td></tr><tr><td>-0.5</td><td>0.39</td><td>0.9236</td><td>0.4098</td><td>0.8541</td><td>0.41</td><td>0.9165</td><td>0.3665</td><td>0.8435</td><td></td></tr><tr><td>0</td><td>0.42</td><td>0.9166</td><td>0.3782</td><td>0.8484</td><td>0.46</td><td>0.9235</td><td>0.4142</td><td>0.8860</td><td></td></tr><tr><td>0.5</td><td>0.46</td><td>0.9131</td><td>0.3826</td><td>0.8345</td><td>0.53</td><td>0.9304</td><td>0.4106</td><td>0.8588</td><td></td></tr><tr><td>1.0</td><td>0.49</td><td>0.8809</td><td>0.2911</td><td>0.7991</td><td>0.60</td><td>0.9133</td><td>0.3671</td><td>0.8361</td><td></td></tr><tr><td>1.5</td><td>0.52</td><td>0.8040</td><td>0.1880</td><td>0.7335</td><td>0.69</td><td>0.8868</td><td>0.0755</td><td>0.8088</td><td></td></tr><tr><td>2</td><td>0.54</td><td>0.7236</td><td>0.1103</td><td>0.6685</td><td>0.74</td><td>0.8221</td><td>0.2080</td><td>0.7533</td><td></td></tr></table>"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.101,
+                0.313,
+                0.475,
+                0.54
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.09,
+                0.554,
+                0.483,
+                0.597
+            ],
+            "angle": 0,
+            "content": "Figure 6. Comparison of AUC values for captions generated per image \\((n)\\). Beam search \\((bs)\\) is used at \\(n = 0\\), while top-\\(p\\) sampling is used for \\(n > 0\\). C refers to MSCOCO."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.659,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "between 0.1 and 0.4 and assess the attack performance using two metrics: AUC and TPR at an FPR of \\(1\\%\\), as shown in the Figure 7. The results show an improvement in attack performance as the member ratio increases. This is evident on both metrics, where higher member ratios lead to better AUC and TPR at a FPR of \\(1\\%\\) values. Intuitively, the performance gains indicate that a higher proportion of members in \\(D_{all}\\) provides the attack model with more representative samples, enhancing its ability to effectively distinguish members from non-members. Notably, even at a low member ratio of 0.1, the C-WSA attack achieves stable performance, with an AUC of 0.68 for BLIP and 0.84 for ViT-GPT2. For the TPR at a FPR of \\(1\\%\\), the values are 0.19 for BLIP and 0.26 for ViT-GPT2. These results demonstrate that the attack remains effective even with a limited proportion of members in \\(D_{all}\\)."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.522,
+                0.313,
+                0.703,
+                0.475
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.715,
+                0.315,
+                0.898,
+                0.474
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.513,
+                0.488,
+                0.908,
+                0.531
+            ],
+            "angle": 0,
+            "content": "Figure 7. Performance comparison in terms of AUC and TPR@FPR=1% across different member ratios in \\( D_{all} \\) for C-WSA configurations. C refers to MSCOCO."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.555,
+                0.792,
+                0.571
+            ],
+            "angle": 0,
+            "content": "6. Conclusions and Future Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.511,
+                0.581,
+                0.907,
+                0.806
+            ],
+            "angle": 0,
+            "content": "We introduce two MIAs, MVTA and C-WSA, specifically designed for image captioning models in the zero-knowledge and partial-knowledge settings, respectively. These attacks leverage the MV metric, which serves as the foundation of our approach. The results obtained in our experiments highlight the vulnerability of image captioning models to MIAs and emphasize the privacy implications inherent in these models. Our evaluation showed that the effectiveness of our attacks is influenced by the text generation decoding strategies employed and the number of captions generated. Specifically, the diversity introduced by top-\\(p\\) sampling and the variance in the number of generated captions provide a more informative set of features for MIAs. These factors affect our attacks' ability to distinguish between member and non-member data."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.807,
+                0.909,
+                0.899
+            ],
+            "angle": 0,
+            "content": "Future research could focus on: (1) developing and evaluating techniques to mitigate the effectiveness of the MVTA and C-WSA; (2) expanding the attacks' applicability to a wider range of models, such as diffusion models, by adapting the MV metric for use on different types of data (this would only require changes to the encoders)."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.482,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "9217"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.093,
+                0.091,
+                0.188,
+                0.106
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.101,
+                0.115,
+                0.484,
+                0.184
+            ],
+            "angle": 0,
+            "content": "[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8948-8957, 2019. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.101,
+                0.186,
+                0.484,
+                0.268
+            ],
+            "angle": 0,
+            "content": "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.271,
+                0.483,
+                0.324
+            ],
+            "angle": 0,
+            "content": "[3] Jyoti Aneja, Aditya Deshpande, and Alexander G Schwing. Convolutional image captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5561-5570, 2018. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.326,
+                0.483,
+                0.38
+            ],
+            "angle": 0,
+            "content": "[4] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE transactions on pattern analysis and machine intelligence, 41(2):423-443, 2018. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.382,
+                0.484,
+                0.436
+            ],
+            "angle": 0,
+            "content": "[5] Siddharth Bhatia, Arjit Jain, and Bryan Hooi. Exgan: Adversarial generation of extreme samples. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6750-6758, 2021. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.439,
+                0.482,
+                0.507
+            ],
+            "angle": 0,
+            "content": "[6] Miles Brundage, Shahar Avin, Jack Clark, Helen Toner, Peter Eckersley, Ben Garfinkel, Allan Dafoe, Paul Scharre, Thomas Zeitzoff, Bobby Filar, et al. The malicious use of artificial intelligence: Forecasting, prevention, and mitigation. arXiv preprint arXiv:1802.07228, 2018. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.509,
+                0.483,
+                0.564
+            ],
+            "angle": 0,
+            "content": "[7] Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yutong Dai, Philip S Yu, and Lichao Sun. A comprehensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt. arXiv preprint arXiv:2303.04226, 2023. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.565,
+                0.483,
+                0.647
+            ],
+            "angle": 0,
+            "content": "[8] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21), pages 2633-2650, 2021. 3, 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.649,
+                0.483,
+                0.717
+            ],
+            "angle": 0,
+            "content": "[9] Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and Florian Tramer. Membership inference attacks from first principles. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1897-1914. IEEE, 2022. 1, 2, 3, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.719,
+                0.483,
+                0.774
+            ],
+            "angle": 0,
+            "content": "[10] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. arXiv preprint arXiv:2202.07646, 2022. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.776,
+                0.483,
+                0.844
+            ],
+            "angle": 0,
+            "content": "[11] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.846,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[12] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.094,
+                0.115,
+                0.484,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.093,
+                0.905,
+                0.133
+            ],
+            "angle": 0,
+            "content": "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.136,
+                0.905,
+                0.19
+            ],
+            "angle": 0,
+            "content": "[14] Jinhao Duan, Fei Kong, Shiqi Wang, Xiaoshuang Shi, and Kaidi Xu. Are diffusion models vulnerable to membership inference attacks? arXiv preprint arXiv:2302.01316, 2023. 1, 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.193,
+                0.905,
+                0.261
+            ],
+            "angle": 0,
+            "content": "[15] Michael Duan, Anshuman Suri, Niloofar Mireshghallah, Sewon Min, Weijia Shi, Luke Zettlemoyer, Yulia Tsvetkov, Yejin Choi, David Evans, and Hannaneh Hajishirzi. Do membership inference attacks work on large language models? arXiv preprint arXiv:2402.07841, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.264,
+                0.905,
+                0.317
+            ],
+            "angle": 0,
+            "content": "[16] Wenjie Fu, Huandong Wang, Chen Gao, Guanghua Liu, Yong Li, and Tao Jiang. Mia-tuner: Adapting large language models as pre-training text detector. arXiv preprint arXiv:2408.08661, 2024.3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.321,
+                0.905,
+                0.375
+            ],
+            "angle": 0,
+            "content": "[17] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International workshop ontoImage, 2006. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.378,
+                0.905,
+                0.418
+            ],
+            "angle": 0,
+            "content": "[18] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.421,
+                0.905,
+                0.46
+            ],
+            "angle": 0,
+            "content": "[19] Hailong Hu and Jun Pang. Membership inference of diffusion models. arXiv preprint arXiv:2301.09956, 2023. 1, 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.464,
+                0.904,
+                0.517
+            ],
+            "angle": 0,
+            "content": "[20] Hongsheng Hu, Zoran Salcic, Lichao Sun, Gillian Dobbie, Philip S Yu, and Xuyun Zhang. Membership inference attacks on machine learning: A survey. ACM Computing Surveys (CSUR), 54(11s):1-37, 2022. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.52,
+                0.905,
+                0.574
+            ],
+            "angle": 0,
+            "content": "[21] Pingyi Hu, Zihan Wang, Ruoxi Sun, Hu Wang, and Minhui Xue. M'4 i: Multi-modal models membership inference. Advances in Neural Information Processing Systems, 35:1867-1882, 2022. 1, 2, 3, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.577,
+                0.905,
+                0.646
+            ],
+            "angle": 0,
+            "content": "[22] Myeongseob Ko, Ming Jin, Chenguang Wang, and Ruoxi Jia. Practical membership inference attacks against large-scale multi-modal models: A pilot study. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4871-4881, 2023. 1, 2, 3, 4, 5, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.649,
+                0.905,
+                0.715
+            ],
+            "angle": 0,
+            "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 1, 2, 3, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.718,
+                0.905,
+                0.786
+            ],
+            "angle": 0,
+            "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.789,
+                0.905,
+                0.83
+            ],
+            "angle": 0,
+            "content": "[25] Songze Li, Ruoxi Cheng, and Xiaojun Jia. Identity inference from clip models using only textual data. arXiv preprint arXiv:2405.14517, 2024. 1, 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.832,
+                0.905,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[26] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. 1"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.517,
+                0.093,
+                0.905,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "9218"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.092,
+                0.482,
+                0.176
+            ],
+            "angle": 0,
+            "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.178,
+                0.482,
+                0.233
+            ],
+            "angle": 0,
+            "content": "[28] Bo Liu, Ming Ding, Sina Shaham, Wenny Rahayu, Farhad Farokhi, and Zihuai Lin. When machine learning meets privacy: A survey and outlook. ACM Computing Surveys (CSUR), 54(2):1-36, 2021. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.234,
+                0.482,
+                0.276
+            ],
+            "angle": 0,
+            "content": "[29] Pratyush Maini, Hengrui Jia, Nicolas Papernot, and Adam Dziedzic. Lm dataset inference: Did you train on my dataset? arXiv preprint arXiv:2406.06443, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.278,
+                0.482,
+                0.319
+            ],
+            "angle": 0,
+            "content": "[30] Tomoya Matsumoto, Takayuki Miura, and Naoto Yanai. Membership inference attacks against diffusion models. arXiv preprint arXiv:2302.03262, 2023. 1, 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.321,
+                0.482,
+                0.389
+            ],
+            "angle": 0,
+            "content": "[31] Justus Mattern, Fatemehsadat Mireshghallah, Zhijing Jin, Bernhard Scholkopf, Mrinmaya Sachan, and Taylor Berg-Kirkpatrick. Membership inference attacks against language models via neighbourhood comparison. arXiv preprint arXiv:2305.18462, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.391,
+                0.482,
+                0.433
+            ],
+            "angle": 0,
+            "content": "[32] Franz Josef Och and Hermann Ney. The alignment template approach to statistical machine translation. Computational linguistics, 30(4):417-449, 2004. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.435,
+                0.482,
+                0.476
+            ],
+            "angle": 0,
+            "content": "[33] Yan Pang and Tianhao Wang. Black-box membership inference attacks against fine-tuned diffusion models. arXiv preprint arXiv:2312.08207, 2023. 1, 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.478,
+                0.482,
+                0.533
+            ],
+            "angle": 0,
+            "content": "[34] Marco Pedersoli, Thomas Lucas, Cordelia Schmid, and Jakob Verbeek. Areas of attention for image captioning. In Proceedings of the IEEE international conference on computer vision, pages 1242-1250, 2017. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.535,
+                0.482,
+                0.617
+            ],
+            "angle": 0,
+            "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.619,
+                0.482,
+                0.689
+            ],
+            "angle": 0,
+            "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.69,
+                0.482,
+                0.746
+            ],
+            "angle": 0,
+            "content": "[37] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP), pages 3-18. IEEE, 2017. 1, 2, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.747,
+                0.482,
+                0.829
+            ],
+            "angle": 0,
+            "content": "[38] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 742–758. Springer, 2020. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.831,
+                0.482,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[39] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1, 2"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.093,
+                0.092,
+                0.482,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.516,
+                0.092,
+                0.906,
+                0.147
+            ],
+            "angle": 0,
+            "content": "[40] Stacey Truex, Ling Liu, Mehmet Emre Gursoy, Lei Yu, and Wenqi Wei. Demystifying membership inference attacks in machine learning as a service. IEEE transactions on services computing, 14(6):2073-2089, 2019. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.516,
+                0.149,
+                0.906,
+                0.19
+            ],
+            "angle": 0,
+            "content": "[41] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.516,
+                0.192,
+                0.906,
+                0.248
+            ],
+            "angle": 0,
+            "content": "[42] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.516,
+                0.249,
+                0.906,
+                0.317
+            ],
+            "angle": 0,
+            "content": "[43] Yuxin Wen, Arpit Bansal, Hamid Kazemi, Eitan Borgnia, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Canary in a coalmine: Better membership inference with ensembled adversarial queries. arXiv preprint arXiv:2210.10750, 2022. 1, 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.516,
+                0.319,
+                0.906,
+                0.373
+            ],
+            "angle": 0,
+            "content": "[44] Yixin Wu, Ning Yu, Zheng Li, Michael Backes, and Yang Zhang. Membership inference attacks against text-to-image generation models. arXiv preprint arXiv:2210.00968, 2022. 1, 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.516,
+                0.375,
+                0.906,
+                0.443
+            ],
+            "angle": 0,
+            "content": "[45] Chugui Xu, Ju Ren, Deyu Zhang, Yaoxue Zhang, Zhan Qin, and Kui Ren. Ganobfuscator: Mitigating information leakage under gan via differential privacy. IEEE Transactions on Information Forensics and Security, 14(9):2358-2371, 2019. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.516,
+                0.446,
+                0.906,
+                0.514
+            ],
+            "angle": 0,
+            "content": "[46] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.516,
+                0.517,
+                0.906,
+                0.584
+            ],
+            "angle": 0,
+            "content": "[47] Samuel Yeom, Irene Giacomelli, Matt Fredrikson, and Somesh Jha. Privacy risk in machine learning: Analyzing the connection to overfitting. In 2018 IEEE 31st computer security foundations symposium (CSF), pages 268-282. IEEE, 2018. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.516,
+                0.587,
+                0.906,
+                0.655
+            ],
+            "angle": 0,
+            "content": "[48] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.516,
+                0.657,
+                0.906,
+                0.738
+            ],
+            "angle": 0,
+            "content": "[49] Minxing Zhang, Ning Yu, Rui Wen, Michael Backes, and Yang Zhang. Generated distributions are all you need for membership inference attacks against generative models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4839-4849, 2024. 1, 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.516,
+                0.741,
+                0.906,
+                0.811
+            ],
+            "angle": 0,
+            "content": "[50] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Houdong Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.516,
+                0.092,
+                0.906,
+                0.811
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.946,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "9219"
+        }
+    ]
+]

2025/Variance-Based Membership Inference Attacks Against Large-Scale Image Captioning Models/26bae7f8-8e37-4862-840e-6578a63d1d11_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19da55a883e7613a4a97f9ae31a511d204539aca97d51edcacb342a065d561d1
+size 484693

2025/Variance-Based Membership Inference Attacks Against Large-Scale Image Captioning Models/full.md

@@ -0,0 +1,290 @@
+# Variance-Based Membership Inference Attacks Against Large-Scale Image Captioning Models
+
+Daniel Samira Edan Habler Yuval Elovici Asaf Shabtai Ben-Gurion University of the Negev, Israel {samirada, habler}@post.bgu.ac.il, {elovici, shabaia}@bgu.ac.il
+
+# Abstract
+
+The proliferation of multi-modal generative models has introduced new privacy and security challenges, especially due to the risks of memorization and unintentional disclosure of sensitive information. This paper focuses on the vulnerability of multi-modal image captioning models to membership inference attacks (MIAs). These models, which synthesize textual descriptions from visual content, could inadvertently reveal personal or proprietary data embedded in their training datasets. We explore the feasibility of MIAs in the context of such models. Specifically, our approach leverages a variance-based strategy tailored for image captioning models, utilizing only image data without knowing the corresponding caption. We introduce the means-of-variance threshold attack (MVTA) and confidence-based weakly supervised attack (C-WSA) based on the metric, means-of-variance (MV), to assess variability among vector embeddings. Our experiments demonstrate that these models are susceptible to MIAs, indicating substantial privacy risks. The effectiveness of our methods is validated through rigorous evaluations on these real-world models, confirming the practical implications of our findings.
+
+# 1. Introduction
+
+The advancement from traditional machine learning (ML) to advanced generative models has progressed significantly in recent years [7]. An example of this development is seen in the field of multi-modal models, which are able to process information from various modalities [4]. These models have become increasingly sophisticated and are employed in challenging tasks involving various types of data (e.g., text, images, video). In this paper we focus on image captioning models [2, 3, 23, 24, 26, 34, 50], a multi-modal model used to generate textual descriptions for images.
+
+While multi-modal models have fostered the development of a variety of new multi-modal applications, they also present new privacy and security challenges [6, 28]. A crit
+
+ical concern is the potential of generative models, particularly those trained on large private datasets, to inadvertently memorize and expose sensitive information. Such risks can lead to issues related to intellectual property misuse and privacy breaches involving data from proprietary images or documents [5, 45]. Among the various threats are membership inference attacks (MIAs), which aim to determine whether specific data records were included in a model's training dataset [20]. The ability to infer membership can result in the exposure of personal health information, financial details, or undisclosed affiliations, thereby posing a threat to individual privacy and security.
+
+Initially, studies on MIAs focused on classification models, revealing heightened privacy risks primarily due to model overfitting [9, 20, 47]. Subsequently, attention turned to generative models, where research highlighted their tendency to memorize training data, raising concerns about digital forgery and the loss of originality [10, 11, 39]. In MIAs, as defined by Shokri et al. [37], shadow models are used to mimic the target model's behavior. However, crafting effective shadow models requires extensive datasets that mirror the original training data, along with substantial computational resources. Additionally, in many MIAs [14, 19, 21, 30, 33, 33, 43, 44, 49], the shadow models are scaled to hundreds of samples, and must share similar architectures with the target model and use identical training algorithms. Their dependence on shadow model training limits their applicability for large-scale models [21, 43].
+
+To the best of our knowledge, the study performed by Hu et al. [21] is the only research exploring MIAs on image captioning models. The authors introduced two attack methods, which shadow models are employed. Furthermore, they did not directly evaluate the methods on operational image captioning models, such as those from Hugging Face. Instead, they were assessed in a simulated environment, using artificial models trained on synthetic configurations, small datasets, and numerous training epochs.
+
+Recent studies have proposed MIAs that do not rely on shadow models [22, 25]. Specifically, Ko et al. [22] introduced the cosine similarity (CS) attack (CSA) and
+
+the weakly supervised attack (WSA). These attacks target large-scale multi-modal contrastive language-image pretraining (CLIP) models [35] that utilize text-image data pairs. Both methods use a CS metric to measure the similarity between the embedding vectors of images and text from the target CLIP model. The second method, WSA, combines ground-truth data about non-members with a pseudomember set, which consists of samples flagged as likely members based on their CS values. A classifier is trained on the dataset to distinguish members from non-members.
+
+In this paper, we introduce two MIAs targeting image captioning models: the means-of-variance (MV) threshold attack (MVTA) and the confidence-based weakly supervised attack (C-WSA). In contrast to the techniques of Ko et al. [22], we tailor our methods specifically for image captioning models and enhance them with a new metric, the MV metric, which measures the variance between vectors. Our methods use only images as input, operating under the realistic assumption that an adversary lacks access to ground-truth reference text (i.e., captions) and instead uses captions generated by the target image captioning model. Additionally, we integrate a confidence threshold training method that leverages the pseudo-dataset's confidence to improve the classifier's performance.
+
+We evaluate the proposed methods under varying levels of attacker knowledge (zero and partial) in a black-box scenario. Additionally, we conduct the evaluation using public models trained on large-scale datasets to simulate a realistic setting. The evaluation results show that our proposed methods identify membership information in multimodal image captioning models, using BLIP [23], ViT-GPT2 [13], and GIT [42], trained on large-scale datasets. Our methods enhance detection capabilities, achieving approximately  $20\%$  to  $30\%$  improvements over current methods. To summarize, our main contributions are as follows:
+
+- New MIAs targeting image captioning models: We introduce the MVTA and C-WSA for the image captioning domain, two new attacks in which the input consists solely of images.  
+- The means-of-variance metric: We introduce a quantitative metric designed to assess the variance between vectors in a multi-dimensional space. While we used it for texts, this metric is general and can be used in various fields that utilize embedding vectors.  
+- Innovative training of the pseudo-member set: We introduce a new step in the process of training the attack model, a confidence-based training step that utilizes a pseudo-member set.  
+- Real-world application: By evaluating our methods on Hugging Face models, we demonstrate the practicality and effectiveness of our methods in real-world scenarios.
+
+# 2. Related Work
+
+Previous studies on MIAs exploited the model's tendency to produce more confident predictions for data included in the training set compared to unseen data [9, 20, 37, 40, 47], which is also referred to as memorization, since they tend to memorize the training data. MIAs are more challenging in the context of generative models, where confidence scores or prediction labels are typically unavailable [10, 11]. In addition, it was shown that the time it takes to train a generative model and the size of the dataset used to train the model influence the level of memorization, where extended training epochs and smaller datasets typically lead to increased memorization, enhancing the success of attacks [39].
+
+Recent studies have demonstrated diffusion models' vulnerability to MIAs [12, 46]. In white-box settings, techniques involving statistical and neural network-based inference have shown high effectiveness, especially in models trained with small datasets and a large number of training epochs [14, 19, 30]. In black-box scenarios, methods leveraging semantic fidelity and novel classifier-based approaches have achieved high accuracy, though they often depend on access to non-training data and known training sets, which can limit their broader applicability [33, 44, 49].
+
+Attacks on multi-modal models. The CLIP model [35], which is one of the many types of multi-model models, utilizes contrastive learning to align text captions with images, leveraging unsupervised learning to interpret and generate image content based on natural language descriptions. Traditionally, MIAs targeting models like CLIP have relied on computationally intensive shadow models [14, 19, 21, 30, 33, 33, 43, 44, 49]. However, recent research has proposed efficient, shadow-free approaches for MIAs against multimodal models. The textual unimodal detector [25] identifies the use of personally identifiable information (PII) in CLIP models using only textual data, reducing both computational costs and privacy risks without the need for images. Another method, the WSA [22], addresses practical MIAs against large-scale multi-modal CLIP models by combining ground-truth information about non-members with noisy member data from a weak attack to develop an attack model. The WSA starts by querying the target model for image and text features of each non-member and evaluates the CS between these features. The CS distribution, typically Gaussian, helps approximate the mean and variance and is used to flag samples with a CS significantly above the mean as likely members. This identified set, known as the pseudo-member set, is then used to construct an attack dataset for training an attack model that determines the membership status of any target sample based on its features. Notably, the attack model does not account for the possibility that labels marked as true during training might be mislabeled, which could potentially improve performance if some of these were excluded from training.
+
+Table 1. An overview of membership inference research on generation models. "Reference Output" indicates the attacker has access to training text references.  
+
+<table><tr><td>Target Model Tasks</td><td>Attacks</td><td>Using Shadow Model</td><td>Final Model Prediction</td><td>Evaluated on Realistic Models</td><td>Other Information</td></tr><tr><td>Diffusion Models</td><td>[14, 19, 30, 33, 44, 49]</td><td>✓ or -</td><td>✓</td><td>✓ or -</td><td>Reference Output or -</td></tr><tr><td>LLMs</td><td>[8, 15, 16, 29, 31]</td><td>✓ or -</td><td>✓ or -</td><td>✓</td><td>-</td></tr><tr><td>Multi-Modal Models - CLIP</td><td>[22, 25]</td><td>-</td><td>✓</td><td>✓</td><td>-</td></tr><tr><td>Multi-Modal Models - Image Captioning</td><td>M4I [21]</td><td>✓</td><td>✓</td><td>-</td><td>Reference Output or -</td></tr><tr><td rowspan="2">Ours</td><td>Means-of-Variance Threshold Attack</td><td>-</td><td>✓</td><td>✓</td><td>-</td></tr><tr><td>Confidence-Based Weakly Supervised Attack</td><td>-</td><td>✓</td><td>✓</td><td>-</td></tr></table>
+
+"✓" - the knowledge is necessary for the attack, "-" the knowledge is unnecessary for the attack
+
+Another type of multi-modal model, BLIP [23, 24], advances vision-language tasks by adapting to both understanding and generation challenges, including image-text retrieval, image captioning, and visual question answering. To the best of our knowledge, the work of Hu et al. [21] is the only one that explores MIAs on image captioning models. Their paper introduces two attack methods: metric-based and feature-based. The metric-based approach uses similarity metrics to infer membership status by comparing the similarity of model outputs with ground-truth data. The feature-based approach utilizes a pre-trained multi-modal feature extractor to analyze similarities between the input and output features. These attacks were not applied directly to operational image captioning models but rather in a controlled setting with synthetic configurations, in terms of small dataset size and a large number of training epochs, using an encoder-decoder framework to simulate the process. In addition, the authors also employed shadow models, which increases the complexity and computational costs.
+
+Unlike previous works, we focuses on the realm of multimodal image captioning models, an area that has received little research attention in the context of MIAs. In this research, we avoid the conventional reliance on computationally intensive shadow models, which necessitate datasets that closely mirror the training data of the target model. Instead, we base our attack methodologies on metrics designed to detect discrepancies in data variance. By applying MIAs in scenarios that only involve image data without access to corresponding textual information, our proposed MIAs have a broader range of applications.
+
+An overview of membership inference research across various generation model tasks is provided in Table 1.
+
+# 3. Threat Model
+
+The primary goal of the attacker is to determine whether a given sample is a member of the target image captioning model's training set or not. The attacker can query the model, submitting images and receiving the corresponding generated text captions in response, but does not have any access to the model itself (i.e., query-only). The attacker does not have any knowledge of the target model's archi
+
+tectural, parameters, or the specific training algorithm used.
+
+We emphasize that the attacker's dataset consists solely of images, allowing the attack to remain effective even in scenarios where access to the model's training captions or other textual references is limited or unavailable. We explore two different settings concerning the attacker's knowledge of the data used to train the target model:
+
+- Zero-knowledge setting: The attacker has no prior knowledge of the distribution of the data used to train the target model, making this approach challenging but realistic in restricted data access scenarios.  
+- Partial-knowledge setting: Following prior research [9, 21, 22, 33, 43], the attacker has access to a publicly available dataset of images,  $D_{all}$ , approximating the data distribution used to train the target model. The attacker also has access to a non-member dataset  $D_{no}$ , while the membership status of  $D_{all}$  is unknown. The non-member dataset can be collected from content published after the model's release or synthesized by image-generating AI models, such as publicly available diffusion models.
+
+# 4. Methodology
+
+Our methodology is designed to leverage our intuition regarding the use of variance in the outputs of a target image captioning model as a means of inferring infer membership information. The pipeline of our methodology consists of three main stages: text generation and multi-modal feature encoding, variance metric computation, and application of proposed attacks. First, the target model generates text captions from images. These captions and images are then encoded into feature embeddings, using pre-trained text and image encoders, respectively. The outputs are then analyzed to determine membership, aiming to ascertain whether a new image was part of the target model's training dataset. Next, we compute a variance-based metric, the MV metric, to quantify the variance among text embeddings generated for each image. This metric serves as the basis for identifying membership by capturing the relationship between variance and membership. Finally, we propose two attacks based on the MV metric: the MVTA and C-WSA. The MVTA operates in a zero-knowledge setting by exploit
+
+ing the reduced variance of text embeddings for training set images. The C-WSA, designed for a partial-knowledge setting, combines non-member ground truth with probabilistic pseudo-member labels and employs a confidence threshold to train an attack classifier model. The stages of our methodology are described below.
+
+# 4.1. Text Generation and Feature Encoding
+
+This stage involves two steps: text generation and multimodal feature encoding, as shown in Figure 1. Typically, image captioning models use two text generation decoding strategies to generate concise descriptions: beam search [32] and top-  $p$  sampling [18]. Beam search is deterministic, consistently generating the same outputs for a given input by keeping the best candidates when choosing the next word. In contrast, top-  $p$  sampling, also known as nucleus sampling, is non-deterministic and introduces diversity by selecting from a probabilistic "nucleus" of options. Carlini et al. [8] showed that when trained on paired data of the prompt and image, an image generation model generates similar images from the same prompt for two different random seeds. Following this principle, we denote the caption generated for an image  $x$  using top-  $p$  sampling by the model Gen as  $\mathrm{Gen}(x)$ . Then, if multiple captions generated for the same image  $x$  are highly similar under a distance measure  $d$  ( $\mathrm{Gen}(x) \approx_d \mathrm{Gen}(x)$ ), it suggests that the model has memorized the image. Based on this, we assume that the variance among the  $n$  captions generated for an image will be lower if the target model has been trained on that image.
+
+During the feature encoding step, features are derived using a pretrained text encoder  $f_{text}$  to obtain text embeddings from the generated captions, and a pre-trained image encoder  $f_{img}$  to extract visual features from images. We note that the widespread availability of large-scale pre-trained models allows adversaries to easily obtain both encoders.
+
+![](images/11bc1b35e06c2bc2051cd015627ef0c9aa01a00665cd16de8d8ce9b00c79b17c.jpg)  
+Figure 1. Overview of our method's text generation and multimodal feature encoding stage. This stage consists of two steps: text generation, in which captions are generated using the target model, and feature encoding, in which features from multiple modalities (text and image) are encoded using pre-trained models.
+
+# 4.2. Means-of-Variance Metric
+
+Based on our intuition, we introduce the new MV metric, which is designed to quantify the variance between embedding vectors. This metric constructs a variance vector of size  $m$ , where each index represents the variance of a specific feature calculated across  $n$  vector embeddings. Once the variance vector is constructed, the metric then computes the mean of the values within it. The resulting single value represents the overall variance among the vectors, providing a concise measure of dispersion. Figure 2 provides an illustration of the MV metric calculation process, demonstrating how the MV metric is computed.
+
+![](images/00ba8d3249090bf9edb1b2235a30344fcb60cfdc3b020da82095f1134bd91960.jpg)  
+Figure 2. MV metric calculation. The feature matrix  $(f)$  consists of  $n$  vectors with  $m$  features. The variance of each feature across these vectors is computed to form the variance vector  $(v)$ , whose mean gives the MV score.
+
+# 4.3. Proposed Attacks
+
+Means-of-variance threshold attack. In a zero-knowledge setting, we compute the MV metric for the set of the embedding texts generated from each image, quantifying the average variance among these captions. If the calculated  $MV$  metric for image  $x$  falls below a predetermined threshold  $\tau$ ,  $x$  is classified as a member of the training set:
+
+$$
+\text {P r e d i c t e d M e m b e r s h i p} = \left\{ \begin{array}{l l} \text {M e m b e r} & \text {i f} M V <   \tau \\ \text {N o n - M e m b e r} & \text {o t h e r w i s e} \end{array} \right. \tag {1}
+$$
+
+The threshold  $\tau$  is empirically determined by analyzing the distribution of MV scores for both members and non-members. In practice, MVTA serves as a baseline for C-WSA, with  $\tau$  calculated using the mean and standard deviation of MV scores from non-member datasets  $(D_{no})$ .
+
+This approach adapts confidence or loss-based thresholds for data membership to multi-modal image captioning.
+
+Confidence-based weakly supervised attack. Building on the WSA method [22], our second attack, the C-WSA, introduces an innovative adaptation in the partial-knowledge setting by utilizing the MV metric to enhance the model's ability to distinguish between member and non-member data. An overview of the attack is provided in Figure 3. Similar to a typical WSA, our attack integrates ground-truth information about non-members, along with probabilistic assessments of membership.
+
+![](images/37ae34c0425bbf0ebadb99588c10e2d8a4a8bbb6edc8803e44bab35fca2e006e.jpg)  
+Figure 3. Overview of the C-WSA.
+
+In our modified attack pipeline, we begin by querying the target model with images from the known non-member dataset  $(D_{no})$  to generate  $n$  captions per image. Given the availability of public or synthetic datasets, an adversary can readily access or create them. Then, we calculate the MV metric for the dataset in order to establish a baseline distribution of variance scores among non-members, which we hypothesized to follow a Gaussian distribution. The MV scores for non-members are denoted as follows:
+
+$$
+M V _ {n o} = \left\{M V \left(f _ {t e x t} (G e n (x))\right) \mid x \in D _ {n o} \right\} \tag {2}
+$$
+
+where  $f_{text}(Gen(x))$  refers to the text embeddings extracted by the text encoder  $f_{text}$  from the captions generated by the target model for each image  $x$ , and  $MV$  is the function that performs the metric on the vectors.
+
+Using  $MV_{no}$ , we calculate the mean  $\mu_{no}$  and standard deviation  $\sigma_{no}$  of the MV scores:
+
+$$
+\mu_ {n o} = \operatorname {m e a n} \left(M V _ {n o}\right), \quad \sigma_ {n o} = \operatorname {s t d} \left(M V _ {n o}\right) \tag {3}
+$$
+
+These parameters  $(\mu_{no}$  and  $\sigma_{no})$  are crucial for identifying potential members by comparing their MV scores to the non-member distribution baseline mentioned above. Intuitively, if a sample's  $MV$  is substantially lower than  $\mu_{no}$ , specifically by more than  $\lambda \sigma_{no}$  for a given constant  $\lambda$ , it suggests that the sample is likely a member.
+
+For the dataset  $D_{all}$ , which comprises potential members, we compute the MV scores using the text embeddings generated by the target model. Samples in  $D_{all}$  whose MV scores significantly deviate from this non-member baseline are identified as potential members. Specifically, if a sample's MV score exceeds a defined threshold calculated as  $\mu_{no} - \lambda \sigma_{no}$ , it is classified as a pseudo-member. The corresponding subset of these identified samples ( $D_{p-mem}$ ), serves as our pseudo-member set. This set comprises a mix of likely members and potential false positives.
+
+The formulas associated with this process are as follows:
+
+$$
+\tau = \mu_ {n o} - \lambda \sigma_ {n o} \tag {4}
+$$
+
+$$
+D _ {p - m e m} = \{x \mid M V (f _ {\text {t e x t}} (\operatorname {G e n} (x))) <   \tau , x \in D _ {\text {a l l}} \} \tag {5}
+$$
+
+With the  $D_{p - mem}$  and  $D_{no}$  in place, we proceed to create the attack training set  $D_{attack}$ , comprising image features from  $D_{no} \cup D_{p - mem}$  with their associated binary labels:
+
+$$
+D _ {\text {a t t a c k}} = \left\{f _ {i m g} (x) \mid x \in D _ {n o} \cup D _ {p - m e m}, b = 1 [ x \in D _ {p - m e m} ] \right\} \tag {6}
+$$
+
+where  $f_{img}(x)$  represents the image features extracted by the image encoder  $f_{img}$  for each image  $x$ , and  $b$  is a binary indicator with 1 representing a potential member and 0 representing a non-member. Finally, using  $D_{all}$ , we train the attack classifier model  $f_{attack}$  so that it is capable of determining the membership status of any target sample by evaluating the sample's image features.
+
+In  $D_{all}$ , samples labeled with 0 are definitively non-members, whereas those labeled with 1 may not actually be members. This distinction guides our training strategy for the attack classifier model  $f_{attack}$ , which incorporates a confidence threshold. Initially, the attack model undergoes a "warm-up" phase, training without interruptions to establish baseline prediction tendencies. This phase lasts for a predefined ratio of epochs, referred to as "warm-epoch." Then, before each training batch, we evaluate the membership confidence of the samples. For those identified as pseudo-members ( $b = 1$ ), training continues only if their membership confidence exceeds a predetermined threshold. This approach refines the training process, focusing primarily on samples most likely to be members, thus enhancing the accuracy and effectiveness of our attack classifier.
+
+# 5. Evaluation
+
+# 5.1. Evaluation Setup
+
+Datasets. To evaluate our attacks, we utilize a composite of datasets including MSCOCO [27], FLICKR30k [48], Textcaps [38], IAPR TC-12 [17], and Nocaps [1] as  $D_{all}$ . These datasets contain approximately 113K, 30K 28K, 20K, and 15K images, respectively. Additionally, we generate artificial dataset images using Stable Diffusion v1.5 [36] based on the captions of FLICKR30k to simulate the creation of a new non-member dataset for the target model. In practice, the target models are trained on one dataset, and we sample from the combination of the other datasets to establish ground-truth non-members. We then further divide these non-members into two disjoint subsets: one provides the ground-truth non-member data as  $D_{no}$ , and another is utilized to assess the attack's performance.
+
+In the evaluation phase, both member and non-member data are split into training and testing sets. The training set is used as  $D_{all}$ , while the testing set is used to evaluate the attack model classifier. Our approach deviates from the traditional MIA evaluation setup, which typically involves splitting a single dataset into member and non-member subsets for small-scale classification models [21, 37]. Instead, similar to Ko et al. [22], we opted for a configuration that offers greater real-world relevance and practical impact.
+
+Target models. We evaluate our attacks using three prominent public image captioning models from Hugging Face: BLIP Large [23], ViT-GPT2 [13], and GIT Base [42], all trained on large-scale datasets.
+
+Evaluation metrics. We employ several key metrics to assess how effective our methods are: area under the curve (AUC), accuracy (ACC), and true positive rate (TPR) at a low false positive rate (FPR), which was used by [9].
+
+# 5.2. Evaluation Results
+
+Attack performance. We evaluate the performance of our proposed attacks, MVTA and C-WSA. To perform a fair comparison with CSA and WSA [22], which utilize imagetext data pairs, we employ captions generated by the target models, producing a single caption per image using the top- $p$  sampling method, aligning with the approach taken in our methods to evaluate performance. We pair CSA with MVTA in a zero-knowledge setting and WSA with C-WSA in a partial-knowledge setting, enabling each method to be evaluated under equivalent data access conditions.
+
+Table 2 presents the evaluation results, showing the relative effectiveness of each attack method across the various metrics: AUC, ACC, and TPR at an FPR of  $1\%$ . Overall, our proposed methods, MVTA and C-WSA, outperform the baseline CSA and WSA attacks across these key metrics, highlighting their robustness and precision in distinguishing member from non-member data.
+
+1. Performance in zero-knowledge setting: MVTA consistently outperforms CSA across all evaluated models. For instance, MVTA achieves AUC scores of 0.66 and 0.77 on BLIP and ViT-GPT2, respectively, compared to CSA's 0.54 and 0.72. This improvement demonstrates the effectiveness of MVTA's MV metric in capturing subtle distinctions in membership status even when data knowledge is minimal.
+
+2. Performance in partial-knowledge setting: C-WSA shows superior performance over WSA, especially on high-precision metrics such as TPR at an FPR of  $1\%$ . C-WSA achieves an AUC of 0.91 on BLIP and 0.93 on ViT-GPT2, surpassing WSA's scores of 0.84 and 0.89, respectively. Additionally, C-WSA's TPR at an FPR of  $1\%$  values are 0.38 for BLIP and 0.43 for ViT-GPT2, which indicates a substantial reduction in the number of false positives, highlighting C-WSA's accuracy in identifying true member instances.
+
+MVTA analysis. In this analysis, we focus on the MVTA and its use of the MV metric to distinguish between member and non-member data. Figure 4a presents the distribution of MV scores for both members and non-members. The histogram shows a separation between the two groups, with members tending to have lower MV scores while non-members exhibit higher values. We can observe that as the
+
+threshold MV score decreases, the proportion of members increases, which is informative for determining an effective threshold for the attack and its consequences.
+
+To illustrate how the captions from member and non-member images behave in the embedding space, Figure 4b presents features from the captions of one member and one non-member image in a two-dimensional space using t-SNE [41]. We find that the captions of member images are closer together than those of non-member images, indicating more consistent embeddings for members.
+
+![](images/f086fabc44e51918508d487d724aa6d1272fb2ae27e2fb2626d77985e9a51bbd.jpg)  
+(a) Distribution of the MV metric values for members and non-members.
+
+![](images/481cfb853e62f8b4fde4ea818bbd21edfde07ef2288845af7c2c03835267b50a.jpg)  
+(b) Projection of member and non-member features from generated texts onto a two-dimensional embedding space using t-SNE [41].  
+Figure 4. Comparative visualization of the MV metric for members and non-members.
+
+C-WSA analysis. In our evaluation of C-WSA, the parameter  $\lambda$  is crucial for setting the threshold that determines whether a sample is classified as a member or non-member. Specifically, if a sample's MV score is lower than the mean score for non-members  $\mu_{no}$  by more than  $\lambda \sigma_{no}$  (where  $\sigma_{no}$  is the standard deviation of non-member MV scores), it is likely to be classified as a member. To identify the optimal value of  $\lambda$ , we varied  $\lambda$  from  $-2$  to 2 and established a baseline by randomly selecting samples from  $D_{all}$  as members for  $D_{p - mem}$ , independent of  $\lambda$ . As shown in Table 3, increasing  $\lambda$  raises the member ratio in  $D_{p - mem}$ , enhancing the accuracy of the labels for the attack model classifier. However, setting  $\lambda$  too high may exclude relevant member
+
+Table 2. Evaluation of attack performance on different public target models (BLIP, ViT-GPT2, and GIT) across the MSCOCO and TextCaps datasets. We bold the best results for each knowledge setting.  
+
+<table><tr><td colspan="2">Dataset [Model]</td><td colspan="3">MSCOCO [BLIP]</td><td colspan="3">MSCOCO [ViT-GPT2]</td><td colspan="3">MSCOCO [GIT]</td><td colspan="3">Textcaps [GIT]</td></tr><tr><td>Method</td><td>AUC</td><td>ACC</td><td>TPR@FPR=1%</td><td>AUC</td><td>ACC</td><td>TPR@FPR=1%</td><td>AUC</td><td>ACC</td><td>TPR@FPR=1%</td><td>AUC</td><td>ACC</td><td>TPR@FPR=1%</td><td></td></tr><tr><td>CSA</td><td>0.5444</td><td>0.5358</td><td>0.0087</td><td>0.7207</td><td>0.6630</td><td>0.0466</td><td>0.5942</td><td>0.6265</td><td>0.0197</td><td>0.4484</td><td>0.5010</td><td>0.0110</td><td></td></tr><tr><td>MVTA</td><td>0.6601</td><td>0.6183</td><td>0.0224</td><td>0.7727</td><td>0.7026</td><td>0.0690</td><td>0.7630</td><td>0.7011</td><td>0.0441</td><td>0.5026</td><td>0.5124</td><td>0.0083</td><td></td></tr><tr><td>WSA</td><td>0.8398</td><td>0.7914</td><td>0.3144</td><td>0.8931</td><td>0.8415</td><td>0.3758</td><td>0.8626</td><td>0.8034</td><td>0.3328</td><td>0.8165</td><td>0.7945</td><td>0.3773</td><td></td></tr><tr><td>C-WSA</td><td>0.9131</td><td>0.8345</td><td>0.3826</td><td>0.9335</td><td>0.8619</td><td>0.4251</td><td>0.9255</td><td>0.8520</td><td>0.4075</td><td>0.8967</td><td>0.8479</td><td>0.4152</td><td></td></tr></table>
+
+data, potentially causing the attack model to miss broader patterns.
+
+Our experiments indicate that  $\lambda = 0.5$  achieves the highest average AUC values, with 0.91 for BLIP and 0.93 for ViT-GPT2. This choice not only maximizes the AUC but also achieves a favorable balance across the rest of the metrics, maintaining a TPR at a  $1\%$  FPR.
+
+We also evaluate the impact of adding a confidence threshold to the attack model's training. Figure 5 presents the average probability over the training epochs for C-WSA with a threshold (Figure 5a) and WSA without a threshold (Figure 5b). There are three possible groups of samples with pseudo-labels: members with label 1, non-members with label 1 (mislabeling), and non-members with label 0.
+
+The inclusion of the confidence threshold enables the attack model to better manage the mislabeling group (non-members with label 1), as reflected in the reduced average probability for this group across the training epochs. When the threshold is applied, the attack model assigns lower probabilities to non-members, unlike the setting without the threshold, where the model continues to train on misclassified samples and assigns higher probabilities to them.
+
+However, this improvement comes at the cost of slightly lowering the average probability for true members with label 1. Despite this trade-off, the confidence threshold refines the training process by prioritizing reliable high-confidence samples. Further details on the selection of the confidence threshold and warm-up epoch parameters can be found in Appendix ??
+
+Evaluation on fine-tuned public models. We also validate our approach on public pre-trained models that we fine-tuned. The results of this experiment (detailed in Appendix ??) show that our attacks outperform the baseline methods across most key metrics, further highlighting their effectiveness.
+
+# 5.3. Sensitivity Analysis
+
+Impact of text generation decoding strategies. Figure 6 presents the performance of various attack methods in terms of the AUC metric as a function of the number of captions generated per image  $(n)$  and the text generation decoding strategy used (beam search or top-  $p$  sampling). The beam search results are placed at  $n = 0$ , while values from  $n = 1$  to  $n = 50$  correspond to captions generated using top-  $p$
+
+![](images/c6b85213944d42dc6c21a8b65239f66b3cedeed368f0e3743a50bc473fec3c45.jpg)  
+(a) C-WSA with threshold.
+
+![](images/4ed5791c254b1487139d2cb9aba117270d134965295cc000efa88ecb6d25df2d.jpg)  
+(b) WSA without threshold.  
+Figure 5. Comparison of C-WSA with a confidence threshold and WSA without one, depicting the average probability of  $D_{p - mem}$  across the training epochs. M denotes membership status, and PL represents the pseudo-label assignment.
+
+sampling. Notably, when there is only one caption generated, we perform our evaluation using the CS metric, because the MV metric, which calculates variance, cannot be applied.
+
+We observe that the performance of the attacks improves when transitioning from beam search to top-  $p$  sampling. This suggests that the diversity introduced by top-  $p$  sampling better captures the nuances of the dataset, resulting in more informative features for the attack models.
+
+Additionally, increasing the number of generated captions  $(n)$  positively impacts the performance of the attacks across all methods. This improvement is more pronounced for MVTA than C-WSA. A likely explanation for this difference is the relative simplicity of MVTA, which relies solely on the MV score calculated from the generated texts. In contrast, the greater complexity of C-WSA enables it to compensate for fewer generated captions, leveraging its advanced mechanisms to maintain competitive performance.
+
+Based on the results, we set  $n = 40$  as a balance parameter between the number of captions generated and the resulting performance. This choice reflects a trade-off that achieves high AUC values while limiting the computational overhead associated with generating more captions.
+
+Impact of member ratio in  $D_{all}$ . In this analysis, we evaluate the impact of the member ratio in  $D_{all}$  on the performance of the C-WSA attack. We vary the member ratio
+
+Table 3. Attack performance vs.  $\lambda$  and the corresponding ratio of true members in pseudo-labeled data  $\left( {\lambda  = {0.5}}\right.$  achieves optimal performance.)  
+
+<table><tr><td colspan="2">Dataset [Model]</td><td colspan="4">MSCOCO [BLIP]</td><td colspan="4">MSCOCO [ViT-GPT2]</td></tr><tr><td colspan="2">Method</td><td colspan="4">C-WSA</td><td colspan="4">C-WSA</td></tr><tr><td>λ</td><td>Member Ratio in Dp-mem</td><td>AUC</td><td>TPR@FPR=1%</td><td>ACC</td><td>Member Ratio in Dp-mem</td><td>AUC</td><td>TPR@FPR=1%</td><td>ACC</td><td></td></tr><tr><td>Baseline</td><td>0.33</td><td>0.9041</td><td>0.3655</td><td>0.8359</td><td>0.33</td><td>0.6024</td><td>0.0755</td><td>0.5994</td><td></td></tr><tr><td>-2</td><td>0.34</td><td>0.8758</td><td>0.3408</td><td>0.8214</td><td>0.34</td><td>0.9152</td><td>0.3896</td><td>0.8428</td><td></td></tr><tr><td>-1.5</td><td>0.35</td><td>0.8867</td><td>0.3551</td><td>0.5997</td><td>0.35</td><td>0.9032</td><td>0.3557</td><td>0.8368</td><td></td></tr><tr><td>-1</td><td>0.36</td><td>0.8917</td><td>0.3622</td><td>0.8339</td><td>0.37</td><td>0.9083</td><td>0.4061</td><td>0.8541</td><td></td></tr><tr><td>-0.5</td><td>0.39</td><td>0.9236</td><td>0.4098</td><td>0.8541</td><td>0.41</td><td>0.9165</td><td>0.3665</td><td>0.8435</td><td></td></tr><tr><td>0</td><td>0.42</td><td>0.9166</td><td>0.3782</td><td>0.8484</td><td>0.46</td><td>0.9235</td><td>0.4142</td><td>0.8860</td><td></td></tr><tr><td>0.5</td><td>0.46</td><td>0.9131</td><td>0.3826</td><td>0.8345</td><td>0.53</td><td>0.9304</td><td>0.4106</td><td>0.8588</td><td></td></tr><tr><td>1.0</td><td>0.49</td><td>0.8809</td><td>0.2911</td><td>0.7991</td><td>0.60</td><td>0.9133</td><td>0.3671</td><td>0.8361</td><td></td></tr><tr><td>1.5</td><td>0.52</td><td>0.8040</td><td>0.1880</td><td>0.7335</td><td>0.69</td><td>0.8868</td><td>0.0755</td><td>0.8088</td><td></td></tr><tr><td>2</td><td>0.54</td><td>0.7236</td><td>0.1103</td><td>0.6685</td><td>0.74</td><td>0.8221</td><td>0.2080</td><td>0.7533</td><td></td></tr></table>
+
+![](images/7fff563a0d60d300b9d31a78fc002947c4146d192c8707e7db29638f075f3388.jpg)  
+Figure 6. Comparison of AUC values for captions generated per image  $(n)$ . Beam search  $(bs)$  is used at  $n = 0$ , while top- $p$  sampling is used for  $n > 0$ . C refers to MSCOCO.
+
+between 0.1 and 0.4 and assess the attack performance using two metrics: AUC and TPR at an FPR of  $1\%$ , as shown in the Figure 7. The results show an improvement in attack performance as the member ratio increases. This is evident on both metrics, where higher member ratios lead to better AUC and TPR at a FPR of  $1\%$  values. Intuitively, the performance gains indicate that a higher proportion of members in  $D_{all}$  provides the attack model with more representative samples, enhancing its ability to effectively distinguish members from non-members. Notably, even at a low member ratio of 0.1, the C-WSA attack achieves stable performance, with an AUC of 0.68 for BLIP and 0.84 for ViT-GPT2. For the TPR at a FPR of  $1\%$ , the values are 0.19 for BLIP and 0.26 for ViT-GPT2. These results demonstrate that the attack remains effective even with a limited proportion of members in  $D_{all}$ .
+
+![](images/e33c9072188bd395704203422f23d6468f29c3ed80b07553dd47f2fd519c49dd.jpg)  
+Figure 7. Performance comparison in terms of AUC and TPR@FPR=1% across different member ratios in  $D_{all}$  for C-WSA configurations. C refers to MSCOCO.
+
+![](images/3137770a689ee9eceb6f1007d6d6a24aed209576c1eb42fddfc0ef3863ef3712.jpg)
+
+# 6. Conclusions and Future Work
+
+We introduce two MIAs, MVTA and C-WSA, specifically designed for image captioning models in the zero-knowledge and partial-knowledge settings, respectively. These attacks leverage the MV metric, which serves as the foundation of our approach. The results obtained in our experiments highlight the vulnerability of image captioning models to MIAs and emphasize the privacy implications inherent in these models. Our evaluation showed that the effectiveness of our attacks is influenced by the text generation decoding strategies employed and the number of captions generated. Specifically, the diversity introduced by top- $p$  sampling and the variance in the number of generated captions provide a more informative set of features for MIAs. These factors affect our attacks' ability to distinguish between member and non-member data.
+
+Future research could focus on: (1) developing and evaluating techniques to mitigate the effectiveness of the MVTA and C-WSA; (2) expanding the attacks' applicability to a wider range of models, such as diffusion models, by adapting the MV metric for use on different types of data (this would only require changes to the encoders).
+
+# References
+
+[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8948-8957, 2019. 5  
+[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 1  
+[3] Jyoti Aneja, Aditya Deshpande, and Alexander G Schwing. Convolutional image captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5561-5570, 2018. 1  
+[4] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE transactions on pattern analysis and machine intelligence, 41(2):423-443, 2018. 1  
+[5] Siddharth Bhatia, Arjit Jain, and Bryan Hooi. Exgan: Adversarial generation of extreme samples. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6750-6758, 2021. 1  
+[6] Miles Brundage, Shahar Avin, Jack Clark, Helen Toner, Peter Eckersley, Ben Garfinkel, Allan Dafoe, Paul Scharre, Thomas Zeitzoff, Bobby Filar, et al. The malicious use of artificial intelligence: Forecasting, prevention, and mitigation. arXiv preprint arXiv:1802.07228, 2018. 1  
+[7] Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yutong Dai, Philip S Yu, and Lichao Sun. A comprehensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt. arXiv preprint arXiv:2303.04226, 2023. 1  
+[8] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21), pages 2633-2650, 2021. 3, 4  
+[9] Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and Florian Tramer. Membership inference attacks from first principles. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1897-1914. IEEE, 2022. 1, 2, 3, 6  
+[10] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. arXiv preprint arXiv:2202.07646, 2022. 1, 2  
+[11] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2  
+[12] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2
+
+[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 6  
+[14] Jinhao Duan, Fei Kong, Shiqi Wang, Xiaoshuang Shi, and Kaidi Xu. Are diffusion models vulnerable to membership inference attacks? arXiv preprint arXiv:2302.01316, 2023. 1, 2, 3  
+[15] Michael Duan, Anshuman Suri, Niloofar Mireshghallah, Sewon Min, Weijia Shi, Luke Zettlemoyer, Yulia Tsvetkov, Yejin Choi, David Evans, and Hannaneh Hajishirzi. Do membership inference attacks work on large language models? arXiv preprint arXiv:2402.07841, 2024. 3  
+[16] Wenjie Fu, Huandong Wang, Chen Gao, Guanghua Liu, Yong Li, and Tao Jiang. Mia-tuner: Adapting large language models as pre-training text detector. arXiv preprint arXiv:2408.08661, 2024.3  
+[17] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International workshop ontoImage, 2006. 5  
+[18] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 4  
+[19] Hailong Hu and Jun Pang. Membership inference of diffusion models. arXiv preprint arXiv:2301.09956, 2023. 1, 2, 3  
+[20] Hongsheng Hu, Zoran Salcic, Lichao Sun, Gillian Dobbie, Philip S Yu, and Xuyun Zhang. Membership inference attacks on machine learning: A survey. ACM Computing Surveys (CSUR), 54(11s):1-37, 2022. 1, 2  
+[21] Pingyi Hu, Zihan Wang, Ruoxi Sun, Hu Wang, and Minhui Xue. M'4 i: Multi-modal models membership inference. Advances in Neural Information Processing Systems, 35:1867-1882, 2022. 1, 2, 3, 5  
+[22] Myeongseob Ko, Ming Jin, Chenguang Wang, and Ruoxi Jia. Practical membership inference attacks against large-scale multi-modal models: A pilot study. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4871-4881, 2023. 1, 2, 3, 4, 5, 6  
+[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 1, 2, 3, 6  
+[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 3  
+[25] Songze Li, Ruoxi Cheng, and Xiaojun Jia. Identity inference from clip models using only textual data. arXiv preprint arXiv:2405.14517, 2024. 1, 2, 3  
+[26] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. 1
+
+[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5  
+[28] Bo Liu, Ming Ding, Sina Shaham, Wenny Rahayu, Farhad Farokhi, and Zihuai Lin. When machine learning meets privacy: A survey and outlook. ACM Computing Surveys (CSUR), 54(2):1-36, 2021. 1  
+[29] Pratyush Maini, Hengrui Jia, Nicolas Papernot, and Adam Dziedzic. Lm dataset inference: Did you train on my dataset? arXiv preprint arXiv:2406.06443, 2024. 3  
+[30] Tomoya Matsumoto, Takayuki Miura, and Naoto Yanai. Membership inference attacks against diffusion models. arXiv preprint arXiv:2302.03262, 2023. 1, 2, 3  
+[31] Justus Mattern, Fatemehsadat Mireshghallah, Zhijing Jin, Bernhard Scholkopf, Mrinmaya Sachan, and Taylor Berg-Kirkpatrick. Membership inference attacks against language models via neighbourhood comparison. arXiv preprint arXiv:2305.18462, 2023. 3  
+[32] Franz Josef Och and Hermann Ney. The alignment template approach to statistical machine translation. Computational linguistics, 30(4):417-449, 2004. 4  
+[33] Yan Pang and Tianhao Wang. Black-box membership inference attacks against fine-tuned diffusion models. arXiv preprint arXiv:2312.08207, 2023. 1, 2, 3  
+[34] Marco Pedersoli, Thomas Lucas, Cordelia Schmid, and Jakob Verbeek. Areas of attention for image captioning. In Proceedings of the IEEE international conference on computer vision, pages 1242-1250, 2017. 1  
+[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2  
+[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5  
+[37] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP), pages 3-18. IEEE, 2017. 1, 2, 5  
+[38] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 742–758. Springer, 2020. 5  
+[39] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1, 2
+
+[40] Stacey Truex, Ling Liu, Mehmet Emre Gursoy, Lei Yu, and Wenqi Wei. Demystifying membership inference attacks in machine learning as a service. IEEE transactions on services computing, 14(6):2073-2089, 2019. 2  
+[41] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 6  
+[42] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2, 6  
+[43] Yuxin Wen, Arpit Bansal, Hamid Kazemi, Eitan Borgnia, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Canary in a coalmine: Better membership inference with ensembled adversarial queries. arXiv preprint arXiv:2210.10750, 2022. 1, 2, 3  
+[44] Yixin Wu, Ning Yu, Zheng Li, Michael Backes, and Yang Zhang. Membership inference attacks against text-to-image generation models. arXiv preprint arXiv:2210.00968, 2022. 1, 2, 3  
+[45] Chugui Xu, Ju Ren, Deyu Zhang, Yaoxue Zhang, Zhan Qin, and Kui Ren. Ganobfuscator: Mitigating information leakage under gan via differential privacy. IEEE Transactions on Information Forensics and Security, 14(9):2358-2371, 2019. 1  
+[46] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 2  
+[47] Samuel Yeom, Irene Giacomelli, Matt Fredrikson, and Somesh Jha. Privacy risk in machine learning: Analyzing the connection to overfitting. In 2018 IEEE 31st computer security foundations symposium (CSF), pages 268-282. IEEE, 2018. 1, 2  
+[48] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5  
+[49] Minxing Zhang, Ning Yu, Rui Wen, Michael Backes, and Yang Zhang. Generated distributions are all you need for membership inference attacks against generative models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4839-4849, 2024. 1, 2, 3  
+[50] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Houdong Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1

2025/Variance-Based Membership Inference Attacks Against Large-Scale Image Captioning Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2ef9cabf3488e11e358046d99ae4629df63e82bf12d74e83e695716df209685
+size 339568

2025/VasTSD_ Learning 3D Vascular Tree-state Space Diffusion Model for Angiography Synthesis/221d762b-beb5-4562-9aa6-3ebe0030c356_content_list.json

@@ -0,0 +1,2162 @@
+[
+    {
+        "type": "text",
+        "text": "VasTSD: Learning 3D Vascular Tree-state Space Diffusion Model for Angiography Synthesis",
+        "text_level": 1,
+        "bbox": [
+            153,
+            128,
+            846,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zhifeng Wang $^{1*}$  Renjiao Yi $^{1*}$  Xin Wen $^{1,2}$  Chenyang Zhu $^{1\\dagger}$  Kai Xu $^{1\\dagger}$ $^{1}$ National University of Defense Technology  \n $^{2}$ Xiangjiang Laboratory",
+        "bbox": [
+            220,
+            200,
+            774,
+            257
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            306,
+            326,
+            321
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Angiography imaging is a medical imaging technique that enhances the visibility of blood vessels within the body by using contrast agents. Angiographic images can effectively assist in the diagnosis of vascular diseases. However, contrast agents may bring extra radiation exposure which is harmful to patients with health risks. To mitigate these concerns, in this paper, we aim to automatically generate angiography from non-angiographic inputs, by leveraging and enhancing the inherent physical properties of vascular structures. Previous methods relying on 2D slice-based angiography synthesis struggle with maintaining continuity in 3D vascular structures and exhibit limited effectiveness across different imaging modalities. We propose VasTSD, a 3D vascular tree-state space diffusion model to synthesize angiography from 3D non-angiographic volumes, with a novel state space serialization approach that dynamically constructs vascular tree topologies, integrating these with a diffusion-based generative model to ensure the generation of anatomically continuous vasculature in 3D volumes. A pre-trained vision embeder is employed to construct vascular state space representations, enabling consistent modeling of vascular structures across multiple modalities. Extensive experiments on various angiographic datasets demonstrate the superiority of VasTSD over prior works, achieving enhanced continuity of blood vessels in synthesized angiographic synthesis for multiple modalities and anatomical regions. Project page: https://jefferyzhifeng.github.io/projects/VasTSD/",
+        "bbox": [
+            89,
+            338,
+            485,
+            763
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            91,
+            791,
+            220,
+            806
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Vascular diseases are widely recognized as major contributors to human mortality worldwide [35, 36]. In clinical practice, angiography serves as the gold standard for diag",
+        "bbox": [
+            89,
+            816,
+            483,
+            864
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/231b90764f243ed183a86917f03a98394f5765cd9e7520c2c358e0fafccf0510.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            290,
+            903,
+            411
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/453c199262ffae16c65c58012308bed179dad75b4844c800d4a70166e1015255.jpg",
+        "image_caption": [
+            "Figure 1. (a) In 3D medical volumes, vascular have relatively consistent contextual semantics and structural connectivity. (b) The proposed tree-state space model for simulating 3D vasculature."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            411,
+            903,
+            494
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "nosing vascular conditions and diseases, such as coronary artery disease, atherosclerosis, and aneurysms. In scientific research, angiography provides high-resolution imaging of vascular anatomy, aiding researchers in the detailed exploration of vascular physiology, hemodynamic features, and more, thereby offering crucial data for fundamental research. Traditional angiographic methods, such as Xray Angiogram (XRA), Computed Tomography Angiography (CTA), and Magnetic Resonance Angiography (MRA), have demonstrated their effectiveness, while they require the use of nephrotoxic iodinated contrast agents and expose patients to ionizing radiation, leading to potential carcinogenic risks [39, 45]. Thus, angiography synthesis from nonangiography data becomes a critical direction.",
+        "bbox": [
+            509,
+            564,
+            906,
+            776
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "With the advancement of generative deep learning technologies, the synthesized medical images become more realistic and diverse. Existing medical image generation methods are broadly categorized into several types based on backbone models, including Variational Autoencoders (VAEs) [19, 29], Generative Adversarial Networks (GANs) [7, 43], and Diffusion Models [13, 18]. Recent approaches [20, 24, 36, 49] also explored cross-modal image",
+        "bbox": [
+            509,
+            779,
+            908,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Co-first authors. † Corresponding authors: zhuchenyang07@nudt.edu.cn, kevin.kai.xu@gmail.com",
+        "bbox": [
+            112,
+            875,
+            410,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "15693",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "transformation. For instance, MedGAN [1] enables multimodal conversion by learning end-to-end image transformation. SynDiff [27] employs a conditional diffusion process to capture the correlations in image distributions, allowing for unsupervised multi-modal synthesis. These methods could all potentially be used for angiography synthesis. However, we observe that these image-based generation approaches synthesize incontinent blood vessels in 3D volume. Moreover, each method only works for transformation for one pair of modalities.",
+        "bbox": [
+            89,
+            90,
+            480,
+            241
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We believe that due to variations in imaging devices and principles, the data distributions of different types of medical images exhibit significant disparities within 3D space. Nevertheless, within the same modality, vascular data distributions demonstrate geometric continuity. Thus, integrating medical data from different modalities into a unified state space may be feasible for different multi-modal transformation tasks, and by incorporating the geometric representation of blood vessels and high-level vascular semantic information, continuity of vessels in angiography synthesis can be achieved.",
+        "bbox": [
+            89,
+            244,
+            480,
+            411
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As shown in Fig. 1(a), we observe that in 3D vascular images, the vascular structures (semantic regions) exhibit consistent positioning across slices and share relatively consistent anatomical characteristics (shape). Therefore, the geometric state space consistency across 3D vascular sequences offers a promising direction. This paper proposes a cross-slice state space vision embedding approach aimed at unifying vascular scenes by encoding different sequences of vascular images into the same state space. This unified visual encoding is adaptable for various modalities of vascular data. As illustrated in Fig. 1(b), the geometric structure of blood vessels exists not only within individual slices but also across slices. Masks are guided by a vision embedder to enhance the cross-slice consistency. A tree topology is introduced based on the spatial relationships of the vessels and their input features within the same slice. This approach is implemented by linear complexity dynamic programming, effectively breaking the traditional sequential constraints. A tree-state space model is introduced to facilitate long-range interactions in 3D angiography, enhancing the geometric integrity of synthesized vessels.",
+        "bbox": [
+            89,
+            414,
+            482,
+            731
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As illustrated in Fig. 2, this paper introduces a 3D Vascular Tree-state Space Diffusion model (VasTSD) for synthesizing vascular structures in angiography. In particular, we first utilize a pre-trained vision embedder to process slices from different slices, extracting token weights that are subsequently integrated into a diffusion model for angiography synthesis. This framework supports multi-modal angiographic synthesis tasks, offering a novel solution for clinical diagnostics and fundamental research in angiography. To quantify the effectiveness, comprehensive comparisons are conducted with several prior approaches of medi",
+        "bbox": [
+            89,
+            734,
+            482,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "cal modality conversion. Experiments demonstrate that we achieved state-of-the-art results on commonly used angiography datasets. Furthermore, we provide a clear 3D visualizations of synthesized angiographic data, showing the continuity of blood vessels within resulting 3D volumes. The contributions are summarized as follows:",
+        "bbox": [
+            511,
+            90,
+            903,
+            180
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- VasTSD is the first framework based on state space model for synthesizing angiographies, achieving high structural fidelity from non-angiographic data across multiple modalities and anatomical regions.",
+            "- A dynamic spanning tree of angiography state space features is introduced based on dynamic programming to achieve continuous and semantic long-range interaction of vascular geometry.",
+            "- This paper presents a pre-trained vision embedder mechanism that ensures the geometric consistency of blood vessels and continuous structure representation with directional semantic propagation in angiography."
+        ],
+        "bbox": [
+            511,
+            181,
+            903,
+            364
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Works",
+        "text_level": 1,
+        "bbox": [
+            511,
+            380,
+            663,
+            396
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Deep Learning for Angiography Synthesis",
+        "text_level": 1,
+        "bbox": [
+            511,
+            407,
+            870,
+            422
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Angiography synthesis has gained attention in computer-aided diagnosis, driven by recent advancements in deep learning. Current research focuses on two main approaches to achieve high-quality synthesis. The first approach centres on modality conversion via generative methods within a single modality [20, 21, 41]. For instance, Lyu et al. [24] proposed a generative adversarial method to synthesize CTA from CT. Similarly, Kamran et al. [17] introduced Attention2AngioGAN, which enables the synthesis of fluorescein angiography from retinal fundus images. The second approach focuses on generating high-quality nonangiographic angiograms. For example, Javier et al. [26] achieved high-quality cardiovascular angiography synthesis using MRI with a simple residual U-Net model. Meanwhile, Lyu et al. [25] introduced a denoising diffusion probabilistic model to enhance contrast agents in Head-Neck Dual-energy CT, achieving high-quality contrast-enhanced vasculature with only  $12.5\\%$  of the normal dosage. These methods are effective for specific single-modality synthesis, and vascular geometry is not considered.",
+        "bbox": [
+            511,
+            430,
+            903,
+            732
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Diffusion Models for Medical Image Synthesis",
+        "text_level": 1,
+        "bbox": [
+            511,
+            741,
+            901,
+            758
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Diffusion models have gained significant attention in medical imaging [4, 18, 27] for their ability to generate high-quality images and model complex data distributions, with successful applications in tasks like image reconstruction [32], denoising [5], and segmentation [22]. Recent studies have demonstrated their potential in generating synthetic medical images, particularly for modalities like MRI, CT, and X-ray, where high-quality data is often limited. For instance, Ho et al. [2] introduced a score-based generative",
+        "bbox": [
+            511,
+            763,
+            903,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "15694",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/4a152e97e2be6d5ea869d90ed6032cd4ed0f119c332691226246ac185a2a732d.jpg",
+        "image_caption": [
+            "Figure 2. The overall framework of VasTSD. VasTSD contains a pre-trained vision embedder and a 3D vascular tree-state space diffusion module. The vision embedder encodes 3D medical data and generates embeddings for the diffusion process. The 3D vascular state space diffusion module consists of a forward diffusion and a 3D denoising process based on tree-state space."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            88,
+            911,
+            330
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "model demonstrating remarkable results in generating realistic MRI scans. DiffuseReg [51] and DRDM [48] generate medical deformation fields for medical image synthesis but are not designed for angiography-related tasks. SynDiff [27] proposed an unsupervised diffusion-based method for high-fidelity medical image synthesis. Fast-DDPM [16] focuses on improving the speed of image denoising and image-to-image translation tasks. DiffMa [37] aims to integrate mamba to achieve modality conversion from CT to MRI. While promising in medical image synthesis, diffusion models are sensitive to noise due to their generation process, which can cause information loss and feature blurring. In tasks like angiographic synthesis, incorporating physical characteristics to guide the process is crucial to mitigate these issues.",
+        "bbox": [
+            88,
+            398,
+            485,
+            625
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3. State Space Models (SSMs)",
+        "text_level": 1,
+        "bbox": [
+            89,
+            636,
+            336,
+            652
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "State Space Models (SSMs) [12, 14, 15, 23, 50] are a novel deep learning approach for sequence-to-sequence tasks, particularly effective with long sequences. Structured State Space Models were introduced to address the challenge of long-range dependencies by utilizing a structured state space framework, sparking a wave of subsequent research in the area. Recently, several SSMs-based approaches have been developed for medical applications, including tasks such as medical image segmentation [30, 34, 38], image denoising [28], and classification [6, 44], achieving notable success. These methods typically employ fixed scanning mechanisms to systematically transfer information between time steps, effectively capturing dependencies within sequences. However, they struggle to capture structures and spatial relationships inherent in 3D vascular data. Thus, we propose a dynamic tree-state space model to enhance long-",
+        "bbox": [
+            89,
+            657,
+            483,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "range modeling of vascular geometry.",
+        "bbox": [
+            511,
+            398,
+            764,
+            414
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            511,
+            430,
+            604,
+            446
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Fig. 2 presents the overall framework of VasTSD. It builds upon the foundational principles of diffusion models, overcoming the heterogeneity between different modalities and anatomical regions while addressing key limitations in 3D structured generation within the medical field. The vision embedder is presented in Sec. 3.1. Given 3D medical angiography and non-angiography volumes as input, we first utilize a Variational Autoencoder (VAE) to transform the data into a 3D latent space representation along the cross-sectional slices, denoted as  $Z_{Slices}$ . Next, we perform a vision embedder on the non-angiographic data to generate the conditional embedding representations of the mask  $M$  and tokens  $T$ , which are then used as inputs to the diffusion process. The diffusion module employed for angiography synthesis is described in Sec. 3.2. At each forward diffusion step, we apply noise to the angiography and perform reverse denoising by incorporating a tree-state space to synthesize new angiograms. Here, a novel vascular dynamic tree topology scanning method is proposed, as shown in Fig. 3, which integrates the 3D physical properties of the vasculature to enable high-quality 3D vascular synthesis, as detailed in Sec. 3.3. At last, the 3D angiographic volumes are reconstructed by VAEs.",
+        "bbox": [
+            511,
+            457,
+            906,
+            805
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Pre-trained Vision Embedder",
+        "text_level": 1,
+        "bbox": [
+            511,
+            816,
+            777,
+            832
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The inherent vascular location features in angiography provide prior knowledge for synthesizing continuous angiography from non-angiography. We leverage this by designing a pre-trained vision embedder that effectively captures vas",
+        "bbox": [
+            511,
+            839,
+            906,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "15695",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "cular structures. This embedder extracts high-quality representations of vascular regions from medical slices, forming the foundation for the reverse diffusion process. Specifically, the vision embedder converts non-angiographic 2D vessel slices into latent feature representations, reducing the data dimensionality and complexity. It employs average-and max-pooling to capture global and local features. These pooled features are then passed through a multi-layer perception, which maps them into a specific embedding space.",
+        "bbox": [
+            89,
+            90,
+            480,
+            226
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Additionally, a contrastive learning framework is incorporated by introducing an Exponential Moving Average mechanism to smooth the vascular sequences. The InfoNCE loss is employed to compute both the similarity and dissimilarity between samples. A normalized feature similarity matrix is then computed from the processed features, and cross-entropy loss is applied to guide the model in learning more discriminative feature representations. Thus, the loss function  $\\mathcal{L}_{\\mathrm{InfoNCE}}$  is formulated as:",
+        "bbox": [
+            89,
+            227,
+            480,
+            363
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {I n f o N C E}} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(z _ {i} , z _ {j}\\right) / \\tau\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(\\operatorname {s i m} \\left(z _ {i} , z _ {k}\\right) / \\tau\\right)}, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            135,
+            368,
+            482,
+            405
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $z_{i}$  represents the latent representation  $Z_{Non-angio}$ , by VAE from non-angiography.  $z_{j}$  denotes the tokenized representation of the same input image obtained through the MLP, and  $\\mathrm{sim}(z_i,z_j)$  represents the similarity measure between  $z_{i}$  and  $z_{j}$ , and  $z_{k}$  represents the latent representations of other images in the batch, serving as negative samples in contrast to the positive pair  $z_{i}$  and  $z_{j}$ .",
+        "bbox": [
+            89,
+            410,
+            482,
+            515
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The visual embedding model was trained on paired angiographic and non-angiographic data of the same modality. From each non-angiographic slice, vascular positional and structural information was extracted to generate embedded representations of the mask  $(M)$  and labels  $(T)$  for all slice elements. These embeddings provide spatial constraints that guide the diffusion process, ensuring the synthesized angiography aligns with the vascular structure and enhancing ability of the model to capture detailed features. More details are in supplementary materials.",
+        "bbox": [
+            89,
+            515,
+            482,
+            667
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Diffusion Module for Angiography Synthesis",
+        "text_level": 1,
+        "bbox": [
+            89,
+            674,
+            470,
+            690
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "A 3d tree-state space diffusion model is introduced to synthesize angiographic data from features of the vision embedder. This module consists of two phases: forward diffusion and 3D denoising. In the reverse denoising phase, we propose a tree-state space model to remove noise, generating high-quality angiography iteratively.",
+        "bbox": [
+            89,
+            696,
+            480,
+            787
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Forward diffusion. The latent representation  $z_{\\text{angio}}$  of the angiography is generated by the Variational Autoencoder (VAE). This representation, along with the time step  $t$ , serves as input to the forward diffusion process. In this process, noise is progressively introduced in successive steps. These processes can be parameterized as:",
+        "bbox": [
+            89,
+            787,
+            480,
+            876
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nz _ {\\text {a n g i o .}} ^ {t} = \\sqrt {\\alpha_ {t}} \\cdot z _ {\\text {a n g i o .}} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            169,
+            883,
+            482,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $t$  represents the time step in the diffusion process, and  $\\alpha_{t}$  is a scaling factor that controls the amount of original signal versus noise at each step,  $\\epsilon$  denotes the noise sampled from a gaussian distribution with the same dimensions as the latent representation  $z_{\\mathrm{angio}}$ .",
+        "bbox": [
+            511,
+            90,
+            903,
+            167
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3D denoising. The angiography noise slices generated by the forward diffusion process are divided into blocks  $z_{\\text{angio}}^{Noise,t}$ , which serve as input for the reverse denoising process. Meanwhile, non-angiography features from the pretrained vision embedder provide crucial embedding representations for reverse diffusion. During the reverse denoising process, noise is gradually removed, and angiography is progressively reconstructed. The vascular tree-state space model enables hierarchical conditional feature aggregation, ensuring the seamless merging of embedded representations  $M$  and  $T$  at each restoration step. Finally, after linear transformation, rearrangement, and optimization via conditional fusion, the final angiography slice  $z_{\\text{gen}}$  is generated.",
+        "bbox": [
+            511,
+            167,
+            903,
+            362
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In the reverse diffusion process, the Mean Squared Error (MSE) loss minimizes the difference between the generated image  $x_{t}$  at each timestep and the target noise-free image  $x_0$ , guiding the model to eliminate noise progressively. The Denoising loss is defined as:",
+        "bbox": [
+            511,
+            363,
+            903,
+            436
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {d i f f}} = \\mathbb {E} _ {t, x _ {0}, \\epsilon} \\left[ \\| x _ {t} - x _ {0} \\| ^ {2} \\right], \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            609,
+            448,
+            903,
+            467
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\epsilon$  denotes noise drawn from a standard Gaussian distribution. Minimizing this loss allows the model to iteratively refine the generated image, progressively reducing noise and bringing it closer to the angiography.",
+        "bbox": [
+            511,
+            477,
+            903,
+            537
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Vascular Geometric State Space Model",
+        "text_level": 1,
+        "bbox": [
+            511,
+            546,
+            848,
+            561
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The geometric structure of vessels is crucial for clinical diagnosis. To emphasize these features in angiography synthesis, we map the volumetric properties of vascular geometry into a three-dimensional state space during reverse diffusion. Additionally, we introduce embedding representations via vision embedder and cross-slice attention, enabling effective geometric synthesis of angiography.",
+        "bbox": [
+            511,
+            568,
+            903,
+            672
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The vascular tree-state space model takes two inputs. Firstly, Sec. 3.1 generates the embedding representations of the non-angiography mask  $M$  and tokens  $T$ , which guide the synthesis process. Secondly, the angiographic latent representation  $z_{\\text{angio}}^{Noise,t}$  from forward diffusion is converted into multiple tokens as the other input, which are enhanced with positional embeddings to maintain spatial context.",
+        "bbox": [
+            511,
+            674,
+            903,
+            779
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The inputs to each vascular tree-state space model include the latent representation  $z_{\\mathrm{angio}}^{\\mathrm{Noise},t}$  of angiography and the embedding representations of the non-angiographic mask  $M$  and tokens  $T$ . A linear mapping layer converts each local feature into fixed-dimensional embeddings, followed by deep convolution for alignment. The mamba block adapts to conditional inputs, and long skip connections provide feedback to avoid numerical issues in multi",
+        "bbox": [
+            511,
+            780,
+            903,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "15696",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/67223a528a9b7c220bacbeac096728edc91d77b29921b0d614f9671feb24a80a.jpg",
+        "image_caption": [
+            "Figure 3. Vascular Tree-state space architecture. A dynamic tree is constructed from vessel features and spatial relationships, with embeddings guided by prior slice information."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            93,
+            87,
+            480,
+            338
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "layer processing. Features after layer normalization are connected residually to form the final output.",
+        "bbox": [
+            89,
+            398,
+            483,
+            428
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The vascular tree-state space model, illustrated in Fig. 3, combines the vascular tree scanning algorithm presented in Sec. 3.3.1 with the cross-slice attention mechanism discussed in Sec. 3.3.2, highlighting the significance of different regions within the slice.",
+        "bbox": [
+            89,
+            429,
+            483,
+            503
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3.1. Vascular Tree Scanning Algorithm",
+        "text_level": 1,
+        "bbox": [
+            89,
+            515,
+            377,
+            530
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The tree scanning algorithm constructs an input-adaptive tree topology to transmit and aggregate features efficiently. Input a feature map  $X = \\{x_{i}\\}_{i=1}^{L}$ , where  $L$  is the sequence length, an undirected,  $m$ -connected graph  $G = (V, E)$  is created, with  $m = 4$  ensuring each node is connected to its four neighbors. The edges of this graph are weighted using feature dissimilarity, typically measured by cosine distance. Kruskal's algorithm is then applied to prune the graph into a Minimum Spanning Tree  $G_{T}$ , which preserves spatial relationships while minimizing edge weights. The specific algorithm flow is shown in Algorithm 1.",
+        "bbox": [
+            89,
+            534,
+            483,
+            700
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "State propagation is performed by treating each node as a root and aggregating features from neighboring nodes to capture long-range dependencies. This process is formalized by a state transition matrix and dynamic programming, ensuring information flows both top-down and bottom-up across the tree. At each node  $i$ , feature aggregation is given by  $h[i] = \\sum_{j \\in \\mathrm{Neigh}(i)} S(E[i,j]) \\cdot B[j] \\cdot X[j]$ , where  $\\mathrm{Neigh}(i)$  is the set of neighboring nodes,  $S(E[i,j])$  represents the edge weight,  $B[j]$  is the input transformation matrix, and  $X[j]$  is the input feature. This method preserves spatial structure and enhances feature representation, enabling efficient aggregation of long-range dependencies.",
+        "bbox": [
+            89,
+            702,
+            483,
+            883
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Tree scan loss is computed via backpropagation over a",
+        "bbox": [
+            109,
+            886,
+            482,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [
+            "Algorithm 1: Vascular Tree Scanning Algorithm"
+        ],
+        "code_body": "Require:  $X = \\{x_{i}\\}_{i = 1}^{L},Y$    \nEnsure : output, gradients   \n1 Initialize:  $G\\gets$  initializeGraph(V,E),   \n $G_{T}\\gets$  kruskalMST(G),  $h[i] = 0,\\forall i\\in G_T;$    \n2 for  $i = 1$  to L do   \n3 foreach  $j\\in$  getNeighbors(i,4) do   \n4 addEdge  $(G,i,j,\\mathrm{cosineDist}(X[i],X[j]))$    \n5 Update:   \n $h[i] = \\sum_{j\\in \\mathrm{Neigh}(i)}S(E[i,j])\\cdot B[j]\\cdot X[j],\\forall i\\in G_T;$    \n6 Compute: output  $= C\\cdot (A\\cdot h + D\\cdot X)$ $\\eta_i\\gets \\frac{\\partial\\mathcal{L}}{\\partial h_i};$    \n7 for  $i\\in G_T$  (Bottom-up & Top-down) do   \n8 Aggregate:  $\\eta_{i}\\leftarrow B_{i}\\eta_{i} + \\sum_{j\\in \\mathrm{Children}(i)}\\eta_{j}A_{j};$    \n9 Propagate:  $\\eta_j\\gets \\bar{B}_j\\eta_i,\\quad \\forall j\\in \\mathrm{Children}(i);$    \n0 Calculate:  $\\mathcal{L}_{\\mathrm{scan}} = \\sum_{i}\\frac{1}{2}\\| \\eta_{i}\\|^{2};$    \n1 return output, gradients;",
+        "bbox": [
+            522,
+            111,
+            883,
+            372
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Minimum Spanning Tree (MST), consisting of two stages: aggregation and state propagation. In the aggregation stage, the gradient for each node  $i$  is computed as the partial derivative of the loss with respect to its hidden state  $h_i$ .",
+        "bbox": [
+            511,
+            402,
+            906,
+            463
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\eta_ {i} = \\frac {\\partial \\text {L o s s}}{\\partial h _ {i}}, \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            663,
+            470,
+            905,
+            503
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\eta_ {i} = B _ {i} \\frac {\\partial \\text {L o s s}}{\\partial h _ {i}} + \\sum_ {j \\in \\text {C h i l d r e n} (i)} \\eta_ {j} A _ {j}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            584,
+            508,
+            903,
+            547
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\eta_{i}$  represents the gradient of the hidden state  $h_i$ ,  $\\bar{B}_i$  is the input matrix for node  $i$ ,  $\\bar{A}_j$  is the state matrix for the child node  $j$ ,  $\\operatorname{Par}(t)$  denotes the parent of node  $t$ .",
+        "bbox": [
+            511,
+            547,
+            905,
+            593
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the state propagation stage, the gradients propagate from the root node to the leaf nodes. The gradients for each node  $i$  and its child node  $j$  are updated as follows:",
+        "bbox": [
+            511,
+            594,
+            905,
+            638
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nA _ {i} ^ {-} = A _ {i} - \\eta \\cdot \\frac {\\partial L}{\\partial A _ {i}}, \\quad B _ {i} ^ {-} = B _ {i} - \\eta \\cdot \\frac {\\partial L}{\\partial B _ {i}}, \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            557,
+            647,
+            906,
+            679
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\eta_ {i} = \\bar {A} _ {i} \\eta_ {i}, \\quad \\eta_ {j} = \\bar {B} _ {j} \\eta_ {i}. \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            625,
+            679,
+            903,
+            696
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The tree scan loss is then defined as the sum of the squared magnitudes of the gradients across all nodes:",
+        "bbox": [
+            511,
+            702,
+            905,
+            732
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {s c a n}} = \\sum_ {i} \\frac {1}{2} \\| \\eta_ {i} \\| ^ {2}. \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            638,
+            741,
+            903,
+            777
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The overall loss function is composed of the InfoNCE loss from the vision embedder, the denoising loss from the denoising module, and the tree scan loss from the vascular scanning in the tree-state space. It can be expressed as:",
+        "bbox": [
+            511,
+            779,
+            906,
+            840
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\alpha \\cdot \\mathcal {L} _ {\\text {I n f o N C E}} + \\beta \\cdot \\mathcal {L} _ {\\text {d i f f}} + \\gamma \\cdot \\mathcal {L} _ {\\text {s c a n}}, \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            566,
+            849,
+            903,
+            867
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\alpha, \\beta,$  and  $\\gamma$  are the respective weights assigned to each component of the loss function.",
+        "bbox": [
+            511,
+            869,
+            905,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "15697",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/db2a551bd3295eeef5ce8e9ab1a12157176abc98dcbccf361110ced7a69b2f5d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"3\">Methods</td><td colspan=\"6\">ITKTubeTK</td><td colspan=\"2\">ISICDM 2020</td><td colspan=\"2\">Topcow 2024</td></tr><tr><td colspan=\"2\">T1-Flash</td><td colspan=\"2\">T2</td><td colspan=\"2\">T1-MPRAGE</td><td colspan=\"2\">CT</td><td colspan=\"2\">CTA</td></tr><tr><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td></tr><tr><td>SAGAN [46]</td><td>27.85</td><td>0.9184</td><td>28.16</td><td>0.9254</td><td>27.96</td><td>0.9234</td><td>26.42</td><td>0.8921</td><td>25.32</td><td>0.9018</td></tr><tr><td>cGAN [3]</td><td>28.45</td><td>0.9248</td><td>28.73</td><td>0.9289</td><td>28.68</td><td>0.9264</td><td>27.87</td><td>0.9163</td><td>25.58</td><td>0.9142</td></tr><tr><td>SynDiff [27]</td><td>28.98</td><td>0.9446</td><td>29.06</td><td>0.9482</td><td>29.12</td><td>0.9516</td><td>28.56</td><td>0.9321</td><td>26.13</td><td>0.9317</td></tr><tr><td>Fast-DDPM [16]</td><td>26.92</td><td>0.8958</td><td>27.51</td><td>0.9024</td><td>27.37</td><td>0.8984</td><td>27.38</td><td>0.9128</td><td>24.15</td><td>0.8924</td></tr><tr><td>DiffMa [37]</td><td>29.12</td><td>0.6936</td><td>29.36</td><td>0.7174</td><td>29.28</td><td>0.7058</td><td>28.94</td><td>0.9428</td><td>26.76</td><td>0.7286</td></tr><tr><td>VasTSD</td><td>29.86</td><td>0.9478</td><td>29.72</td><td>0.9384</td><td>29.92</td><td>0.9454</td><td>29.48</td><td>0.9562</td><td>27.36</td><td>0.9372</td></tr></table>",
+        "bbox": [
+            119,
+            87,
+            880,
+            229
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Comparison of modal conversion methods for angiography synthesis. PSNR ↑ and SSIM ↑ results across different datasets. The best and second-best results are highlighted in bold and underlined.",
+        "bbox": [
+            89,
+            239,
+            903,
+            268
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.3.2. Cross-Slice Attention",
+        "text_level": 1,
+        "bbox": [
+            89,
+            273,
+            284,
+            286
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Traditional State Space Models (SSMs) map input sequences  $\\mathbf{u}(t)$  to output sequences  $\\mathbf{y}(t)$ . Recent advancements, such as S4 [9], S5 [31], and S6 [8], have improved the efficiency of SSMs. For instance, S6 introduces a dynamic selection mechanism in the  $A$ -matrix that adapts based on the input content. However, the serialization method in S6 processes each slice independently, making it difficult to capture the 3D vascular structure.",
+        "bbox": [
+            89,
+            291,
+            483,
+            412
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To overcome this limitation, this paper proposes a cross-slice attention mechanism that leverages pre-trained non-angiography embeddings and spatial constraints from slice masks. Within each slice mask, weights are high in vascular regions and low in non-vascular areas. Each vascular region is represented as a feature vector, and the similarity between across slices is computed using cosine similarity:",
+        "bbox": [
+            89,
+            412,
+            483,
+            518
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS _ {i j} = \\frac {\\left(k _ {i} \\cdot k _ {j}\\right)}{\\| k _ {i} \\| \\| k _ {j} \\|}, \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            225,
+            526,
+            480,
+            559
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $k_{i}$  and  $k_{j}$  are the feature vectors for slices  $i$  and  $j$ . To this end, the cross-slice attention coefficient  $\\alpha_{ij}$  represents the attention from node  $i$  to node  $j$  across slices:",
+        "bbox": [
+            89,
+            559,
+            482,
+            604
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\alpha_ {i j} = \\frac {\\exp \\left(\\mathrm {L R e L U} \\left(S _ {i j} + w \\cdot \\operatorname {m a s k} _ {i} \\cdot \\operatorname {m a s k} _ {j}\\right)\\right)}{\\sum_ {k} \\exp \\left(\\mathrm {L R e L U} \\left(S _ {i k} + w \\cdot \\operatorname {m a s k} _ {i} \\cdot \\operatorname {m a s k} _ {k}\\right)\\right)}, \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            98,
+            611,
+            482,
+            646
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $w$  is a hyperparameter controlling mask influence. LReLU represents the LeakyReLU activation function. This cross-slice attention captures the spatial continuity of vascular structures, enabling dynamic state transition matrix construction for 3D vascular synthesis.",
+        "bbox": [
+            89,
+            646,
+            483,
+            720
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            733,
+            220,
+            750
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Experimental Details",
+        "text_level": 1,
+        "bbox": [
+            89,
+            757,
+            289,
+            773
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Dataset. This paper conducts several experiments on brain and lung vessels. The brain vascular data include the ITK-TubeTK dataset [10] and the Topcow2024 dataset [42]. The pulmonary vessels from ISICDM 2020 dataset [11, 33]. All datasets undergo specific preprocessing steps, as detailed in the supplementary material.",
+        "bbox": [
+            89,
+            780,
+            482,
+            869
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. The vision embedder is pretrained on an NVIDIA A100 GPU using non-angiography,",
+        "bbox": [
+            89,
+            869,
+            482,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "with the Variational Auto-Encoder (VAE) employing an exponential moving average (EMA) mode. During the training of the diffusion model, two A100 GPUs are used with the learning rate of  $1 \\times 10^{-4}$  and batch size of 256.",
+        "bbox": [
+            511,
+            273,
+            903,
+            333
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Metrics. In this paper, we evaluate the quality of synthetic vascular 2D slices generated by different methods using two quantitative metrics: Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). The Dice and Jaccard scores evaluate the 3D vascular morphological similarity, while the Connectivity score measures spatial connectivity. Details are provided in the supplementary.",
+        "bbox": [
+            511,
+            334,
+            905,
+            441
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Comparisons",
+        "text_level": 1,
+        "bbox": [
+            511,
+            453,
+            653,
+            469
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "This work evaluates the effectiveness of the proposed VasTSD through experiments on both 2D vascular slices and 3D synthetic angiography volumes applied to brain and lung vasculature.",
+        "bbox": [
+            511,
+            476,
+            905,
+            536
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "2D Vascular Slice Evaluation. To assess the microstructure of synthetic angiography, we compare VasTSD with classic modality conversion methods, including GAN-based SAGAN [46], cGAN [3], and several diffusion-based methods, such as SynDiff [27], Fast-DDPM [16], and DiffMa [37]. All input data are aligned to ensure fairness in comparison. We visualize synthetic angiography results generated from non-angiographic MRI sequences in the ITKTubeTK dataset in Fig. 4. As shown in the red-marked part, the VasTSD obviously outperforms these modality conversion approaches, particularly in synthesizing small vessels and capturing vascular curvature. Based on the ISICDM 2020 pulmonary artery dataset, as shown in Fig. 5, we visualize the 2D slice performance of synthesizing CTA from CT in the body coordinate system. In terms of the branching structure of the vessels, we observe that VasTSD can achieve more accurate and continuous blood branches. The synthesis performance of 2D vascular slices is summarized in Tab. 1. It can be observed that our method achieves state-of-the-art (SOTA) performance across most experiments. Compared to other methods, our approach achieves the highest PSNR across all datasets, with the maximum value reaching 29.92. Although the SSIM for T2 and T1-MPRAGE are slightly lower than those achieved by",
+        "bbox": [
+            511,
+            537,
+            906,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "15698",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d50d9e466f8ef87266478c0913dcf54d76f053bc7eb28a68a015cb95a0d2add9.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            93,
+            89,
+            361,
+            143
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/41a5010109bdc32e17f2c7ba7f30c8bc8bc1f19a122c1d2fc4e706a035e7122c.jpg",
+        "image_caption": [
+            "GT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            120,
+            151,
+            150,
+            174
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4a88bb98fb071e79d1047258ff7b10f8d7b14cbfcbecbdfab09fa29ddb50bfa0.jpg",
+        "image_caption": [
+            "cGAN"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            184,
+            152,
+            215,
+            174
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/6039bd0400f6b1298b661bf99cf1d3517524a32f9517b8824c70c80b6345fc35.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            253,
+            152,
+            282,
+            174
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4ea931b191c749a6b4f07ffadd75e663c19ec7aa0aeb3c30a9e3ca75dc97243e.jpg",
+        "image_caption": [
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            318,
+            152,
+            348,
+            174
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/519bd7c8f3f02745f0f17cb9371e420a013dbbe6cf960f63d00861c3b8e0bfeb.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            364,
+            90,
+            630,
+            143
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/8aba460758d7111854a59e5a4bb11da6741ce015698549cf508ecbbbbbe02b9e.jpg",
+        "image_caption": [
+            "GT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            383,
+            152,
+            413,
+            174
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/df4e2bc25d4869dfcdbc66dea3e24388789af4a49d97c8e911bb9afb3af4d8bf.jpg",
+        "image_caption": [
+            "cGAN"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            449,
+            152,
+            478,
+            174
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7326d2006054eb7cab0c8b68e931f1f8d5ad95f6e7d58713494922c7b701d16a.jpg",
+        "image_caption": [
+            "SynDiff"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            152,
+            547,
+            174
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/8a6f25c1750327c395576ab826477df7e5bb46ea5369f995774fdb233d337d89.jpg",
+        "image_caption": [
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            584,
+            152,
+            612,
+            174
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/cf57630276c88053d6be40e6083b7b838666f9de1da13993d5c1f6a8faeb447f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            630,
+            90,
+            901,
+            143
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/588e8c34232eccb5e963bcdd7ecd6fc18dae6f8306201dd51d97e441b2136fc2.jpg",
+        "image_caption": [
+            "GT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            653,
+            152,
+            683,
+            174
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/69c0addb08600d2243659ed92633a5c419e593e87efceffed6ff9ee45e3192d8.jpg",
+        "image_caption": [
+            "cGAN"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            718,
+            152,
+            748,
+            174
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/0908226374327d96fec8c2a9066d9c1eb5aa732e8eb50a2372a025e1937bf828.jpg",
+        "image_caption": [
+            "SynDiff"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            784,
+            152,
+            815,
+            174
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/5e8f6af4f97b7d1ffc6a812bae46a33846526367d34d4bf9f0ac6b1505c2f405.jpg",
+        "image_caption": [
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            854,
+            152,
+            883,
+            174
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/40794f31bf7cde79681c51aab9d373b4e7b41f57b48f67ffa1aedfc222dc2d4f.jpg",
+        "image_caption": [
+            "SynDiff"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            188,
+            361,
+            239
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/3ee17f9dfd8c53980a60dded7b487d8ef09319614dfc52d2734dc1f672b90191.jpg",
+        "image_caption": [
+            "GT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            120,
+            247,
+            150,
+            270
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e70bdd6498697853c6fb3cd69231bdc302f4c3b7b30f143cd60735fe9e882728.jpg",
+        "image_caption": [
+            "W"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            184,
+            247,
+            215,
+            270
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/90ce7597519df31963659de291170bc6a40259dbf8010824de467368a29a9457.jpg",
+        "image_caption": [
+            "SynDiff"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            251,
+            247,
+            282,
+            270
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/874fea7762e4a1637e568517009b8acb0f531adfc3f268cd27812ecd2976c99c.jpg",
+        "image_caption": [
+            "Ours",
+            "(a) T1-Flash ---> MRA"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            320,
+            247,
+            349,
+            270
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/9dc674a26335b17e872b5609f965802abac482fed6df0be00415d713c7f69d3b.jpg",
+        "image_caption": [
+            "GT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            392,
+            247,
+            413,
+            270
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/968faf77be181a1033bc5af425ff9a99e01fdb7752e52aad37734dd92901b79d.jpg",
+        "image_caption": [
+            "cGAN",
+            "(b) T1-MPRAGE ---> MRA"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            449,
+            247,
+            478,
+            270
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f89d0e62bf0d6562b9c9d2267b76914828cf748037eef6ba6a8ded0d25d7aa55.jpg",
+        "image_caption": [
+            "SynDiff"
+        ],
+        "image_footnote": [
+            "ynDiff [27], and our method."
+        ],
+        "bbox": [
+            517,
+            247,
+            547,
+            270
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/ec812392557a39a40ee4f798bbc3d8ebed6864068b54e47638b039dfd4f6941e.jpg",
+        "image_caption": [
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            583,
+            247,
+            612,
+            270
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/89155d9e830bd651611fd769211ba08c3c5afd4384089b142a39c2e69e90273a.jpg",
+        "image_caption": [
+            "GT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            630,
+            247,
+            681,
+            270
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/c9c35f59fc2c75bde1a147333ae17f2b978daa2fbd60fe0ccf1f16745730f3bb.jpg",
+        "image_caption": [
+            "cGAN",
+            "(c) T2--> MRA"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            718,
+            247,
+            750,
+            270
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/bb7bf185400d3fd88cd285e21d2ab19b841239b59770c510d3081ad2809f94ce.jpg",
+        "image_caption": [
+            "SynDiff"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            784,
+            247,
+            815,
+            270
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/14182b9c005f08d84155d2f6c5e8cb8c0719892e9f61b9d7a42052a7f574ce04.jpg",
+        "image_caption": [
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            852,
+            247,
+            883,
+            270
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/1effd593622218848308ea57bcb3a3644e1aaebebdba72b1ece03f95c1ea57e8.jpg",
+        "image_caption": [
+            "Figure 4. Comparison of 2D angiographic slice modality transformation and ours on the ITKTubeTK dataset. From left to right:",
+            "(a) Axial Plane"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            99,
+            354,
+            475,
+            425
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2c5b5244dc0833872d5b5d2842216f83fc2792096b3ad2f44b1061bd3abfa2bc.jpg",
+        "image_caption": [
+            "(b) Coronal Plane"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            101,
+            435,
+            473,
+            491
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/3e80bff311ad1ad6a68aeb2bb6dea899cf9a73e53eb5f9d6e09b0276b08f6cfd.jpg",
+        "image_caption": [
+            "(c) Sagittal Plane",
+            "Figure 5. Synthesized CTA from the pulmonary artery (nonangiographic) CTs, by cGAN [3], SynDiff [27], DiffMa [37], and our method. From top to bottom, visualized from three different anatomical coordinate systems of the body."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            101,
+            503,
+            473,
+            560
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "SynDiff, it is important to note that SynDiff is a 2D approach. In contrast, our method focuses on the connectivity of vascular structures, which is beneficial for disease diagnosis and holds the potential for improved clinical utility.",
+        "bbox": [
+            89,
+            654,
+            482,
+            715
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3D Vascular Volume Evaluation. The spatial topology of vessels is crucial for diagnosing vascular diseases and surgical planning. This work evaluates VasTSD from the perspective of 3D vascular volume. Due to the lack of previous research on synthetic 3D angiography, the comparative method is based on existing 2D vascular synthesis methods by stacking 2d slices according to their spatial positions. Inspired by multi-view NeRF [40], we visualize the 3D vascular volume from four perspectives, as shown in Fig. 6. Compared to SynDiff and DiffMa, VasTSD generates a more integrated 3D vasculature, demonstrating superior integrity. To quantify the connectivity of the synthetic",
+        "bbox": [
+            89,
+            719,
+            483,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3D vessels and their morphological similarity to real blood vessels, the evaluation results for Dice, Jaccard, and Connectivity Scores are presented in Fig. 7. The results indicate that VasTSD excels in these metrics, demonstrating its superior geometric continuity across vascular slices. Statistical analyses of 3D volumes are provided in the supplementary.",
+        "bbox": [
+            511,
+            356,
+            906,
+            448
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3. Ablation Studies",
+        "text_level": 1,
+        "bbox": [
+            511,
+            459,
+            678,
+            473
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We conduct ablation studies to analyze the contribution of components in the VasTSD to its performance. Specifically, we investigate the impact of the vision embedder, the scanning of the state space, and the loss function. The ablation study on the loss is provided in the supplementary materials. Ablation Comparison of Vision Embedder. In this work, to achieve consistent vascular geometry across different slices, we design a pre-trained vision embedder to capture conditional information  $m, t$ , which enhances the reverse diffusion process of the model. Here we test two other configurations for comparison: one without the pretrained mechanism and another where BioMedCLIP [47] is used to generate the conditional information directly. As shown in Tab. 2, the pre-trained vision embedder representation significantly improves the accuracy of the generated angiography. Compared to the BioMedCLIP vision embedder, which is pre-trained on large-scale medical datasets. Our method demonstrates superior performance in PSNR and SSIM on 2D slices synthesized for multi-modal and anatomical blood vessels. This further underscores the effectiveness of our approach for various types of vessels.",
+        "bbox": [
+            511,
+            482,
+            906,
+            800
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/4de3258871cee8e6c261dfcf8f1e50eb646625196ab70876bf0cd500209093d1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">ITKTubeTK (T1-Flash)</td><td colspan=\"2\">ISICDM 2020</td></tr><tr><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td></tr><tr><td>w/o pre-training</td><td>26.36</td><td>0.9172</td><td>27.57</td><td>0.8993</td></tr><tr><td>BioMedCLIP</td><td>28.68</td><td>0.9286</td><td>28.53</td><td>0.9296</td></tr><tr><td>Ours</td><td>29.86</td><td>0.9478</td><td>29.48</td><td>0.9562</td></tr></table>",
+        "bbox": [
+            517,
+            813,
+            905,
+            878
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Compare the effects of different vision embeddings. PSNR  $\\uparrow$  and SSIM  $\\uparrow$  results across the dataset.",
+        "bbox": [
+            511,
+            890,
+            903,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "15699",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/0742f3c5a5ee55b067511c7b397cb344e9d34e8d315b8bcb4a434e4a8d8c0f61.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            88,
+            297,
+            388
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ddfff90712f4b7533f751bdb6d243fff6a0a157677a1219472622d1af927c5b7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            336,
+            88,
+            426,
+            176
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/936e3755277555e0757dfa3fe73e9057ddbedb9864b77a58e02100e467260574.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            454,
+            88,
+            542,
+            172
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/cb3ee49fda693ba0bd07c7fee63b98b6b48f13a3d2490fd318dbd4ca3a900385.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            558,
+            99,
+            614,
+            166
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/712e6aca641a59d0ca0e7e6a1c5f574e117047c4389ba6f41df5a75df032dbfc.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            627,
+            101,
+            681,
+            167
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/4e6525521d9efab30da8209920522e69cd610e723a380d77a64c3349a4bfcf2f.jpg",
+        "image_caption": [
+            "(a) SynDiff-based method"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            689,
+            90,
+            774,
+            174
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/e20b83cd8b15e4d6f23247ddaa668ba87a8eec438a29dd38b37090b67ae9afe1.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            803,
+            89,
+            898,
+            174
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/cb39f94e11e8b306dc0213263319a6774be5474b050183214a85aafdb9a20b9e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            339,
+            185,
+            424,
+            268
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/e15be5b02a6055bf37677499690acd89d994602dda89b3af08bdb6c5a784b911.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            455,
+            186,
+            540,
+            267
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/7da338af9e88a6ca6876eda11d0af6cddbd78ea2db38e618939340d647f3508d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            560,
+            196,
+            614,
+            258
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/0506659e71e0b2b690988fb00046cf6e1e86a4d093b282ad51cd0e15eb0ff8c7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            627,
+            196,
+            679,
+            258
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/e3b041f16080631bbad2d5537a2610323bfb6261f6cb53b3c9650fcd6f977438.jpg",
+        "image_caption": [
+            "(b) DiffMa-based method"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            689,
+            186,
+            774,
+            268
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6747aaf851c52ffc8f9f297935d55ff02b98ca32a24350cdc2631fae86eb6de0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            803,
+            186,
+            897,
+            268
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/bf0652bd5ddadf56b524c22c285e2a7ec9e6952698ffe13a9919cb77eea1536f.jpg",
+        "image_caption": [
+            "view1-45°"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            339,
+            280,
+            424,
+            363
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/8c0a38347bc9eee1ede17177913bb1baa37eb91b3e4c11eec41baea239314b95.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            455,
+            280,
+            540,
+            359
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/11366afc944aae711d5042a4434294637040be9c80bc374974312926f47df6ba.jpg",
+        "image_caption": [
+            "view2-135°",
+            "(c) Our VasTSD method"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            560,
+            290,
+            614,
+            354
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b29df8153b9ce9eb3e1699581db219c7b7a4d86ff6489462def1e42afe267d1f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            627,
+            290,
+            681,
+            354
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/749ed86552b2b4257fa57828cd249393e8ac6cf5975d0eea84645ed632c51466.jpg",
+        "image_caption": [
+            "view3-225°"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            689,
+            280,
+            774,
+            362
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/4a43830151b2b68101602e29e65c5f35a927f873f49b1920627c833801817093.jpg",
+        "image_caption": [
+            "view4-325°"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            803,
+            279,
+            897,
+            362
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ad76bf26ea116bf31a158f231b7246b7605dd8018173919a1523b90aab4fc223.jpg",
+        "image_caption": [
+            "Figure 6. Comparison of synthetic angiographic 3D volumes from four perspectives. Left: Schematic of the 3D angiography synthesis process. Right: To enhance the visualization of vascular regions, the original data is processed. The first and second rows are based on  $\\S$  ds, while the third row presents the synthetic outcomes of our approach.",
+            "Figure 7. 3D vascular morphology and connectivity evaluation. Higher Dice/Jaccard scores (0-1) indicate greater similarity to real blood vessels, while a higher connectivity score (0-10) reflects better spatial connectivity, smoothness, and structural quality."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            99,
+            444,
+            450,
+            646
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Ablation Comparison of Vascular State Space Scans. In this work, we design a vascular state space scanning method based on the physical principles of vascular geometry, so we investigate the impact of different vascular serialization methods on vascular continuity synthesis. We evaluate the synthesized vascular angiograms and corresponding model parameter counts on the ITKTubeTK dataset, with the results shown in Tab. 3. Compared with two classic statespace serialization methods, our approach achieves higher angiogram synthesis accuracy, especially with greater improvement in structural similarity. At the same time, our model employs the fewest number of parameters.",
+        "bbox": [
+            89,
+            719,
+            483,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/51404f26b6c63080271077d8d38d7f187f055e8e5a4b6702a4ec8be823fa3538.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Backbone</td><td colspan=\"2\">T1-Flash</td><td colspan=\"2\">T2</td><td rowspan=\"2\">Params(M)</td></tr><tr><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td></tr><tr><td>Tree-scan</td><td>29.86</td><td>0.9478</td><td>29.72</td><td>0.9384</td><td>79.24</td></tr><tr><td>Zigzag-scan-B</td><td>26.53</td><td>0.8216</td><td>26.25</td><td>0.8138</td><td>133.8</td></tr><tr><td>Local-scan-S</td><td>28.96</td><td>0.8729</td><td>28.54</td><td>0.8726</td><td>81.83</td></tr></table>",
+        "bbox": [
+            517,
+            445,
+            905,
+            513
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Ablation effects of different SSM serialization methods. Best results are in bold, with PSNR  $\\uparrow$  and SSIM  $\\uparrow$  for ITK-TubeTK (T1-Flash).",
+        "bbox": [
+            511,
+            523,
+            905,
+            565
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusions",
+        "text_level": 1,
+        "bbox": [
+            513,
+            570,
+            638,
+            585
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This work introduces a method VasTSD for angiography synthesis, which can be used to reduce patient exposure to contrast agents in clinical settings. VasTSD dynamically constructs vascular tree topology and integrates it with a diffusion-based generative model, ensuring continuity in synthesized 3D vessels, working for different modalities.",
+        "bbox": [
+            511,
+            595,
+            906,
+            686
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitations. The anisotropy inherent in medical data acquisition may lead to varying image resolution, texture, and tissue response properties in different directions due to physical characteristics and differences in tissue structures across various orientations. Our experiments based on different slice orientations also exhibit some degree of independence. Further exploration of this issue is essential.",
+        "bbox": [
+            511,
+            686,
+            905,
+            792
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements. This work is supported in part by the NSFC (62325211, 62132021, 62372457), the Major Program of Xiangjiang Laboratory (23XJ01009), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), the Natural Science Foundation of Hunan Province of China (2022RC1104) and the NUDT Research Grants (ZK22-52).",
+        "bbox": [
+            511,
+            795,
+            905,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "15700",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            91,
+            89,
+            187,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Karim Armanious, Chenming Jiang, Marc Fischer, Thomas Küstner, Tobias Hepp, Konstantin Nikolaou, Sergios Gatidis, and Bin Yang. Medgan: Medical image translation using gans. Computerized medical imaging and graphics, 79:101684, 2020. 2",
+            "[2] Hyungjin Chung and Jong Chul Ye. Score-based diffusion models for accelerated mri. Medical image analysis, 80: 102479, 2022. 2",
+            "[3] Salman UH Dar, Mahmut Yurt, Levent Karacan, Aykut Erdem, Erkut Erdem, and Tolga Cukur. Image synthesis in multi-contrast mri with conditional generative adversarial networks. IEEE transactions on medical imaging, 38(10): 2375-2388, 2019. 6, 7",
+            "[4] Rucha Deshpande, Muzaffer Özbey, Hua Li, Mark A Anastasio, and Frank J Brooks. Assessing the capacity of a denoising diffusion probabilistic model to reproduce spatial context. IEEE Transactions on Medical Imaging, 2024. 2",
+            "[5] Qi Gao, Zilong Li, Junping Zhang, Yi Zhang, and Hongming Shan. Corediff: Contextual error-modulated generalized diffusion model for low-dose ct denoising and generalization. IEEE Transactions on Medical Imaging, 2023. 2",
+            "[6] Haifan Gong, Luoyao Kang, Yitao Wang, Xiang Wan, and Haofeng Li. nnmamba: 3d biomedical image segmentation, classification and landmark detection with state space model. arXiv preprint arXiv:2402.03526, 2024. 3",
+            "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 1",
+            "[8] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 6",
+            "[9] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 6",
+            "[10] Alper Güngör, Baris Askin, Damla Alptekin Soydan, Can Bariş Top, Emine Ulku Saritas, and Tolga Çukur. Deqpmpi: A deep equilibrium reconstruction with learned consistency for magnetic particle imaging. IEEE Transactions on Medical Imaging, 43(1):321-334, 2023. 6",
+            "[11] Jiachen Han, Naixin He, Qiang Zheng, Lin Li, and Chaoqing Ma. 3d pulmonary vessel segmentation based on improved residual attention u-net. Medicine in Novel Technology and Devices, 20:100268, 2023. 6",
+            "[12] Yuhong He, Long Peng, Qiaosi Yi, Chen Wu, and Lu Wang. Multi-scale representation learning for image restoration with state-space model. arXiv preprint arXiv:2408.10145, 2024.3",
+            "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1",
+            "[14] Vincent Tao Hu, Stefan Andreas Baumann, Ming Gui, Olga Grebenkova, Pingchuan Ma, Johannes S Fischer, and Björn Ommer. *Zigma: A dit-style zigzag mamba diffusion model.* arXiv preprint arXiv:2403.13802, 2024. 3"
+        ],
+        "bbox": [
+            93,
+            114,
+            483,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[15] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 3",
+            "[16] Hongxu Jiang, Muhammad Imran, Linhai Ma, Teng Zhang, Yuyin Zhou, Muxuan Liang, Kuang Gong, and Wei Shao. Fast denoising diffusion probabilistic models for medical image-to-image generation. arXiv preprint arXiv:2405.14802, 2024. 3, 6",
+            "[17] Sharif Amit Kamran, Khondker Fariha Hossain, Alireza Tavakkoli, and Stewart Lee Zuckerbrod. Attention2angiogan: Synthesizing fluorescein angiography from retinal fundus images using generative adversarial networks. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 9122-9129. IEEE, 2021. 2",
+            "[18] Amirhossein Kazerouni, Ehsan Khodapanah Aghdam, Moein Heidari, Reza Azad, Mohsen Fayyaz, Ilker Hacihaliloglu, and Dorit Merhof. Diffusion models in medical imaging: A comprehensive survey. Medical Image Analysis, 88:102846, 2023. 1, 2",
+            "[19] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1",
+            "[20] Alexander Koch, Orhun Utku Aydin, Adam Hilbert, Jana Rieger, Satoru Tanioka, Fujimaro Ishida, and Dietmar Frey. Cross-modality image synthesis from tof-mra to cta using diffusion-based models. arXiv preprint arXiv:2409.10089, 2024.1,2",
+            "[21] Jay Kshirsagar, John McNulty, Bahareh Taji, Derek So, AunYeong Chong, Pascal Theriault-Lauzier, AJ Wisniewski, and Shervin Shrimohammadi. Generative ai-assisted novel view synthesis of coronary arteries for angiography. In 2024 IEEE International Symposium on Medical Measurements and Applications (MeMeA), pages 1-6. IEEE, 2024. 2",
+            "[22] Xinyu Liu, Wuyang Li, and Yixuan Yuan. Diffrect: Latent diffusion label rectification for semi-supervised medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, 2024. 2",
+            "[23] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024.3",
+            "[24] Jinhao Lyu, Ying Fu, Mingliang Yang, Yongqin Xiong, Qi Duan, Caohui Duan, Xueyang Wang, Xinbo Xing, Dong Zhang, Jiaji Lin, et al. Generative adversarial network-based noncontrast ct angiography for aorta and carotid arteries. Radiology, 309(2):e230681, 2023. 1, 2",
+            "[25] Qing Lyu, Josh Tan, Megan E Lipford, Chuang Niu, Micheal E Zapadka, Christopher M Lack, Jonathan D Clemente, Christopher T Whitlow, and Ge Wang. Head-neck dual-energy ct contrast media reduction using diffusion models. arXiv preprint arXiv:2308.13002, 2023. 2",
+            "[26] Javier Montalt-Tordera, Michael Quail, Jennifer A Steeden, and Vivek Muthurangu. Reducing contrast agent dose in cardiovascular mr angiography with deep learning. Journal of Magnetic Resonance Imaging, 54(3):795-805, 2021. 2",
+            "[27] Muzaffer Özbey, Onat Dalmaz, Salman UH Dar, Hasan A Bedel, Şaban Öztürk, Alper Güngör, and Tolga Çukur. Un-"
+        ],
+        "bbox": [
+            516,
+            92,
+            906,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "15701",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "supervised medical image translation with adversarial diffusion models. IEEE Transactions on Medical Imaging, 2023, 2, 3, 6, 7, 8",
+            "[28] Saban Öztürk, Oğuz Can Duran, and Tolga Çukur. Denomamba: A fused state-space model for low-dose ct denoising. arXiv preprint arXiv:2409.13094, 2024. 3",
+            "[29] Mehran Pesteie, Purang Abolmaesumi, and Robert N Rohling. Adaptive augmentation of medical data using independently conditional variational auto-encoders. IEEE transactions on medical imaging, 38(12):2807-2820, 2019. 1",
+            "[30] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 3",
+            "[31] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 6",
+            "[32] Bowen Song, Jason Hu, Zhaoxu Luo, Jeffrey A Fessler, and Liyue Shen. Diffusionblend: Learning 3d image prior through position-aware diffusion score blending for 3d computed tomography reconstruction. arXiv preprint arXiv:2406.10211, 2024. 2",
+            "[33] Wenjun Tan, Luyu Zhou, Xiaoshuo Li, Xiaoyu Yang, Yufei Chen, and Jinzhu Yang. Automated vessel segmentation in lung ct and cta images via deep neural networks. Journal of X-ray science and technology, 29(6):1123-1137, 2021. 6",
+            "[34] Ting Yu Tsai, Li Lin, Shu Hu, Hongtu Zhu, Xin Wang, et al. Uu-mamba: Uncertainty-aware u-mamba for cardiac image segmentation. arXiv preprint arXiv:2405.17496, 2024. 3",
+            "[35] AHA Statistical Update. Heart disease and stroke statistics-2017 update. Circulation, 135:e146-e603, 2017. 1",
+            "[36] Zhifeng Wang, Renjiao Yi, Xin Wen, Chenyang Zhu, and Kai Xu. Cardiovascular medical image and analysis based on 3d vision: A comprehensive survey. *Meta-Radiology*, page 100102, 2024. 1",
+            "[37] Zhenbin Wang, Lei Zhang, Lituan Wang, and Zhenwei Zhang. Soft masked mamba diffusion model for ct to mri conversion. arXiv preprint arXiv:2406.15910, 2024. 3, 6, 7, 8",
+            "[38] Ziyang Wang, Jian-Qing Zheng, Yichi Zhang, Ge Cui, and Lei Li. Mamba-unet: Unet-like pure visual mamba for medical image segmentation. arXiv preprint arXiv:2402.05079, 2024.3",
+            "[39] Steven D Weisbord, Paul M Palevsky, James S Kaufman, Hongsheng Wu, Maria Androsenko, Ryan E Ferguson, Chirag R Parikh, Deepak L Bhatt, and Martin Gallagher. Contrast-associated acute kidney injury and serious adverse outcomes following angiography. Journal of the American College of Cardiology, 75(11):1311-1320, 2020. 1",
+            "[40] Xin Wen, Xuening Zhu, Renjiao Yi, Zhifeng Wang, Chenyang Zhu, and Kai Xu. Cad-nerf: Learning nerfs from uncalibrated few-view images by cad model retrieval. arXiv preprint arXiv:2411.02979, 2024. 7",
+            "[41] Yan Xia, Nishant Ravikumar, Toni Lassila, and Alejandro F Frangi. Virtual high-resolution mr angiography from non-angiographic multi-contrast mris: synthetic vascular model populations for in-silico trials. Medical Image Analysis, 87: 102814, 2023. 2"
+        ],
+        "bbox": [
+            91,
+            90,
+            482,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[42] Kaiyuan Yang, Fabio Musio, Yihui Ma, Norman Juchler, Johannes C Paetzold, Rami Al-Maskari, Luciano Höher, Hongwei Bran Li, Ibrahim Ethem Hamamci, Anjany Sekuboyina, et al. Benchmarking the cow with the topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra. ArXiv, 2023. 6",
+            "[43] Xin Yi, Ekta Walia, and Paul Babyn. Generative adversarial network in medical imaging: A review. Medical image analysis, 58:101552, 2019. 1",
+            "[44] Yubiao Yue and Zhenzhang Li. Medmamba: Vision mamba for medical image classification. arXiv preprint arXiv:2403.03849, 2024. 3",
+            "[45] Pat Zanzonico, Lawrence N Rothenberg, and H William Strauss. Radiation exposure of computed tomography and direct intracoronary angiography: risk has its reward. Journal of the American College of Cardiology, 47(9):1846-1849, 2006. 1",
+            "[46] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In International conference on machine learning, pages 7354–7363. PMLR, 2019. 6",
+            "[47] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023. 7",
+            "[48] Jian-Qing Zheng, Yuanhan Mo, Yang Sun, Jiahua Li, Fuping Wu, Ziyang Wang, Tonia Vincent, and Bartlomiej W Papiez. Deformation-recovery diffusion model (drdm): Instance deformation for image manipulation and synthesis. arXiv preprint arXiv:2407.07295, 2024. 3",
+            "[49] Tao Zhou, Huazhu Fu, Geng Chen, Jianbing Shen, and Ling Shao. Hi-net: hybrid-fusion network for multi-modal mr image synthesis. IEEE transactions on medical imaging, 39(9): 2772-2781, 2020. 1",
+            "[50] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 3",
+            "[51] Yongtai Zhuo and Yiqing Shen. Diffusereg: Denoising diffusion model for obtaining deformation fields in unsupervised deformable image registration. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 597-607. Springer, 2024. 3"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            726
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "15702",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2025/VasTSD_ Learning 3D Vascular Tree-state Space Diffusion Model for Angiography Synthesis/221d762b-beb5-4562-9aa6-3ebe0030c356_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fcd8957b3e5ce647bf1fc9902dc235ab47e3d48de27cd4f8f9c74fc8f8a3768
+size 2795084

2025/VasTSD_ Learning 3D Vascular Tree-state Space Diffusion Model for Angiography Synthesis/full.md

@@ -0,0 +1,444 @@
+# VasTSD: Learning 3D Vascular Tree-state Space Diffusion Model for Angiography Synthesis
+
+Zhifeng Wang $^{1*}$  Renjiao Yi $^{1*}$  Xin Wen $^{1,2}$  Chenyang Zhu $^{1\dagger}$  Kai Xu $^{1\dagger}$ $^{1}$ National University of Defense Technology  
+ $^{2}$ Xiangjiang Laboratory
+
+# Abstract
+
+Angiography imaging is a medical imaging technique that enhances the visibility of blood vessels within the body by using contrast agents. Angiographic images can effectively assist in the diagnosis of vascular diseases. However, contrast agents may bring extra radiation exposure which is harmful to patients with health risks. To mitigate these concerns, in this paper, we aim to automatically generate angiography from non-angiographic inputs, by leveraging and enhancing the inherent physical properties of vascular structures. Previous methods relying on 2D slice-based angiography synthesis struggle with maintaining continuity in 3D vascular structures and exhibit limited effectiveness across different imaging modalities. We propose VasTSD, a 3D vascular tree-state space diffusion model to synthesize angiography from 3D non-angiographic volumes, with a novel state space serialization approach that dynamically constructs vascular tree topologies, integrating these with a diffusion-based generative model to ensure the generation of anatomically continuous vasculature in 3D volumes. A pre-trained vision embeder is employed to construct vascular state space representations, enabling consistent modeling of vascular structures across multiple modalities. Extensive experiments on various angiographic datasets demonstrate the superiority of VasTSD over prior works, achieving enhanced continuity of blood vessels in synthesized angiographic synthesis for multiple modalities and anatomical regions. Project page: https://jefferyzhifeng.github.io/projects/VasTSD/
+
+# 1. Introduction
+
+Vascular diseases are widely recognized as major contributors to human mortality worldwide [35, 36]. In clinical practice, angiography serves as the gold standard for diag
+
+![](images/231b90764f243ed183a86917f03a98394f5765cd9e7520c2c358e0fafccf0510.jpg)
+
+![](images/453c199262ffae16c65c58012308bed179dad75b4844c800d4a70166e1015255.jpg)  
+Figure 1. (a) In 3D medical volumes, vascular have relatively consistent contextual semantics and structural connectivity. (b) The proposed tree-state space model for simulating 3D vasculature.
+
+nosing vascular conditions and diseases, such as coronary artery disease, atherosclerosis, and aneurysms. In scientific research, angiography provides high-resolution imaging of vascular anatomy, aiding researchers in the detailed exploration of vascular physiology, hemodynamic features, and more, thereby offering crucial data for fundamental research. Traditional angiographic methods, such as Xray Angiogram (XRA), Computed Tomography Angiography (CTA), and Magnetic Resonance Angiography (MRA), have demonstrated their effectiveness, while they require the use of nephrotoxic iodinated contrast agents and expose patients to ionizing radiation, leading to potential carcinogenic risks [39, 45]. Thus, angiography synthesis from nonangiography data becomes a critical direction.
+
+With the advancement of generative deep learning technologies, the synthesized medical images become more realistic and diverse. Existing medical image generation methods are broadly categorized into several types based on backbone models, including Variational Autoencoders (VAEs) [19, 29], Generative Adversarial Networks (GANs) [7, 43], and Diffusion Models [13, 18]. Recent approaches [20, 24, 36, 49] also explored cross-modal image
+
+transformation. For instance, MedGAN [1] enables multimodal conversion by learning end-to-end image transformation. SynDiff [27] employs a conditional diffusion process to capture the correlations in image distributions, allowing for unsupervised multi-modal synthesis. These methods could all potentially be used for angiography synthesis. However, we observe that these image-based generation approaches synthesize incontinent blood vessels in 3D volume. Moreover, each method only works for transformation for one pair of modalities.
+
+We believe that due to variations in imaging devices and principles, the data distributions of different types of medical images exhibit significant disparities within 3D space. Nevertheless, within the same modality, vascular data distributions demonstrate geometric continuity. Thus, integrating medical data from different modalities into a unified state space may be feasible for different multi-modal transformation tasks, and by incorporating the geometric representation of blood vessels and high-level vascular semantic information, continuity of vessels in angiography synthesis can be achieved.
+
+As shown in Fig. 1(a), we observe that in 3D vascular images, the vascular structures (semantic regions) exhibit consistent positioning across slices and share relatively consistent anatomical characteristics (shape). Therefore, the geometric state space consistency across 3D vascular sequences offers a promising direction. This paper proposes a cross-slice state space vision embedding approach aimed at unifying vascular scenes by encoding different sequences of vascular images into the same state space. This unified visual encoding is adaptable for various modalities of vascular data. As illustrated in Fig. 1(b), the geometric structure of blood vessels exists not only within individual slices but also across slices. Masks are guided by a vision embedder to enhance the cross-slice consistency. A tree topology is introduced based on the spatial relationships of the vessels and their input features within the same slice. This approach is implemented by linear complexity dynamic programming, effectively breaking the traditional sequential constraints. A tree-state space model is introduced to facilitate long-range interactions in 3D angiography, enhancing the geometric integrity of synthesized vessels.
+
+As illustrated in Fig. 2, this paper introduces a 3D Vascular Tree-state Space Diffusion model (VasTSD) for synthesizing vascular structures in angiography. In particular, we first utilize a pre-trained vision embedder to process slices from different slices, extracting token weights that are subsequently integrated into a diffusion model for angiography synthesis. This framework supports multi-modal angiographic synthesis tasks, offering a novel solution for clinical diagnostics and fundamental research in angiography. To quantify the effectiveness, comprehensive comparisons are conducted with several prior approaches of medi
+
+cal modality conversion. Experiments demonstrate that we achieved state-of-the-art results on commonly used angiography datasets. Furthermore, we provide a clear 3D visualizations of synthesized angiographic data, showing the continuity of blood vessels within resulting 3D volumes. The contributions are summarized as follows:
+
+- VasTSD is the first framework based on state space model for synthesizing angiographies, achieving high structural fidelity from non-angiographic data across multiple modalities and anatomical regions.  
+- A dynamic spanning tree of angiography state space features is introduced based on dynamic programming to achieve continuous and semantic long-range interaction of vascular geometry.  
+- This paper presents a pre-trained vision embedder mechanism that ensures the geometric consistency of blood vessels and continuous structure representation with directional semantic propagation in angiography.
+
+# 2. Related Works
+
+# 2.1. Deep Learning for Angiography Synthesis
+
+Angiography synthesis has gained attention in computer-aided diagnosis, driven by recent advancements in deep learning. Current research focuses on two main approaches to achieve high-quality synthesis. The first approach centres on modality conversion via generative methods within a single modality [20, 21, 41]. For instance, Lyu et al. [24] proposed a generative adversarial method to synthesize CTA from CT. Similarly, Kamran et al. [17] introduced Attention2AngioGAN, which enables the synthesis of fluorescein angiography from retinal fundus images. The second approach focuses on generating high-quality nonangiographic angiograms. For example, Javier et al. [26] achieved high-quality cardiovascular angiography synthesis using MRI with a simple residual U-Net model. Meanwhile, Lyu et al. [25] introduced a denoising diffusion probabilistic model to enhance contrast agents in Head-Neck Dual-energy CT, achieving high-quality contrast-enhanced vasculature with only  $12.5\%$  of the normal dosage. These methods are effective for specific single-modality synthesis, and vascular geometry is not considered.
+
+# 2.2. Diffusion Models for Medical Image Synthesis
+
+Diffusion models have gained significant attention in medical imaging [4, 18, 27] for their ability to generate high-quality images and model complex data distributions, with successful applications in tasks like image reconstruction [32], denoising [5], and segmentation [22]. Recent studies have demonstrated their potential in generating synthetic medical images, particularly for modalities like MRI, CT, and X-ray, where high-quality data is often limited. For instance, Ho et al. [2] introduced a score-based generative
+
+![](images/4a152e97e2be6d5ea869d90ed6032cd4ed0f119c332691226246ac185a2a732d.jpg)  
+Figure 2. The overall framework of VasTSD. VasTSD contains a pre-trained vision embedder and a 3D vascular tree-state space diffusion module. The vision embedder encodes 3D medical data and generates embeddings for the diffusion process. The 3D vascular state space diffusion module consists of a forward diffusion and a 3D denoising process based on tree-state space.
+
+model demonstrating remarkable results in generating realistic MRI scans. DiffuseReg [51] and DRDM [48] generate medical deformation fields for medical image synthesis but are not designed for angiography-related tasks. SynDiff [27] proposed an unsupervised diffusion-based method for high-fidelity medical image synthesis. Fast-DDPM [16] focuses on improving the speed of image denoising and image-to-image translation tasks. DiffMa [37] aims to integrate mamba to achieve modality conversion from CT to MRI. While promising in medical image synthesis, diffusion models are sensitive to noise due to their generation process, which can cause information loss and feature blurring. In tasks like angiographic synthesis, incorporating physical characteristics to guide the process is crucial to mitigate these issues.
+
+# 2.3. State Space Models (SSMs)
+
+State Space Models (SSMs) [12, 14, 15, 23, 50] are a novel deep learning approach for sequence-to-sequence tasks, particularly effective with long sequences. Structured State Space Models were introduced to address the challenge of long-range dependencies by utilizing a structured state space framework, sparking a wave of subsequent research in the area. Recently, several SSMs-based approaches have been developed for medical applications, including tasks such as medical image segmentation [30, 34, 38], image denoising [28], and classification [6, 44], achieving notable success. These methods typically employ fixed scanning mechanisms to systematically transfer information between time steps, effectively capturing dependencies within sequences. However, they struggle to capture structures and spatial relationships inherent in 3D vascular data. Thus, we propose a dynamic tree-state space model to enhance long-
+
+range modeling of vascular geometry.
+
+# 3. Method
+
+Fig. 2 presents the overall framework of VasTSD. It builds upon the foundational principles of diffusion models, overcoming the heterogeneity between different modalities and anatomical regions while addressing key limitations in 3D structured generation within the medical field. The vision embedder is presented in Sec. 3.1. Given 3D medical angiography and non-angiography volumes as input, we first utilize a Variational Autoencoder (VAE) to transform the data into a 3D latent space representation along the cross-sectional slices, denoted as  $Z_{Slices}$ . Next, we perform a vision embedder on the non-angiographic data to generate the conditional embedding representations of the mask  $M$  and tokens  $T$ , which are then used as inputs to the diffusion process. The diffusion module employed for angiography synthesis is described in Sec. 3.2. At each forward diffusion step, we apply noise to the angiography and perform reverse denoising by incorporating a tree-state space to synthesize new angiograms. Here, a novel vascular dynamic tree topology scanning method is proposed, as shown in Fig. 3, which integrates the 3D physical properties of the vasculature to enable high-quality 3D vascular synthesis, as detailed in Sec. 3.3. At last, the 3D angiographic volumes are reconstructed by VAEs.
+
+# 3.1. Pre-trained Vision Embedder
+
+The inherent vascular location features in angiography provide prior knowledge for synthesizing continuous angiography from non-angiography. We leverage this by designing a pre-trained vision embedder that effectively captures vas
+
+cular structures. This embedder extracts high-quality representations of vascular regions from medical slices, forming the foundation for the reverse diffusion process. Specifically, the vision embedder converts non-angiographic 2D vessel slices into latent feature representations, reducing the data dimensionality and complexity. It employs average-and max-pooling to capture global and local features. These pooled features are then passed through a multi-layer perception, which maps them into a specific embedding space.
+
+Additionally, a contrastive learning framework is incorporated by introducing an Exponential Moving Average mechanism to smooth the vascular sequences. The InfoNCE loss is employed to compute both the similarity and dissimilarity between samples. A normalized feature similarity matrix is then computed from the processed features, and cross-entropy loss is applied to guide the model in learning more discriminative feature representations. Thus, the loss function  $\mathcal{L}_{\mathrm{InfoNCE}}$  is formulated as:
+
+$$
+\mathcal {L} _ {\text {I n f o N C E}} = - \log \frac {\exp \left(\operatorname {s i m} \left(z _ {i} , z _ {j}\right) / \tau\right)}{\sum_ {k = 1} ^ {N} \exp \left(\operatorname {s i m} \left(z _ {i} , z _ {k}\right) / \tau\right)}, \tag {1}
+$$
+
+where  $z_{i}$  represents the latent representation  $Z_{Non-angio}$ , by VAE from non-angiography.  $z_{j}$  denotes the tokenized representation of the same input image obtained through the MLP, and  $\mathrm{sim}(z_i,z_j)$  represents the similarity measure between  $z_{i}$  and  $z_{j}$ , and  $z_{k}$  represents the latent representations of other images in the batch, serving as negative samples in contrast to the positive pair  $z_{i}$  and  $z_{j}$ .
+
+The visual embedding model was trained on paired angiographic and non-angiographic data of the same modality. From each non-angiographic slice, vascular positional and structural information was extracted to generate embedded representations of the mask  $(M)$  and labels  $(T)$  for all slice elements. These embeddings provide spatial constraints that guide the diffusion process, ensuring the synthesized angiography aligns with the vascular structure and enhancing ability of the model to capture detailed features. More details are in supplementary materials.
+
+# 3.2. Diffusion Module for Angiography Synthesis
+
+A 3d tree-state space diffusion model is introduced to synthesize angiographic data from features of the vision embedder. This module consists of two phases: forward diffusion and 3D denoising. In the reverse denoising phase, we propose a tree-state space model to remove noise, generating high-quality angiography iteratively.
+
+Forward diffusion. The latent representation  $z_{\text{angio}}$  of the angiography is generated by the Variational Autoencoder (VAE). This representation, along with the time step  $t$ , serves as input to the forward diffusion process. In this process, noise is progressively introduced in successive steps. These processes can be parameterized as:
+
+$$
+z _ {\text {a n g i o .}} ^ {t} = \sqrt {\alpha_ {t}} \cdot z _ {\text {a n g i o .}} + \sqrt {1 - \alpha_ {t}} \epsilon , \tag {2}
+$$
+
+where  $t$  represents the time step in the diffusion process, and  $\alpha_{t}$  is a scaling factor that controls the amount of original signal versus noise at each step,  $\epsilon$  denotes the noise sampled from a gaussian distribution with the same dimensions as the latent representation  $z_{\mathrm{angio}}$ .
+
+3D denoising. The angiography noise slices generated by the forward diffusion process are divided into blocks  $z_{\text{angio}}^{Noise,t}$ , which serve as input for the reverse denoising process. Meanwhile, non-angiography features from the pretrained vision embedder provide crucial embedding representations for reverse diffusion. During the reverse denoising process, noise is gradually removed, and angiography is progressively reconstructed. The vascular tree-state space model enables hierarchical conditional feature aggregation, ensuring the seamless merging of embedded representations  $M$  and  $T$  at each restoration step. Finally, after linear transformation, rearrangement, and optimization via conditional fusion, the final angiography slice  $z_{\text{gen}}$  is generated.
+
+In the reverse diffusion process, the Mean Squared Error (MSE) loss minimizes the difference between the generated image  $x_{t}$  at each timestep and the target noise-free image  $x_0$ , guiding the model to eliminate noise progressively. The Denoising loss is defined as:
+
+$$
+\mathcal {L} _ {\text {d i f f}} = \mathbb {E} _ {t, x _ {0}, \epsilon} \left[ \| x _ {t} - x _ {0} \| ^ {2} \right], \tag {3}
+$$
+
+where  $\epsilon$  denotes noise drawn from a standard Gaussian distribution. Minimizing this loss allows the model to iteratively refine the generated image, progressively reducing noise and bringing it closer to the angiography.
+
+# 3.3. Vascular Geometric State Space Model
+
+The geometric structure of vessels is crucial for clinical diagnosis. To emphasize these features in angiography synthesis, we map the volumetric properties of vascular geometry into a three-dimensional state space during reverse diffusion. Additionally, we introduce embedding representations via vision embedder and cross-slice attention, enabling effective geometric synthesis of angiography.
+
+The vascular tree-state space model takes two inputs. Firstly, Sec. 3.1 generates the embedding representations of the non-angiography mask  $M$  and tokens  $T$ , which guide the synthesis process. Secondly, the angiographic latent representation  $z_{\text{angio}}^{Noise,t}$  from forward diffusion is converted into multiple tokens as the other input, which are enhanced with positional embeddings to maintain spatial context.
+
+The inputs to each vascular tree-state space model include the latent representation  $z_{\mathrm{angio}}^{\mathrm{Noise},t}$  of angiography and the embedding representations of the non-angiographic mask  $M$  and tokens  $T$ . A linear mapping layer converts each local feature into fixed-dimensional embeddings, followed by deep convolution for alignment. The mamba block adapts to conditional inputs, and long skip connections provide feedback to avoid numerical issues in multi
+
+![](images/67223a528a9b7c220bacbeac096728edc91d77b29921b0d614f9671feb24a80a.jpg)  
+Figure 3. Vascular Tree-state space architecture. A dynamic tree is constructed from vessel features and spatial relationships, with embeddings guided by prior slice information.
+
+layer processing. Features after layer normalization are connected residually to form the final output.
+
+The vascular tree-state space model, illustrated in Fig. 3, combines the vascular tree scanning algorithm presented in Sec. 3.3.1 with the cross-slice attention mechanism discussed in Sec. 3.3.2, highlighting the significance of different regions within the slice.
+
+# 3.3.1. Vascular Tree Scanning Algorithm
+
+The tree scanning algorithm constructs an input-adaptive tree topology to transmit and aggregate features efficiently. Input a feature map  $X = \{x_{i}\}_{i=1}^{L}$ , where  $L$  is the sequence length, an undirected,  $m$ -connected graph  $G = (V, E)$  is created, with  $m = 4$  ensuring each node is connected to its four neighbors. The edges of this graph are weighted using feature dissimilarity, typically measured by cosine distance. Kruskal's algorithm is then applied to prune the graph into a Minimum Spanning Tree  $G_{T}$ , which preserves spatial relationships while minimizing edge weights. The specific algorithm flow is shown in Algorithm 1.
+
+State propagation is performed by treating each node as a root and aggregating features from neighboring nodes to capture long-range dependencies. This process is formalized by a state transition matrix and dynamic programming, ensuring information flows both top-down and bottom-up across the tree. At each node  $i$ , feature aggregation is given by  $h[i] = \sum_{j \in \mathrm{Neigh}(i)} S(E[i,j]) \cdot B[j] \cdot X[j]$ , where  $\mathrm{Neigh}(i)$  is the set of neighboring nodes,  $S(E[i,j])$  represents the edge weight,  $B[j]$  is the input transformation matrix, and  $X[j]$  is the input feature. This method preserves spatial structure and enhances feature representation, enabling efficient aggregation of long-range dependencies.
+
+Tree scan loss is computed via backpropagation over a
+
+Algorithm 1: Vascular Tree Scanning Algorithm  
+Require:  $X = \{x_{i}\}_{i = 1}^{L},Y$    
+Ensure : output, gradients   
+1 Initialize:  $G\gets$  initializeGraph(V,E),   
+ $G_{T}\gets$  kruskalMST(G),  $h[i] = 0,\forall i\in G_T;$    
+2 for  $i = 1$  to L do   
+3 foreach  $j\in$  getNeighbors(i,4) do   
+4 addEdge  $(G,i,j,\mathrm{cosineDist}(X[i],X[j]))$    
+5 Update:   
+ $h[i] = \sum_{j\in \mathrm{Neigh}(i)}S(E[i,j])\cdot B[j]\cdot X[j],\forall i\in G_T;$    
+6 Compute: output  $= C\cdot (A\cdot h + D\cdot X)$ $\eta_i\gets \frac{\partial\mathcal{L}}{\partial h_i};$    
+7 for  $i\in G_T$  (Bottom-up & Top-down) do   
+8 Aggregate:  $\eta_{i}\leftarrow B_{i}\eta_{i} + \sum_{j\in \mathrm{Children}(i)}\eta_{j}A_{j};$    
+9 Propagate:  $\eta_j\gets \bar{B}_j\eta_i,\quad \forall j\in \mathrm{Children}(i);$    
+0 Calculate:  $\mathcal{L}_{\mathrm{scan}} = \sum_{i}\frac{1}{2}\| \eta_{i}\|^{2};$    
+1 return output, gradients;
+
+Minimum Spanning Tree (MST), consisting of two stages: aggregation and state propagation. In the aggregation stage, the gradient for each node  $i$  is computed as the partial derivative of the loss with respect to its hidden state  $h_i$ .
+
+$$
+\eta_ {i} = \frac {\partial \text {L o s s}}{\partial h _ {i}}, \tag {4}
+$$
+
+$$
+\eta_ {i} = B _ {i} \frac {\partial \text {L o s s}}{\partial h _ {i}} + \sum_ {j \in \text {C h i l d r e n} (i)} \eta_ {j} A _ {j}, \tag {5}
+$$
+
+where  $\eta_{i}$  represents the gradient of the hidden state  $h_i$ ,  $\bar{B}_i$  is the input matrix for node  $i$ ,  $\bar{A}_j$  is the state matrix for the child node  $j$ ,  $\operatorname{Par}(t)$  denotes the parent of node  $t$ .
+
+In the state propagation stage, the gradients propagate from the root node to the leaf nodes. The gradients for each node  $i$  and its child node  $j$  are updated as follows:
+
+$$
+A _ {i} ^ {-} = A _ {i} - \eta \cdot \frac {\partial L}{\partial A _ {i}}, \quad B _ {i} ^ {-} = B _ {i} - \eta \cdot \frac {\partial L}{\partial B _ {i}}, \tag {6}
+$$
+
+$$
+\eta_ {i} = \bar {A} _ {i} \eta_ {i}, \quad \eta_ {j} = \bar {B} _ {j} \eta_ {i}. \tag {7}
+$$
+
+The tree scan loss is then defined as the sum of the squared magnitudes of the gradients across all nodes:
+
+$$
+\mathcal {L} _ {\text {s c a n}} = \sum_ {i} \frac {1}{2} \| \eta_ {i} \| ^ {2}. \tag {8}
+$$
+
+The overall loss function is composed of the InfoNCE loss from the vision embedder, the denoising loss from the denoising module, and the tree scan loss from the vascular scanning in the tree-state space. It can be expressed as:
+
+$$
+\mathcal {L} _ {\text {t o t a l}} = \alpha \cdot \mathcal {L} _ {\text {I n f o N C E}} + \beta \cdot \mathcal {L} _ {\text {d i f f}} + \gamma \cdot \mathcal {L} _ {\text {s c a n}}, \tag {9}
+$$
+
+where  $\alpha, \beta,$  and  $\gamma$  are the respective weights assigned to each component of the loss function.
+
+<table><tr><td rowspan="3">Methods</td><td colspan="6">ITKTubeTK</td><td colspan="2">ISICDM 2020</td><td colspan="2">Topcow 2024</td></tr><tr><td colspan="2">T1-Flash</td><td colspan="2">T2</td><td colspan="2">T1-MPRAGE</td><td colspan="2">CT</td><td colspan="2">CTA</td></tr><tr><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td></tr><tr><td>SAGAN [46]</td><td>27.85</td><td>0.9184</td><td>28.16</td><td>0.9254</td><td>27.96</td><td>0.9234</td><td>26.42</td><td>0.8921</td><td>25.32</td><td>0.9018</td></tr><tr><td>cGAN [3]</td><td>28.45</td><td>0.9248</td><td>28.73</td><td>0.9289</td><td>28.68</td><td>0.9264</td><td>27.87</td><td>0.9163</td><td>25.58</td><td>0.9142</td></tr><tr><td>SynDiff [27]</td><td>28.98</td><td>0.9446</td><td>29.06</td><td>0.9482</td><td>29.12</td><td>0.9516</td><td>28.56</td><td>0.9321</td><td>26.13</td><td>0.9317</td></tr><tr><td>Fast-DDPM [16]</td><td>26.92</td><td>0.8958</td><td>27.51</td><td>0.9024</td><td>27.37</td><td>0.8984</td><td>27.38</td><td>0.9128</td><td>24.15</td><td>0.8924</td></tr><tr><td>DiffMa [37]</td><td>29.12</td><td>0.6936</td><td>29.36</td><td>0.7174</td><td>29.28</td><td>0.7058</td><td>28.94</td><td>0.9428</td><td>26.76</td><td>0.7286</td></tr><tr><td>VasTSD</td><td>29.86</td><td>0.9478</td><td>29.72</td><td>0.9384</td><td>29.92</td><td>0.9454</td><td>29.48</td><td>0.9562</td><td>27.36</td><td>0.9372</td></tr></table>
+
+Table 1. Comparison of modal conversion methods for angiography synthesis. PSNR ↑ and SSIM ↑ results across different datasets. The best and second-best results are highlighted in bold and underlined.
+
+# 3.3.2. Cross-Slice Attention
+
+Traditional State Space Models (SSMs) map input sequences  $\mathbf{u}(t)$  to output sequences  $\mathbf{y}(t)$ . Recent advancements, such as S4 [9], S5 [31], and S6 [8], have improved the efficiency of SSMs. For instance, S6 introduces a dynamic selection mechanism in the  $A$ -matrix that adapts based on the input content. However, the serialization method in S6 processes each slice independently, making it difficult to capture the 3D vascular structure.
+
+To overcome this limitation, this paper proposes a cross-slice attention mechanism that leverages pre-trained non-angiography embeddings and spatial constraints from slice masks. Within each slice mask, weights are high in vascular regions and low in non-vascular areas. Each vascular region is represented as a feature vector, and the similarity between across slices is computed using cosine similarity:
+
+$$
+S _ {i j} = \frac {\left(k _ {i} \cdot k _ {j}\right)}{\| k _ {i} \| \| k _ {j} \|}, \tag {10}
+$$
+
+where  $k_{i}$  and  $k_{j}$  are the feature vectors for slices  $i$  and  $j$ . To this end, the cross-slice attention coefficient  $\alpha_{ij}$  represents the attention from node  $i$  to node  $j$  across slices:
+
+$$
+\alpha_ {i j} = \frac {\exp \left(\mathrm {L R e L U} \left(S _ {i j} + w \cdot \operatorname {m a s k} _ {i} \cdot \operatorname {m a s k} _ {j}\right)\right)}{\sum_ {k} \exp \left(\mathrm {L R e L U} \left(S _ {i k} + w \cdot \operatorname {m a s k} _ {i} \cdot \operatorname {m a s k} _ {k}\right)\right)}, \tag {11}
+$$
+
+where  $w$  is a hyperparameter controlling mask influence. LReLU represents the LeakyReLU activation function. This cross-slice attention captures the spatial continuity of vascular structures, enabling dynamic state transition matrix construction for 3D vascular synthesis.
+
+# 4. Experiments
+
+# 4.1. Experimental Details
+
+Dataset. This paper conducts several experiments on brain and lung vessels. The brain vascular data include the ITK-TubeTK dataset [10] and the Topcow2024 dataset [42]. The pulmonary vessels from ISICDM 2020 dataset [11, 33]. All datasets undergo specific preprocessing steps, as detailed in the supplementary material.
+
+Implementation Details. The vision embedder is pretrained on an NVIDIA A100 GPU using non-angiography,
+
+with the Variational Auto-Encoder (VAE) employing an exponential moving average (EMA) mode. During the training of the diffusion model, two A100 GPUs are used with the learning rate of  $1 \times 10^{-4}$  and batch size of 256.
+
+Metrics. In this paper, we evaluate the quality of synthetic vascular 2D slices generated by different methods using two quantitative metrics: Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). The Dice and Jaccard scores evaluate the 3D vascular morphological similarity, while the Connectivity score measures spatial connectivity. Details are provided in the supplementary.
+
+# 4.2. Comparisons
+
+This work evaluates the effectiveness of the proposed VasTSD through experiments on both 2D vascular slices and 3D synthetic angiography volumes applied to brain and lung vasculature.
+
+2D Vascular Slice Evaluation. To assess the microstructure of synthetic angiography, we compare VasTSD with classic modality conversion methods, including GAN-based SAGAN [46], cGAN [3], and several diffusion-based methods, such as SynDiff [27], Fast-DDPM [16], and DiffMa [37]. All input data are aligned to ensure fairness in comparison. We visualize synthetic angiography results generated from non-angiographic MRI sequences in the ITKTubeTK dataset in Fig. 4. As shown in the red-marked part, the VasTSD obviously outperforms these modality conversion approaches, particularly in synthesizing small vessels and capturing vascular curvature. Based on the ISICDM 2020 pulmonary artery dataset, as shown in Fig. 5, we visualize the 2D slice performance of synthesizing CTA from CT in the body coordinate system. In terms of the branching structure of the vessels, we observe that VasTSD can achieve more accurate and continuous blood branches. The synthesis performance of 2D vascular slices is summarized in Tab. 1. It can be observed that our method achieves state-of-the-art (SOTA) performance across most experiments. Compared to other methods, our approach achieves the highest PSNR across all datasets, with the maximum value reaching 29.92. Although the SSIM for T2 and T1-MPRAGE are slightly lower than those achieved by
+
+![](images/d50d9e466f8ef87266478c0913dcf54d76f053bc7eb28a68a015cb95a0d2add9.jpg)
+
+![](images/41a5010109bdc32e17f2c7ba7f30c8bc8bc1f19a122c1d2fc4e706a035e7122c.jpg)  
+GT
+
+![](images/4a88bb98fb071e79d1047258ff7b10f8d7b14cbfcbecbdfab09fa29ddb50bfa0.jpg)  
+cGAN
+
+![](images/6039bd0400f6b1298b661bf99cf1d3517524a32f9517b8824c70c80b6345fc35.jpg)
+
+![](images/4ea931b191c749a6b4f07ffadd75e663c19ec7aa0aeb3c30a9e3ca75dc97243e.jpg)  
+Ours
+
+![](images/519bd7c8f3f02745f0f17cb9371e420a013dbbe6cf960f63d00861c3b8e0bfeb.jpg)
+
+![](images/8aba460758d7111854a59e5a4bb11da6741ce015698549cf508ecbbbbbe02b9e.jpg)  
+GT
+
+![](images/df4e2bc25d4869dfcdbc66dea3e24388789af4a49d97c8e911bb9afb3af4d8bf.jpg)  
+cGAN
+
+![](images/7326d2006054eb7cab0c8b68e931f1f8d5ad95f6e7d58713494922c7b701d16a.jpg)  
+SynDiff
+
+![](images/8a6f25c1750327c395576ab826477df7e5bb46ea5369f995774fdb233d337d89.jpg)  
+Ours
+
+![](images/cf57630276c88053d6be40e6083b7b838666f9de1da13993d5c1f6a8faeb447f.jpg)
+
+![](images/588e8c34232eccb5e963bcdd7ecd6fc18dae6f8306201dd51d97e441b2136fc2.jpg)  
+GT
+
+![](images/69c0addb08600d2243659ed92633a5c419e593e87efceffed6ff9ee45e3192d8.jpg)  
+cGAN
+
+![](images/0908226374327d96fec8c2a9066d9c1eb5aa732e8eb50a2372a025e1937bf828.jpg)  
+SynDiff
+
+![](images/5e8f6af4f97b7d1ffc6a812bae46a33846526367d34d4bf9f0ac6b1505c2f405.jpg)  
+Ours
+
+![](images/40794f31bf7cde79681c51aab9d373b4e7b41f57b48f67ffa1aedfc222dc2d4f.jpg)  
+SynDiff
+
+![](images/3ee17f9dfd8c53980a60dded7b487d8ef09319614dfc52d2734dc1f672b90191.jpg)  
+GT
+
+![](images/e70bdd6498697853c6fb3cd69231bdc302f4c3b7b30f143cd60735fe9e882728.jpg)  
+W
+
+![](images/90ce7597519df31963659de291170bc6a40259dbf8010824de467368a29a9457.jpg)  
+SynDiff
+
+![](images/874fea7762e4a1637e568517009b8acb0f531adfc3f268cd27812ecd2976c99c.jpg)  
+Ours  
+(a) T1-Flash ---> MRA
+
+![](images/9dc674a26335b17e872b5609f965802abac482fed6df0be00415d713c7f69d3b.jpg)  
+GT
+
+![](images/968faf77be181a1033bc5af425ff9a99e01fdb7752e52aad37734dd92901b79d.jpg)  
+cGAN  
+(b) T1-MPRAGE ---> MRA
+
+SynDiff  
+![](images/f89d0e62bf0d6562b9c9d2267b76914828cf748037eef6ba6a8ded0d25d7aa55.jpg)  
+ynDiff [27], and our method.
+
+![](images/ec812392557a39a40ee4f798bbc3d8ebed6864068b54e47638b039dfd4f6941e.jpg)  
+Ours
+
+![](images/89155d9e830bd651611fd769211ba08c3c5afd4384089b142a39c2e69e90273a.jpg)  
+GT
+
+![](images/c9c35f59fc2c75bde1a147333ae17f2b978daa2fbd60fe0ccf1f16745730f3bb.jpg)  
+cGAN  
+(c) T2--> MRA
+
+![](images/bb7bf185400d3fd88cd285e21d2ab19b841239b59770c510d3081ad2809f94ce.jpg)  
+SynDiff
+
+![](images/14182b9c005f08d84155d2f6c5e8cb8c0719892e9f61b9d7a42052a7f574ce04.jpg)  
+Ours
+
+![](images/1effd593622218848308ea57bcb3a3644e1aaebebdba72b1ece03f95c1ea57e8.jpg)  
+Figure 4. Comparison of 2D angiographic slice modality transformation and ours on the ITKTubeTK dataset. From left to right:  
+(a) Axial Plane
+
+![](images/2c5b5244dc0833872d5b5d2842216f83fc2792096b3ad2f44b1061bd3abfa2bc.jpg)  
+(b) Coronal Plane
+
+![](images/3e80bff311ad1ad6a68aeb2bb6dea899cf9a73e53eb5f9d6e09b0276b08f6cfd.jpg)  
+(c) Sagittal Plane  
+Figure 5. Synthesized CTA from the pulmonary artery (nonangiographic) CTs, by cGAN [3], SynDiff [27], DiffMa [37], and our method. From top to bottom, visualized from three different anatomical coordinate systems of the body.
+
+SynDiff, it is important to note that SynDiff is a 2D approach. In contrast, our method focuses on the connectivity of vascular structures, which is beneficial for disease diagnosis and holds the potential for improved clinical utility.
+
+3D Vascular Volume Evaluation. The spatial topology of vessels is crucial for diagnosing vascular diseases and surgical planning. This work evaluates VasTSD from the perspective of 3D vascular volume. Due to the lack of previous research on synthetic 3D angiography, the comparative method is based on existing 2D vascular synthesis methods by stacking 2d slices according to their spatial positions. Inspired by multi-view NeRF [40], we visualize the 3D vascular volume from four perspectives, as shown in Fig. 6. Compared to SynDiff and DiffMa, VasTSD generates a more integrated 3D vasculature, demonstrating superior integrity. To quantify the connectivity of the synthetic
+
+3D vessels and their morphological similarity to real blood vessels, the evaluation results for Dice, Jaccard, and Connectivity Scores are presented in Fig. 7. The results indicate that VasTSD excels in these metrics, demonstrating its superior geometric continuity across vascular slices. Statistical analyses of 3D volumes are provided in the supplementary.
+
+# 4.3. Ablation Studies
+
+We conduct ablation studies to analyze the contribution of components in the VasTSD to its performance. Specifically, we investigate the impact of the vision embedder, the scanning of the state space, and the loss function. The ablation study on the loss is provided in the supplementary materials. Ablation Comparison of Vision Embedder. In this work, to achieve consistent vascular geometry across different slices, we design a pre-trained vision embedder to capture conditional information  $m, t$ , which enhances the reverse diffusion process of the model. Here we test two other configurations for comparison: one without the pretrained mechanism and another where BioMedCLIP [47] is used to generate the conditional information directly. As shown in Tab. 2, the pre-trained vision embedder representation significantly improves the accuracy of the generated angiography. Compared to the BioMedCLIP vision embedder, which is pre-trained on large-scale medical datasets. Our method demonstrates superior performance in PSNR and SSIM on 2D slices synthesized for multi-modal and anatomical blood vessels. This further underscores the effectiveness of our approach for various types of vessels.
+
+<table><tr><td rowspan="2">Methods</td><td colspan="2">ITKTubeTK (T1-Flash)</td><td colspan="2">ISICDM 2020</td></tr><tr><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td></tr><tr><td>w/o pre-training</td><td>26.36</td><td>0.9172</td><td>27.57</td><td>0.8993</td></tr><tr><td>BioMedCLIP</td><td>28.68</td><td>0.9286</td><td>28.53</td><td>0.9296</td></tr><tr><td>Ours</td><td>29.86</td><td>0.9478</td><td>29.48</td><td>0.9562</td></tr></table>
+
+Table 2. Compare the effects of different vision embeddings. PSNR  $\uparrow$  and SSIM  $\uparrow$  results across the dataset.
+
+![](images/0742f3c5a5ee55b067511c7b397cb344e9d34e8d315b8bcb4a434e4a8d8c0f61.jpg)
+
+![](images/ddfff90712f4b7533f751bdb6d243fff6a0a157677a1219472622d1af927c5b7.jpg)
+
+![](images/936e3755277555e0757dfa3fe73e9057ddbedb9864b77a58e02100e467260574.jpg)
+
+![](images/cb3ee49fda693ba0bd07c7fee63b98b6b48f13a3d2490fd318dbd4ca3a900385.jpg)
+
+![](images/712e6aca641a59d0ca0e7e6a1c5f574e117047c4389ba6f41df5a75df032dbfc.jpg)
+
+![](images/4e6525521d9efab30da8209920522e69cd610e723a380d77a64c3349a4bfcf2f.jpg)  
+(a) SynDiff-based method
+
+![](images/e20b83cd8b15e4d6f23247ddaa668ba87a8eec438a29dd38b37090b67ae9afe1.jpg)
+
+![](images/cb39f94e11e8b306dc0213263319a6774be5474b050183214a85aafdb9a20b9e.jpg)
+
+![](images/e15be5b02a6055bf37677499690acd89d994602dda89b3af08bdb6c5a784b911.jpg)
+
+![](images/7da338af9e88a6ca6876eda11d0af6cddbd78ea2db38e618939340d647f3508d.jpg)
+
+![](images/0506659e71e0b2b690988fb00046cf6e1e86a4d093b282ad51cd0e15eb0ff8c7.jpg)
+
+![](images/e3b041f16080631bbad2d5537a2610323bfb6261f6cb53b3c9650fcd6f977438.jpg)  
+(b) DiffMa-based method
+
+![](images/6747aaf851c52ffc8f9f297935d55ff02b98ca32a24350cdc2631fae86eb6de0.jpg)
+
+![](images/bf0652bd5ddadf56b524c22c285e2a7ec9e6952698ffe13a9919cb77eea1536f.jpg)  
+view1-45°
+
+![](images/8c0a38347bc9eee1ede17177913bb1baa37eb91b3e4c11eec41baea239314b95.jpg)
+
+![](images/11366afc944aae711d5042a4434294637040be9c80bc374974312926f47df6ba.jpg)  
+view2-135°  
+(c) Our VasTSD method
+
+![](images/b29df8153b9ce9eb3e1699581db219c7b7a4d86ff6489462def1e42afe267d1f.jpg)
+
+![](images/749ed86552b2b4257fa57828cd249393e8ac6cf5975d0eea84645ed632c51466.jpg)  
+view3-225°
+
+![](images/4a43830151b2b68101602e29e65c5f35a927f873f49b1920627c833801817093.jpg)  
+view4-325°
+
+![](images/ad76bf26ea116bf31a158f231b7246b7605dd8018173919a1523b90aab4fc223.jpg)  
+Figure 6. Comparison of synthetic angiographic 3D volumes from four perspectives. Left: Schematic of the 3D angiography synthesis process. Right: To enhance the visualization of vascular regions, the original data is processed. The first and second rows are based on  $\S$  ds, while the third row presents the synthetic outcomes of our approach.  
+Figure 7. 3D vascular morphology and connectivity evaluation. Higher Dice/Jaccard scores (0-1) indicate greater similarity to real blood vessels, while a higher connectivity score (0-10) reflects better spatial connectivity, smoothness, and structural quality.
+
+Ablation Comparison of Vascular State Space Scans. In this work, we design a vascular state space scanning method based on the physical principles of vascular geometry, so we investigate the impact of different vascular serialization methods on vascular continuity synthesis. We evaluate the synthesized vascular angiograms and corresponding model parameter counts on the ITKTubeTK dataset, with the results shown in Tab. 3. Compared with two classic statespace serialization methods, our approach achieves higher angiogram synthesis accuracy, especially with greater improvement in structural similarity. At the same time, our model employs the fewest number of parameters.
+
+<table><tr><td rowspan="2">Backbone</td><td colspan="2">T1-Flash</td><td colspan="2">T2</td><td rowspan="2">Params(M)</td></tr><tr><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td></tr><tr><td>Tree-scan</td><td>29.86</td><td>0.9478</td><td>29.72</td><td>0.9384</td><td>79.24</td></tr><tr><td>Zigzag-scan-B</td><td>26.53</td><td>0.8216</td><td>26.25</td><td>0.8138</td><td>133.8</td></tr><tr><td>Local-scan-S</td><td>28.96</td><td>0.8729</td><td>28.54</td><td>0.8726</td><td>81.83</td></tr></table>
+
+Table 3. Ablation effects of different SSM serialization methods. Best results are in bold, with PSNR  $\uparrow$  and SSIM  $\uparrow$  for ITK-TubeTK (T1-Flash).
+
+# 5. Conclusions
+
+This work introduces a method VasTSD for angiography synthesis, which can be used to reduce patient exposure to contrast agents in clinical settings. VasTSD dynamically constructs vascular tree topology and integrates it with a diffusion-based generative model, ensuring continuity in synthesized 3D vessels, working for different modalities.
+
+Limitations. The anisotropy inherent in medical data acquisition may lead to varying image resolution, texture, and tissue response properties in different directions due to physical characteristics and differences in tissue structures across various orientations. Our experiments based on different slice orientations also exhibit some degree of independence. Further exploration of this issue is essential.
+
+Acknowledgements. This work is supported in part by the NSFC (62325211, 62132021, 62372457), the Major Program of Xiangjiang Laboratory (23XJ01009), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), the Natural Science Foundation of Hunan Province of China (2022RC1104) and the NUDT Research Grants (ZK22-52).
+
+# References
+
+[1] Karim Armanious, Chenming Jiang, Marc Fischer, Thomas Küstner, Tobias Hepp, Konstantin Nikolaou, Sergios Gatidis, and Bin Yang. Medgan: Medical image translation using gans. Computerized medical imaging and graphics, 79:101684, 2020. 2  
+[2] Hyungjin Chung and Jong Chul Ye. Score-based diffusion models for accelerated mri. Medical image analysis, 80: 102479, 2022. 2  
+[3] Salman UH Dar, Mahmut Yurt, Levent Karacan, Aykut Erdem, Erkut Erdem, and Tolga Cukur. Image synthesis in multi-contrast mri with conditional generative adversarial networks. IEEE transactions on medical imaging, 38(10): 2375-2388, 2019. 6, 7  
+[4] Rucha Deshpande, Muzaffer Özbey, Hua Li, Mark A Anastasio, and Frank J Brooks. Assessing the capacity of a denoising diffusion probabilistic model to reproduce spatial context. IEEE Transactions on Medical Imaging, 2024. 2  
+[5] Qi Gao, Zilong Li, Junping Zhang, Yi Zhang, and Hongming Shan. Corediff: Contextual error-modulated generalized diffusion model for low-dose ct denoising and generalization. IEEE Transactions on Medical Imaging, 2023. 2  
+[6] Haifan Gong, Luoyao Kang, Yitao Wang, Xiang Wan, and Haofeng Li. nnmamba: 3d biomedical image segmentation, classification and landmark detection with state space model. arXiv preprint arXiv:2402.03526, 2024. 3  
+[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 1  
+[8] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 6  
+[9] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 6  
+[10] Alper Güngör, Baris Askin, Damla Alptekin Soydan, Can Bariş Top, Emine Ulku Saritas, and Tolga Çukur. Deqpmpi: A deep equilibrium reconstruction with learned consistency for magnetic particle imaging. IEEE Transactions on Medical Imaging, 43(1):321-334, 2023. 6  
+[11] Jiachen Han, Naixin He, Qiang Zheng, Lin Li, and Chaoqing Ma. 3d pulmonary vessel segmentation based on improved residual attention u-net. Medicine in Novel Technology and Devices, 20:100268, 2023. 6  
+[12] Yuhong He, Long Peng, Qiaosi Yi, Chen Wu, and Lu Wang. Multi-scale representation learning for image restoration with state-space model. arXiv preprint arXiv:2408.10145, 2024.3  
+[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1  
+[14] Vincent Tao Hu, Stefan Andreas Baumann, Ming Gui, Olga Grebenkova, Pingchuan Ma, Johannes S Fischer, and Björn Ommer. *Zigma: A dit-style zigzag mamba diffusion model.* arXiv preprint arXiv:2403.13802, 2024. 3
+
+[15] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 3  
+[16] Hongxu Jiang, Muhammad Imran, Linhai Ma, Teng Zhang, Yuyin Zhou, Muxuan Liang, Kuang Gong, and Wei Shao. Fast denoising diffusion probabilistic models for medical image-to-image generation. arXiv preprint arXiv:2405.14802, 2024. 3, 6  
+[17] Sharif Amit Kamran, Khondker Fariha Hossain, Alireza Tavakkoli, and Stewart Lee Zuckerbrod. Attention2angiogan: Synthesizing fluorescein angiography from retinal fundus images using generative adversarial networks. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 9122-9129. IEEE, 2021. 2  
+[18] Amirhossein Kazerouni, Ehsan Khodapanah Aghdam, Moein Heidari, Reza Azad, Mohsen Fayyaz, Ilker Hacihaliloglu, and Dorit Merhof. Diffusion models in medical imaging: A comprehensive survey. Medical Image Analysis, 88:102846, 2023. 1, 2  
+[19] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1  
+[20] Alexander Koch, Orhun Utku Aydin, Adam Hilbert, Jana Rieger, Satoru Tanioka, Fujimaro Ishida, and Dietmar Frey. Cross-modality image synthesis from tof-mra to cta using diffusion-based models. arXiv preprint arXiv:2409.10089, 2024.1,2  
+[21] Jay Kshirsagar, John McNulty, Bahareh Taji, Derek So, AunYeong Chong, Pascal Theriault-Lauzier, AJ Wisniewski, and Shervin Shrimohammadi. Generative ai-assisted novel view synthesis of coronary arteries for angiography. In 2024 IEEE International Symposium on Medical Measurements and Applications (MeMeA), pages 1-6. IEEE, 2024. 2  
+[22] Xinyu Liu, Wuyang Li, and Yixuan Yuan. Diffrect: Latent diffusion label rectification for semi-supervised medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, 2024. 2  
+[23] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024.3  
+[24] Jinhao Lyu, Ying Fu, Mingliang Yang, Yongqin Xiong, Qi Duan, Caohui Duan, Xueyang Wang, Xinbo Xing, Dong Zhang, Jiaji Lin, et al. Generative adversarial network-based noncontrast ct angiography for aorta and carotid arteries. Radiology, 309(2):e230681, 2023. 1, 2  
+[25] Qing Lyu, Josh Tan, Megan E Lipford, Chuang Niu, Micheal E Zapadka, Christopher M Lack, Jonathan D Clemente, Christopher T Whitlow, and Ge Wang. Head-neck dual-energy ct contrast media reduction using diffusion models. arXiv preprint arXiv:2308.13002, 2023. 2  
+[26] Javier Montalt-Tordera, Michael Quail, Jennifer A Steeden, and Vivek Muthurangu. Reducing contrast agent dose in cardiovascular mr angiography with deep learning. Journal of Magnetic Resonance Imaging, 54(3):795-805, 2021. 2  
+[27] Muzaffer Özbey, Onat Dalmaz, Salman UH Dar, Hasan A Bedel, Şaban Öztürk, Alper Güngör, and Tolga Çukur. Un-
+
+supervised medical image translation with adversarial diffusion models. IEEE Transactions on Medical Imaging, 2023, 2, 3, 6, 7, 8  
+[28] Saban Öztürk, Oğuz Can Duran, and Tolga Çukur. Denomamba: A fused state-space model for low-dose ct denoising. arXiv preprint arXiv:2409.13094, 2024. 3  
+[29] Mehran Pesteie, Purang Abolmaesumi, and Robert N Rohling. Adaptive augmentation of medical data using independently conditional variational auto-encoders. IEEE transactions on medical imaging, 38(12):2807-2820, 2019. 1  
+[30] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 3  
+[31] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 6  
+[32] Bowen Song, Jason Hu, Zhaoxu Luo, Jeffrey A Fessler, and Liyue Shen. Diffusionblend: Learning 3d image prior through position-aware diffusion score blending for 3d computed tomography reconstruction. arXiv preprint arXiv:2406.10211, 2024. 2  
+[33] Wenjun Tan, Luyu Zhou, Xiaoshuo Li, Xiaoyu Yang, Yufei Chen, and Jinzhu Yang. Automated vessel segmentation in lung ct and cta images via deep neural networks. Journal of X-ray science and technology, 29(6):1123-1137, 2021. 6  
+[34] Ting Yu Tsai, Li Lin, Shu Hu, Hongtu Zhu, Xin Wang, et al. Uu-mamba: Uncertainty-aware u-mamba for cardiac image segmentation. arXiv preprint arXiv:2405.17496, 2024. 3  
+[35] AHA Statistical Update. Heart disease and stroke statistics-2017 update. Circulation, 135:e146-e603, 2017. 1  
+[36] Zhifeng Wang, Renjiao Yi, Xin Wen, Chenyang Zhu, and Kai Xu. Cardiovascular medical image and analysis based on 3d vision: A comprehensive survey. *Meta-Radiology*, page 100102, 2024. 1  
+[37] Zhenbin Wang, Lei Zhang, Lituan Wang, and Zhenwei Zhang. Soft masked mamba diffusion model for ct to mri conversion. arXiv preprint arXiv:2406.15910, 2024. 3, 6, 7, 8  
+[38] Ziyang Wang, Jian-Qing Zheng, Yichi Zhang, Ge Cui, and Lei Li. Mamba-unet: Unet-like pure visual mamba for medical image segmentation. arXiv preprint arXiv:2402.05079, 2024.3  
+[39] Steven D Weisbord, Paul M Palevsky, James S Kaufman, Hongsheng Wu, Maria Androsenko, Ryan E Ferguson, Chirag R Parikh, Deepak L Bhatt, and Martin Gallagher. Contrast-associated acute kidney injury and serious adverse outcomes following angiography. Journal of the American College of Cardiology, 75(11):1311-1320, 2020. 1  
+[40] Xin Wen, Xuening Zhu, Renjiao Yi, Zhifeng Wang, Chenyang Zhu, and Kai Xu. Cad-nerf: Learning nerfs from uncalibrated few-view images by cad model retrieval. arXiv preprint arXiv:2411.02979, 2024. 7  
+[41] Yan Xia, Nishant Ravikumar, Toni Lassila, and Alejandro F Frangi. Virtual high-resolution mr angiography from non-angiographic multi-contrast mris: synthetic vascular model populations for in-silico trials. Medical Image Analysis, 87: 102814, 2023. 2
+
+[42] Kaiyuan Yang, Fabio Musio, Yihui Ma, Norman Juchler, Johannes C Paetzold, Rami Al-Maskari, Luciano Höher, Hongwei Bran Li, Ibrahim Ethem Hamamci, Anjany Sekuboyina, et al. Benchmarking the cow with the topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra. ArXiv, 2023. 6  
+[43] Xin Yi, Ekta Walia, and Paul Babyn. Generative adversarial network in medical imaging: A review. Medical image analysis, 58:101552, 2019. 1  
+[44] Yubiao Yue and Zhenzhang Li. Medmamba: Vision mamba for medical image classification. arXiv preprint arXiv:2403.03849, 2024. 3  
+[45] Pat Zanzonico, Lawrence N Rothenberg, and H William Strauss. Radiation exposure of computed tomography and direct intracoronary angiography: risk has its reward. Journal of the American College of Cardiology, 47(9):1846-1849, 2006. 1  
+[46] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In International conference on machine learning, pages 7354–7363. PMLR, 2019. 6  
+[47] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023. 7  
+[48] Jian-Qing Zheng, Yuanhan Mo, Yang Sun, Jiahua Li, Fuping Wu, Ziyang Wang, Tonia Vincent, and Bartlomiej W Papiez. Deformation-recovery diffusion model (drdm): Instance deformation for image manipulation and synthesis. arXiv preprint arXiv:2407.07295, 2024. 3  
+[49] Tao Zhou, Huazhu Fu, Geng Chen, Jianbing Shen, and Ling Shao. Hi-net: hybrid-fusion network for multi-modal mr image synthesis. IEEE transactions on medical imaging, 39(9): 2772-2781, 2020. 1  
+[50] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 3  
+[51] Yongtai Zhuo and Yiqing Shen. Diffusereg: Denoising diffusion model for obtaining deformation fields in unsupervised deformable image registration. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 597-607. Springer, 2024. 3

2025/VasTSD_ Learning 3D Vascular Tree-state Space Diffusion Model for Angiography Synthesis/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:582728d85d2d05e858ba9bc0b7656b0240b69e80ea2de3c78b1054090d96183e
+size 626076

2025/VerbDiff_ Text-Only Diffusion Models with Enhanced Interaction Awareness/ad57dfe8-0121-4da2-a044-841b75f6e661_content_list.json

@@ -0,0 +1,1407 @@
+[
+    {
+        "type": "text",
+        "text": "VerbDiff: Text-Only Diffusion Models with Enhanced Interaction Awareness",
+        "text_level": 1,
+        "bbox": [
+            112,
+            130,
+            883,
+            151
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "SeungJu Cha Kwanyoung Lee Ye-Chan Kim Hyunwoo Oh Dong-Jin Kim Hanyang University, South Korea",
+        "bbox": [
+            166,
+            180,
+            830,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{sju9020, mobled37, dpcksd178, komji, djdkim}@hanyang.ac.kr",
+        "bbox": [
+            233,
+            218,
+            754,
+            234
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/6762b1d33f38b4c1d57bbdf425112cd776d156f09d0a48b026c728e205ac7464.jpg",
+        "image_caption": [
+            "\"A basketball player jumping and throwing a basketball and a man blocking the ball\"",
+            "Figure 1. Generated samples illustrating multiple human-object interactions. Each color represents distinct humans, objects, and interaction words. GLIGEN [15] and InteractDiffusion [10] use grounding boxes as additional conditions, whereas Stable Diffusion [25] and VerbDiff rely solely on text."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            107,
+            273,
+            890,
+            575
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            667,
+            328,
+            684
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent large-scale text-to-image diffusion models generate photorealistic images but often struggle to accurately depict interactions between humans and objects due to their limited ability to differentiate various interaction words. In this work, we propose VerbDiff to address the challenge of capturing nuanced interactions within text-to-image diffusion models. VerbDiff is a novel text-to-image generation model that weakens the bias between interaction words and objects, enhancing the understanding of interactions. Specifically, we disentangle various interaction words from frequency-based anchor words and leverage localized interaction regions from generated images to help the model better capture semantics in distinctive words without extra",
+        "bbox": [
+            88,
+            704,
+            485,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "conditions. Our approach enables the model to accurately understand the intended interaction between humans and objects, producing high-quality images with accurate interactions aligned with specified verbs. Extensive experiments on the HICO-DET dataset demonstrate the effectiveness of our method compared to previous approaches.",
+        "bbox": [
+            511,
+            670,
+            908,
+            762
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            513,
+            797,
+            645,
+            811
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recently, text-to-image diffusion models have demonstrated the ability to produce photorealistic images from natural texts [25]. However, these models sometimes struggle to accurately depict interactions between humans and objects [10]. For example, they often fail to differentiate",
+        "bbox": [
+            511,
+            825,
+            906,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "8041",
+        "bbox": [
+            482,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/0266c3ca1bcd8f7ab298f6f0e703e45274128c8435b0270bd8e5974da1596d53.jpg",
+        "image_caption": [
+            "Figure 2. Examples of interactions from InteractDiffusion [10], showing a lack of understanding of interaction words. The model relies on precise bounding boxes rather than understanding interaction words to exhibit accurate interactions, as shown in the results for the \"red\" and \"blue\" boxes."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            90,
+            480,
+            303
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "between semantically distinct prompts, such as \"A person walking a bicycle\" and \"A person riding a bicycle,\" particularly in capturing the semantics in verbs that are crucial for accurately depicting the intended interaction.",
+        "bbox": [
+            89,
+            412,
+            483,
+            474
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We hypothesize that this problem arises from the strong object bias of CLIP [23], which fails to adequately capture semantic differences between verbs [22]. Specifically, CLIP [23] often focuses on objects in prompts, overlooking semantic distinctions in interaction verbs. To enhance prompt interpretation ability of models, previous methods have leveraged additional modules such as large language models (LLMs) [6, 7, 16] or extra bounding boxes [10, 15, 30, 34] to provide explicit relational visual information.",
+        "bbox": [
+            89,
+            474,
+            483,
+            611
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "For example, InteractDiffusion [10] introduces controllable image generation using Human-Object Interaction (HOI) information with bounding boxes for human, relation, and object  $<\\mathrm{h}, \\mathrm{r}, \\mathrm{o}>$  triplets. Nevertheless, as shown in Fig. 2, the model still struggles with accurate interactions even with these additional conditions. For instance, given the prompt \"walking, bicycle,\" the model fails to depict the intended interaction when semantically distinctive interaction words share a similar bounding box (e.g. walking and riding). This suggests that the model lacks an understanding of interaction semantics and relies heavily on precise bounding boxes, which are labor-intensive to provide.",
+        "bbox": [
+            89,
+            612,
+            483,
+            792
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To improve the model's interaction understandability without additional conditions, we propose VerbDiff, a novel text-to-image generation method that better distinguishes interaction words via several objectives. First, we apply Relation Disentanglement Guidance (RDG) using frequency-based anchor texts from the dataset for each human-object pair to reduce interaction bias in generated images. We",
+        "bbox": [
+            89,
+            795,
+            483,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "observe that the exhibited interactions in generated images have bias, which follows frequent verbs corresponding to each human-object pair in the data distribution. To address this, we align interaction features between real and generated images while separating those of generated images from anchor text features, enabling a more effective distinction of interaction words. Additionally, we introduce a balancing factor to address the extreme long-tail distribution of interaction words across human-object pairs, adjusting interaction modification during training. This allows the model to learn diverse interactions in a balanced way.",
+        "bbox": [
+            511,
+            90,
+            906,
+            258
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Secondly, to enforce the model to better focus on the interaction regions between humans and objects, we introduce Interaction Direction Guidance (IDG). This approach emphasizes fine-grained semantic distinctions in localized interactions. We design an Interaction Region (IR) module to capture specific interaction areas in generated images, leveraging cross-attention maps in text-only diffusion models. The IR module extracts interaction regions using the centroids of cross-attention maps associated with  $<\\mathrm{h}$ , r, o> tokens. We obtain biased interaction features from these regions and apply interaction direction guidance. This guidance effectively modifies interaction regions within images, guiding Stable Diffusion (SD) to concentrate more accurately on these regions while capturing nuanced interaction details. We validate our method's effectiveness on the HICO-DET dataset [2]. Our approach accurately captures semantic differences between interaction words and generates high-quality images with accurate interactions compared to SD [25], InteractDiffusion [10], and other prior methods. Additionally, we introduce a new metric using Sentence BERT [24] for image generation tasks. This metric offers improved capability in discerning subtle and nuanced meanings of interaction words compared to CLIP. By focusing on semantic distinctions in interactions, this metric enables more precise evaluation, complementing CLIP similarity. Our contributions are as follows:",
+        "bbox": [
+            511,
+            268,
+            906,
+            661
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "(1) We introduce VerbDiff, a novel text-to-image model that enhances the semantic interaction understanding by distinguishing interaction verbs and generating images with accurate human-object interactions without any additional conditions. (2) To reduce generating images with biased interactions, we apply relation disentanglement guidance using frequency-based anchor text extracted from real data distributions. (3) Our proposed IR module captures detailed interaction regions using cross-attention map centroids, providing specified-interaction region guidance without explicit bounding boxes. (4) We implement a Sentence-BERT metric to effectively capture semantic differences between interaction sentences, addressing CLIP's limitations. Extensive experiments demonstrate that our approach outperforms previous methods across various settings.",
+        "bbox": [
+            511,
+            674,
+            908,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "8042",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            91,
+            89,
+            233,
+            104
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.1. Diffusion Models",
+        "text_level": 1,
+        "bbox": [
+            89,
+            114,
+            259,
+            130
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Recently, diffusion models have achieved state-of-the-art performance in text-to-image generation [25] and controllable image generation [10, 15, 33]. While Stable Diffusion (SD) [25] integrates conditioning embeddings into cross-attention layers for text-to-image generation, it often struggles to fully align generated images with text instructions. This arises from inherent limitations of texts to convey exact visual information and strong object bias of CLIP [22].",
+        "bbox": [
+            89,
+            138,
+            482,
+            258
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To address these, recent research has focused on controllable image generation using bounding boxes [15, 30, 34]. For example, GLIGEN [15] extends SD to incorporate grounding inputs such as bounding boxes, image prompts, key points, and spatially aligned conditions. LayoutDiffusion [34] enables multiple bounding box inputs for layout-specific generation. Although these methods better generate images, they require additional conditions beyond text input to achieve the intended image outputs. In contrast, our proposed method achieves comparable results without additional grounding information, providing a more efficient alternative for precise and flexible generation.",
+        "bbox": [
+            89,
+            261,
+            483,
+            441
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2. Relation Generation",
+        "text_level": 1,
+        "bbox": [
+            89,
+            453,
+            284,
+            468
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Relation generation between objects has been explored leveraging LLMs [6, 7, 16] and scene graph [20, 26, 28, 29, 31, 32]. While LLM-based methods can generate accurate object layouts, they primarily focus on compositional relations rather than interactions [6, 7, 16]. Scene graph-based approaches have addressed these limitations by combining scene graph encoders to enhance prompt understandability [20, 26, 28, 29, 31, 32]. However, they require additional training to extract scene graphs from prompts and struggle to capture subtle semantic differences in interactions.",
+        "bbox": [
+            89,
+            477,
+            482,
+            628
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To depict accurate interactions during generation, researchers have focused on human-object interactions (HOI) [10, 12-14]. Some methods use inversion-based frameworks to capture interaction semantics but require optimization for each interaction word [12, 13]. More generalized approaches leverage additional conditions such as human poses [14] or bounding boxes [10]. Specifically, InteractDiffusion [10] leverages the HOI information with bounding boxes to generate interactions. However, this method still relies on additional information rather than semantic differences between prompts. Our work is similar to InteractDiffusion [10] by leveraging HOI information but differs from focusing on semantic differences between interaction words to enhance interaction understanding in SD. This approach aims to improve the generation of nuanced interactions without relying on additional conditions. It could potentially lead to more accurate and diverse depictions of human-object interactions in synthesized images.",
+        "bbox": [
+            89,
+            628,
+            482,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            513,
+            89,
+            604,
+            104
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our goal is to generate images with accurate interactions by enhancing the model understandability of the interaction word without additional conditions. Our proposed method, VerbDiff as illustrated in Fig. 3 contains two main parts: (1) Relation disentanglement guidance leverages frequency-based anchor texts from human and object pairs along with input texts and real and generated image features. This guidance helps visually align the generated interaction features with real ones and distinguish the semantic differences between interaction words. (2) Interaction direction guidance that helps the model concentrate more on the interaction region while modifying the image features. To capture the interaction region in arbitrarily generated images, we further design an interaction region module (IR module) to extract a centroid-based interaction region from the cross-attention map corresponding to  $< \\mathrm{h},\\mathrm{r},\\mathrm{o} >$  tokens.",
+        "bbox": [
+            511,
+            114,
+            906,
+            357
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Preliminary",
+        "text_level": 1,
+        "bbox": [
+            511,
+            364,
+            643,
+            382
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "SD Training Objective. Text-to-image generation models, such as Stable Diffusion (SD) [25], show photo-realistic images that align with given text prompts through an efficient two-stage training scheme. The first stage learns to map the input images  $I$  into latent space  $z$ . The second stage learns the denoising UNet to gradually predict the real noise in the diffusion forward process. Starting from the original latent image  $z_0$ , the diffusion model gradually adds noise  $\\epsilon$  to produce noisy sample  $z_t$  through uniformly sampled time steps  $t$ . Then, with text condition  $T$ , the UNet predicts the noise  $\\epsilon_{\\theta}$  for each time step to reconstruct  $z_t$  to  $z_0$  as follows:",
+        "bbox": [
+            511,
+            388,
+            906,
+            555
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\min  _ {\\theta} \\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {z, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| (\\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, T)) \\| _ {2} ^ {2} \\right]. \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            527,
+            564,
+            906,
+            592
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Cross-Attention Layer. When given text, the CLIP [23] text encoder produces text embedding, and the UNet in SD takes embedding and computes cross-attention between encoded text  $e$  and intermediate image features  $\\phi(z_{t})$ :",
+        "bbox": [
+            511,
+            601,
+            905,
+            662
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {A} = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d}}\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            622,
+            672,
+            906,
+            709
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where query  $Q = W_{Q} \\cdot \\phi(z_{t})$  and key  $K = W_{K} \\cdot e$  are computed with projection matrix  $W_{Q}$ ,  $W_{K}$  and  $d$  is the dimension of  $Q$ . These cross-attention maps  $\\mathcal{A}$  enable the text-to-image diffusion models to depict texts into image features. In addition, each  $\\mathcal{A}_{j}$ , a cross-attention map corresponding to  $j^{th}$  token in prompts, represents the attention weight on intermediate image features in various resolutions.",
+        "bbox": [
+            511,
+            717,
+            905,
+            823
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Relation Disentanglement Guidance",
+        "text_level": 1,
+        "bbox": [
+            511,
+            832,
+            826,
+            848
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To depict accurate interaction in generated images without additional conditions, we first focus on the semantic difference between interaction words, especially the verbs. As the",
+        "bbox": [
+            511,
+            854,
+            905,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "8043",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/1f78b73afcb0a51fc75ba1ffc3ba614dd9f9b589ba3cbf7de08546515d27a8ca.jpg",
+        "image_caption": [
+            "Figure 3. Pipeline of VerbDiff. VerbDiff uses Relation Disentanglment Guidance (Sec. 3.2) that separates the interaction features from the anchor text for each human-object pair. Additionally, it contains the IR module (Sec. 3.3) (right) which extracts localized interaction regions from generated images without explicit bounding boxes and Interaction Direction Guidance (Sec. 3.3) that guides the model to focus more on the fine-grained interaction regions."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            98,
+            88,
+            669,
+            398
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/402f7dc2c45e45c40a2f9962c5008383d634ad746dbbb00349cad9eb8b99b42c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            681,
+            88,
+            901,
+            398
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "SD is known to have limitations in capturing the semantic nuances in interaction words, we apply relation disentangle-ment guidances. The goal of this guidance is to alleviate the relation bias within human-object pairs and improve interaction comprehension through anchor text, input text, and real and generated images.",
+        "bbox": [
+            88,
+            492,
+            480,
+            583
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "First, given the ground-truth text  $T^{gt} =$  \"A photo of a  $\\{h\\} \\{r^{gt}\\} a / an\\{o\\}$ \" (e.g. \"A photo of a person holding a backpack\"), SD generates images  $I^{gen}$  during training. Although  $T^{gt}$  specifies the target interaction,  $I^{gen}$  often contains inaccurate and biased interactions that fail to meet verbs between each human and object pair. We define a sentence with accurate interaction verbs corresponding to inaccurate interaction image  $I^{gen}$  as  $T^{anc}$ .",
+        "bbox": [
+            89,
+            584,
+            482,
+            705
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In particular, we observe biased interaction images tend to depict verbs that appear more frequently in the data along human-object pair. For instance, when generating images with the object \"backpack,\" the images commonly show a person wearing a backpack, which is the most frequent verb in the dataset. Based on this observation, we hypothesize that the biased interaction images typically align with frequent verbs in the real data distribution. Therefore, we define  $r^{anc}$  as the verb with the highest sample count for each human-object pair as follows:",
+        "bbox": [
+            89,
+            705,
+            482,
+            857
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nr ^ {a n c} = \\underset {r \\in R _ {o}} {\\arg \\max } \\mathcal {C} (r | o), \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            202,
+            871,
+            482,
+            897
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $R_{o}$  denotes the set of verbs related to each human-object pair in dataset  $\\mathcal{D}$  and  $\\mathcal{C}(\\cdot)$  represents the number of samples corresponding to each verb  $r$  when given  $o$ .",
+        "bbox": [
+            511,
+            492,
+            906,
+            537
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Given the anchor verb  $r^{anc}$ , to reduce the generation of biased interaction images, we encourage SD to create images that with distant semantics from  $T^{anc} = \"A photo of a \\{h\\} \\{r^{anc}\\} a/an \\{o\\}\"$  (e.g. \"wearing a backpack\"). Specifically, we guide the  $I^{gen}$  closer to the accurate interaction sentence  $T^{gt}$  while pushing away from the sentence with anchor interaction  $T^{anc}$  by leveraging triplet loss [11].",
+        "bbox": [
+            511,
+            539,
+            906,
+            645
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We extract normalized feature  $f^{gen}$ ,  $e^{gt}$ , and  $e^{anc}$  through CLIP [23] encoder  $E_{T}$  and  $E_{I}$  from  $I^{gen}$ ,  $T^{gt}$ , and  $T^{anc}$ , respectively. Then, we set  $e^{gt}$  as positive and  $e^{anc}$  as negative when given anchor  $f^{gen}$  and compute the cosine similarity with the generated image features,  $f^{gen}$ , and each text feature,  $e^{gt}$  and  $e^{anc}$ . Finally, we introduce a triplet loss to maximize the cosine similarity between  $f^{gen}$  and  $e^{gt}$  larger than the similarity between  $f^{gen}$  and  $e^{anc}$  with a margin of  $m$  as follows:",
+        "bbox": [
+            511,
+            645,
+            908,
+            781
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {t r i p l e}} = \\max  \\left(0, m + \\operatorname {s i m} \\left(f ^ {g e n}, e ^ {g t}\\right) - \\operatorname {s i m} \\left(f ^ {g e n}, e ^ {a n c}\\right)\\right). \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            511,
+            792,
+            906,
+            825
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In addition to distinguishing the semantic difference between verbs, to generate accurate interaction at the image level as well, we add image aligning loss to make the model follow the interaction in  $I^{gt}$ . However,  $I^{gt}$  often contains",
+        "bbox": [
+            511,
+            839,
+            908,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "8044",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "multiple humans, objects, and interactions within a single image while we only use a single human, object, and interaction triplet as input during image generation. This input discrepancy makes it difficult for the model to focus solely on the relevant interaction specified by the input prompts, causing confusion when depicting the specified interaction words. To focus on the part in the image corresponding to the interaction label, we extract mask region  $\\mathcal{M}$  combining the human and object bounding boxes in a dataset and obtain  $I_{\\mathcal{M}}^{gt} = I^{gt} \\odot \\mathcal{M}$ . Then, we extract  $f_{\\mathcal{M}}^{gt}$  through image encoder and minimize the cosine similarity between  $f_{\\mathcal{M}}^{gt}$  and  $f^{gen}$  as below:",
+        "bbox": [
+            89,
+            90,
+            483,
+            272
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = 1 - \\frac {f _ {\\mathcal {M}} ^ {g t} \\cdot f ^ {g e n}}{| f _ {\\mathcal {M}} ^ {g t} | | f ^ {g e n} |}. \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            200,
+            282,
+            483,
+            321
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Adaptive Interaction Modification. When applying  $\\mathcal{L}_{\\mathrm{triple}}$  and  $\\mathcal{L}_{\\mathrm{align}}$  to capture the semantic differences between interaction words, we find that the modification extent applied to the generated images differs across interaction words. As each interaction text  $T^{gt}$  has widely varying sample counts, we also focus on balancing the modification extent across interaction words. Specifically, we adaptively scale the loss considering the number of samples following [4]. We apply an effective number, which is defined as:",
+        "bbox": [
+            89,
+            328,
+            483,
+            464
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\alpha (k) = \\frac {1 - \\beta^ {n _ {k}}}{1 - \\beta}, \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            225,
+            472,
+            482,
+            503
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\beta = (N - 1) / N$ ,  $n_k$  is the number of samples in class  $k$ , and  $\\mathbf{N}$  is the total number of samples in the dataset.",
+        "bbox": [
+            89,
+            513,
+            483,
+            542
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Relation Disentanglement Guidance. Finally, we multiply  $\\alpha$  on  $\\mathcal{L}_{\\mathrm{triple}}$  and  $\\mathcal{L}_{\\mathrm{align}}$ , defining relation disentanglement guidance as follows:",
+        "bbox": [
+            89,
+            544,
+            483,
+            588
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R D G}} = \\alpha \\cdot \\left(\\mathcal {L} _ {\\text {t r i p l e}} + \\mathcal {L} _ {\\text {a l i g n}}\\right). \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            599,
+            482,
+            618
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This guidance induces SD to generate images with accurate interactions for each verb and to distinguish the semantic difference between interactions.",
+        "bbox": [
+            89,
+            626,
+            483,
+            671
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3. IR Module & Interaction Direction Guidance",
+        "text_level": 1,
+        "bbox": [
+            89,
+            681,
+            480,
+            696
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Relation Disentanglement Guidance (RDG) effectively enables the model to better distinguish and represent each interaction word, helping to exhibit the intended interactions at the image level. However, the model sometimes shows images that misalign with human expectations, especially in the local regions where human-object interactions occur. We hypothesize that focusing on finer, localized interaction regions helps the model improve the detailed interaction expressions, thereby allowing the model to better capture more region-specific interaction information.",
+        "bbox": [
+            89,
+            703,
+            482,
+            854
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To focus more on specific interaction regions, we apply interaction direction guidance (IDG), leveraging interaction regions extracted from the interaction region module with",
+        "bbox": [
+            89,
+            854,
+            483,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "cross-attention maps corresponding to the  $<\\mathrm{h}, \\mathrm{r}, \\mathrm{o}>$  tokens in generated images. IDG enhances the model's ability to depict region-level details, guiding the depicted interactions to align more perceptually close to human intentions.",
+        "bbox": [
+            511,
+            90,
+            905,
+            151
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Interaction Region Module. To achieve a more detailed interaction expression, we first extract interaction regions from both real and generated images. Although real images contain explicit human and object bounding boxes in the dataset, it is challenging to extract specific interaction regions without additional conditions from generated images. Recently, [9] demonstrated that the cross-attention maps capture the human or object's existence corresponding to each token in prompts. We leverage the cross-attention maps  $\\mathcal{A}_h$ ,  $\\mathcal{A}_r$  and  $\\mathcal{A}_o$  corresponding to  $<\\mathrm{h}, \\mathrm{r}, \\mathrm{o}>$  token to extract interaction region between human and objects.",
+        "bbox": [
+            511,
+            152,
+            906,
+            319
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "However, it is nontrivial to directly extract the specific interaction region from a continuous attention map. We apply the centroid extraction mechanism proposed in [5] to compute specific points within attention maps. Specifically, we calculate  $c_{h}$ ,  $c_{r}$  and  $c_{o}$  from each cross-attention maps as follows:",
+        "bbox": [
+            511,
+            320,
+            905,
+            409
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nc = \\frac {1}{\\sum_ {h , w} \\mathcal {A}} \\left[ \\begin{array}{l} \\sum_ {h, w} w \\cdot \\mathcal {A} \\\\ \\sum_ {h, w} h \\cdot \\mathcal {A} \\end{array} \\right], \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            607,
+            422,
+            905,
+            459
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $h$  and  $w$  denote the height and width in the attention map  $\\mathcal{A}$ . Then, we define the interaction center  $c_{rel}$  as the centroid of a triangle defined by  $c_h$ ,  $c_r$ , and  $c_o$ . Finally, we extract interaction region  $B_{rel}^{gen}$  by leveraging the distance between human and object centers with  $B_{rel}^{gen} = c_{rel} \\pm \\|c_h - c_o\\|_2^2$ . Additionally, we extract  $B_{rel}^{gt}$  in the same manner, leveraging the human and object bounding boxes provided in the dataset. We treat  $c_{rel}$  as the midpoint between the human and object centers from the given boxes  $B^{gt}$ .",
+        "bbox": [
+            511,
+            470,
+            905,
+            609
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Interaction Direction Guidance. To generate images with more realistic interactions, we design guidance that aligns feature differences at the image level with those at the interaction region. With interaction region  $B_{rel}^{gt}$  and  $B_{rel}^{gen}$ , we obtain interaction region image by masking the real and generated images:  $I_{rel}^{g} = B_{rel}^{g} \\odot I^{g}$ , where  $g \\in \\{gt, gen\\}$ . We then encode both interaction region images through a CLIP image encoder and extract interaction region features  $f_{rel}^{gt}$  and  $f_{rel}^{gen}$ . Leveraging interaction region features, we design the interaction direction guidance to focus the model on modifying the image feature, especially within the interaction region. Specifically, we calculate the direction between  $f_{rel}^{gt}$  and  $f_{rel}^{gen}$ , referred to as biased relation feature  $f_{rel}^{bias} = f_{rel}^{gt} - f_{rel}^{gen}$ , to align the direction of biased interaction features with that of the real images. The interaction direction guidance is as follows:",
+        "bbox": [
+            511,
+            609,
+            906,
+            853
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {I D G}} = 1 - \\frac {\\left(f _ {\\mathcal {M}} ^ {g t} - f ^ {g e n}\\right) \\cdot \\left(f _ {r e l} ^ {b i a s}\\right)}{\\left| f _ {\\mathcal {M}} ^ {g t} - f ^ {g e n} \\right| \\left| f _ {r e l} ^ {b i a s} \\right|}. \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            586,
+            867,
+            905,
+            906
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "8045",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4. Training Phase",
+        "text_level": 1,
+        "bbox": [
+            89,
+            90,
+            243,
+            107
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We train only the cross-attention layer in SD to capture the semantically distinct relation words, as the cross-attention reflects the existence of each word token [18]. We adopt the reconstruction loss (Eq. 1) to train T2I models in addition to relation disentanglement guidance and interaction direction guidance. However, when an image contains multiple humans and objects, it can lead to multiple interactions that do not match a single target interaction verb. To accurately separate the human and object corresponding to interaction words, we apply mask  $\\mathcal{M}$  when calculating the reconstruction loss as follows:",
+        "bbox": [
+            89,
+            113,
+            483,
+            279
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {r e c}} = \\mathbb {E} _ {z, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| (\\epsilon \\odot \\mathcal {M} - \\epsilon_ {\\theta} (z _ {t}, t, T) \\odot \\mathcal {M}) \\| _ {2} ^ {2} \\right]. \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            96,
+            285,
+            483,
+            325
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Finally, we combine reconstruction loss with disentanglement guidance and interaction direction guidance leveraging  $\\lambda_1, \\lambda_2$ , and  $\\lambda_3$  as below:",
+        "bbox": [
+            89,
+            335,
+            483,
+            380
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\cdot \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {2} \\cdot \\mathcal {L} _ {\\text {R D G}} + \\lambda_ {3} \\cdot \\mathcal {L} _ {\\text {I D G}}. \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            130,
+            391,
+            482,
+            407
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            417,
+            223,
+            435
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We train our model on SD v1.4 at a resolution of  $512 \\times 512$ . Using the Adam optimizer with a learning rate of  $4 \\times 10^{-6}$ , we trained for a single epoch over 17 hours with a batch size of 12 on 3 Nvidia A6000 GPUs. For image generation during training, we utilize the DDIM scheduler [27] with 30 sampling steps, and during inference, we increase sampling steps to 50. Finally,  $\\lambda_{1}, \\lambda_{2}$ , and  $\\lambda_{3}$  are set to 1.0, 10, and 0.8, respectively, and the margin  $m$  is set to 0.2.",
+        "bbox": [
+            89,
+            443,
+            483,
+            565
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Dataset",
+        "text_level": 1,
+        "bbox": [
+            89,
+            573,
+            186,
+            588
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We conduct our experiments on the HICO-DET dataset [2], which consists of 47,776 images (38,118 for training and 9,658 for testing). The dataset provides annotations for 600 HOI categories using 80 object categories and 117 verbs.",
+        "bbox": [
+            89,
+            595,
+            482,
+            657
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To train our model, we realign the training set according to the 600 HOI text descriptions and modify annotations to match each image with its corresponding text descriptions. This realignment yields 69,404 images across 600 prompts. We exclude text classes containing \"and\" (e.g. \"A photo of a person and an airplane\") to avoid ambiguity from various interactions within the same class, resulting in a final training set of 61,114 images across 501 prompts. More details about the dataset are in the supplementary materials.",
+        "bbox": [
+            89,
+            657,
+            483,
+            792
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Metrics",
+        "text_level": 1,
+        "bbox": [
+            89,
+            801,
+            186,
+            816
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "CLIP & Sentence-BERT Similarity. We assess the model understandability of interaction words by calculating text-to-text similarity scores using both CLIP [23] and SentenceBERT (S-BERT) [24] that we newly adopt to image generation task. After generating captions with InstructBLIP [19],",
+        "bbox": [
+            89,
+            824,
+            483,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "we compute cosine similarity against ground-truth texts. We also measure the similarity between the input prompts and generated images through CLIP. Details about the S-BERT metric are in the supplementary materials.",
+        "bbox": [
+            511,
+            90,
+            903,
+            151
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "HOI Classification Accuracy. To evaluate the robustness of the model understanding interaction words, we measure interaction accuracy with generated images using the SOV-STG [3] HOI detection model pre-trained on HICODET [2]. We report scores on two different weights, SOV-STG-S and SOV-STG-Swin-L, with different backbones (e.g. ResNet [8] and Swin-L [21]). Accuracy is reported in two main settings: Known Object (KO), which measures verb accuracy assuming the objects are correct, and Default (Def.), which assesses both verb and object correctness, with each setting further divided into Full (all classes) and Rare (classes with fewer than 10 samples).",
+        "bbox": [
+            511,
+            152,
+            905,
+            333
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "VQA-Score. We additionally use VQAScore [17], a large image-text alignment benchmark for generative models, to evaluate comprehensive interaction alignment scores. This score measures the model's ability to understand more complex compositional prompts, such as spatial and action relationships. For assessing the accuracy of depicting interaction between humans and objects in images, we use two main metrics: the Image-to-Text (I2T) alignment score and the VQA score. The I2T alignment score improves similarity measurement between complex prompts containing interaction words and generated images by incorporating a large language model. The VQA score measures the probability that a VQA model will answer \"yes\" to questions containing interaction words paired with generated images.",
+        "bbox": [
+            511,
+            334,
+            906,
+            546
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Qualitative Results",
+        "text_level": 1,
+        "bbox": [
+            511,
+            556,
+            696,
+            573
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Fig. 4 shows a comparison of generated images focusing on the interactions between humans and objects across our method and other models. Compared to previous approaches, our results exhibit more detailed interactions, closely resembling ground-truth images and producing high-quality outputs comparable to images from DALL-E 3 [1], a large-scale text-to-image generation model.",
+        "bbox": [
+            511,
+            580,
+            905,
+            686
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For example, in Fig. 4 (a)-(f), GLIGEN [15] mainly focuses on generating humans and objects, failing to depict the accurate interactions related to the interaction words. While InteractDiffusion [10] improves interaction representation, its results show inaccurate or ambiguous interactions that lack both accurate object rendering (e) or fine-grained interaction details (a)-(d) and (f).",
+        "bbox": [
+            511,
+            688,
+            905,
+            794
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In particular, in Fig.4 (g), our model generates a visually enhanced interaction that closely resembles the ground truth interaction region, showing fluid coming out from a bottle. In the multi-humans with one interaction scenario (f), our results depict people holding foods similar to the image from Dall-E 3, while other models only show people sitting at the table. This demonstrates that our model fur",
+        "bbox": [
+            511,
+            795,
+            905,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "8046",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/ea8e80d0108754070695e5b12030661ea5f49784a7303a807c350731f7f80887.jpg",
+        "image_caption": [
+            "Figure 4. Interaction comparison in generated images with other models. We generate images using a fixed template: “A {H} {R} a/an {O}”. The top row displays the input human, interaction word, and object. Green and blue boxes represent additional grounding boxes used during image generation for each human and object, respectively. Our results produce images with more accurate interactions than other models, closely resembling ground-truth interactions while maintaining high image quality comparable to large-scale generative models like DALL-E 3 [1]."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            135,
+            89,
+            864,
+            592
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "ther concentrates on detailed interaction regions and distinguishes subtle differences between interaction words with similar spatial arrangements (e.g. \"sitting at\" and \"eating at\"). Overall, our model demonstrates a similar level of interaction understandability as a large-scale model such as Dall-E 3 and more accurately exhibits the intended interactions, closely matching with the ground-truth interactions.",
+        "bbox": [
+            89,
+            698,
+            482,
+            803
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We further evaluate our model with complex prompts with multiple interactions to test the ability to distinguish interaction words. As shown in Fig. 1, when given multiple interactions, our model accurately depicts each one, while other models struggle. For example, GLIGEN places objects correctly but fails to represent interactions like \"walk",
+        "bbox": [
+            89,
+            810,
+            482,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "ing,\" \"jumping,\" and \"throwing.\" InteractDiffusion captures some interactions but misses key objects, like an umbrella. In contrast, our model generates accurate images that reflect the intended interactions. This shows that our model successfully captures the semantic difference between interaction words, generating more detailed interactions. More qualitative results are in the supplementary materials.",
+        "bbox": [
+            511,
+            698,
+            906,
+            806
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4.Quantitative Results",
+        "text_level": 1,
+        "bbox": [
+            511,
+            816,
+            707,
+            830
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Tab. 1 compares similarity scores between VerbDiff and previous methods. Our model achieves the highest scores across nearly all evaluation settings. In the text-to-text (T2T) similarity evaluation, VerbDiff outperforms existing",
+        "bbox": [
+            511,
+            839,
+            906,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "8047",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/57221e730efcd1ef30aae6a68075f988da187b941201dc32c005926df7363e1e.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Models</td><td colspan=\"2\">CLIP</td><td>S-BERT</td><td colspan=\"2\">VQA-Score</td></tr><tr><td>T2T</td><td>T2I</td><td>T2T</td><td>I2T</td><td>VQA</td></tr><tr><td>HICO-DET</td><td>0.604</td><td>0.233</td><td>0.446</td><td>0.755</td><td>0.754</td></tr><tr><td>SD [25]</td><td>0.725</td><td>0.251</td><td>0.620</td><td>0.769</td><td>0.765</td></tr><tr><td>GLIGEN [15]</td><td>0.683</td><td>0.238</td><td>0.554</td><td>0.679</td><td>0.674</td></tr><tr><td>InteractDiffusion [10]</td><td>0.703</td><td>0.232</td><td>0.575</td><td>0.728</td><td>0.734</td></tr><tr><td>VerbDiff (Ours)</td><td>0.733</td><td>0.242</td><td>0.633</td><td>0.771</td><td>0.766</td></tr></table>",
+        "bbox": [
+            94,
+            88,
+            482,
+            196
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/61ccea430662648917cb6f95db509c14a26c15baabfab8e1b6beba49e5b4fcaf.jpg",
+        "table_caption": [
+            "Table 1. Similarity comparison between VerbDiff and other models. We evaluate scores on CLIP, S-BERT [24] and a large vision-language alignment benchmark VQA-Score [17]."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"3\">Models</td><td colspan=\"4\">SOV-STG-S (Acc ↑)</td><td colspan=\"4\">SOV-STG-Swin-L (Acc ↑)</td></tr><tr><td colspan=\"2\">Def.</td><td colspan=\"2\">KO.</td><td colspan=\"2\">Def.</td><td colspan=\"2\">KO.</td></tr><tr><td>Full</td><td>Rare</td><td>Full</td><td>Rare</td><td>Full</td><td>Rare</td><td>Full</td><td>Rare</td></tr><tr><td>HICO-DET</td><td>26.52</td><td>6.78</td><td>28.68</td><td>7.29</td><td>29.98</td><td>12.66</td><td>31.16</td><td>13.43</td></tr><tr><td>SD [25]</td><td>16.09</td><td>4.59</td><td>18.22</td><td>4.85</td><td>20.08</td><td>8.07</td><td>21.69</td><td>8.66</td></tr><tr><td>GLIGEN [15]</td><td>15.88</td><td>4.85</td><td>17.91</td><td>5.24</td><td>17.83</td><td>7.00</td><td>19.35</td><td>7.57</td></tr><tr><td>InteractDiffusion [10]</td><td>19.67</td><td>7.00</td><td>21.31</td><td>7.69</td><td>23.53</td><td>10.27</td><td>24.86</td><td>11.18</td></tr><tr><td>VerbDiff (Ours)</td><td>22.59</td><td>7.62</td><td>24.79</td><td>7.83</td><td>27.05</td><td>12.60</td><td>28.43</td><td>13.18</td></tr></table>",
+        "bbox": [
+            93,
+            263,
+            483,
+            371
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "methods, and the difference becomes even more significant in the S-BERT evaluation. This highlights that the S-BERT metric better captures the semantic differences within interaction verbs. Although the text-to-image (T2I) score is lower than SD, our model achieves higher scores than the models that leverage additional bounding boxes, demonstrating the ability of our model to understand interactions. To assess a more comprehensive interaction understanding of VerbDiff, we also report the VQA-Score, which evaluates how well generative models understand more complex compositional prompts, such as interactions. Although our method shows a slight improvement in the I2T and VQA scores compared to SD, it scores significantly higher than InteractDiffusion.",
+        "bbox": [
+            88,
+            438,
+            482,
+            648
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Tab. 2 presents the HOI accuracy scores, where VerbDiff achieves the highest accuracy across all settings compared to previous methods. In particular, although different backbones are used for computing accuracy, VerbDiff consistently shows the highest accuracy across all settings. This demonstrates that our method has robust interaction word understanding and produces images with precise interactions. Overall, VerbDiff consistently outperforms other models across nearly all evaluation metrics, demonstrating its strong ability to comprehend interaction words and generate high-quality images with accurate human-object interactions, even without extra conditions.",
+        "bbox": [
+            89,
+            651,
+            482,
+            833
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.5. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            89,
+            845,
+            243,
+            862
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We show the effectiveness of RDG and IDG in Tab. 3 and Fig. 5. As shown, both RDG (w/  $\\mathcal{L}_{\\text {triple }}$  and  $\\mathcal{L}_{\\text {align }}$ ) (b)",
+        "bbox": [
+            89,
+            869,
+            482,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b4fde7740f8d83a829e41e22ad41dfce07c740751338bab4626517108a7f9e31.jpg",
+        "image_caption": [
+            "Figure 5. Image interaction comparison with and without guidance. IDG focuses on the specific localized interaction region (red box) in the image."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            516,
+            90,
+            905,
+            178
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/e97177ea776be776e0d5861768a718fdd0c31627276ff1db95888fe46516921c.jpg",
+        "table_caption": [
+            "Table 2. HOI accuracy comparison between VerbDiff and previous methods. Def. and KO. refer to Default and Known Object."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>VerbDiff</td><td>Lrec</td><td>Ltriple</td><td>Lalign</td><td>LIDG</td><td>CLTP2T</td><td>S-BERT T2T</td><td>HOI Def.</td><td>Acc KO.</td></tr><tr><td>(a)</td><td>✓</td><td></td><td></td><td></td><td>0.691</td><td>0.582</td><td>19.38</td><td>20.89</td></tr><tr><td>(b)</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>0.700</td><td>0.589</td><td>20.32</td><td>21.87</td></tr><tr><td>(c)</td><td>✓</td><td></td><td></td><td>✓</td><td>0.699</td><td>0.588</td><td>20.21</td><td>21.67</td></tr><tr><td>(d)</td><td>✓</td><td>✓</td><td></td><td>✓</td><td>0.710</td><td>0.610</td><td>23.39</td><td>24.51</td></tr><tr><td>All (Ours)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>0.733</td><td>0.633</td><td>27.05</td><td>28.43</td></tr></table>",
+        "bbox": [
+            519,
+            246,
+            903,
+            329
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Ablation of the VerbDiff guidance settings. We score the similarity and HOI accuracy on the Full setting. We use the SOV-STG-Swin-L model. Combining all the proposed loss functions shows the best performance.",
+        "bbox": [
+            511,
+            339,
+            906,
+            395
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "and IDG (c) enhance the similarity score and HOI accuracy. Specifically, (b) scores slightly higher than (c) demonstrating that the model RDG distinguishes the semantics in interaction words. While scores improve significantly when using IDG and RDG without  $\\mathcal{L}_{\\text {align }}$ , the model with all guidance terms achieves the highest overall performance. In Fig. 5, we compare images generated with and without each type of guidance. Notably, the enlarged interaction region (red box) shows that the model with IDG produces detailed representations (e.g. human hands), unlike models with only RDG. This indicates that IDG effectively captures fine-grained details in localized interaction regions. More ablations are in the supplementary materials.",
+        "bbox": [
+            511,
+            422,
+            906,
+            619
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            513,
+            633,
+            633,
+            650
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We propose VerbDiff, a novel text-to-image (T2I) diffusion models that address the interaction misunderstanding problem in SD without additional conditions. Although previous methods leverage additional conditions to help the model understand the interaction between humans and objects, they still rely on precise instructions and lack an understanding of the semantic differences between interaction verbs. Our model successfully captures the semantic meanings inherent in interaction words and generates high-quality images with accurate interactions. Extensive experiments demonstrate the effectiveness of our method in enhancing the ability to understand interaction words of T2I models, achieving better interaction comprehension compared to the previous approaches.",
+        "bbox": [
+            511,
+            660,
+            906,
+            871
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments. This was supported by the Institute of Information & Communications Technology Plan-",
+        "bbox": [
+            511,
+            872,
+            905,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "8048",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "ning & Evaluation (IITP) grant funded by the Korean government(MSIT) (No.RS-2020-II201373, Artificial Intelligence Graduate School Program(Hanyang University)).",
+        "bbox": [
+            89,
+            90,
+            485,
+            135
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            91,
+            162,
+            187,
+            178
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 6, 7",
+            "[2] Yu-Wei Chao, Yunfan Liu, Xieyang Liu, Huayi Zeng, and Jia Deng. Learning to detect human-object interactions. In 2018 IEEE winter conference on applications of computer vision (wacv), pages 381-389. IEEE, 2018. 2, 6",
+            "[3] Junwen Chen, Yingcheng Wang, and Keiji Yanai. Focusing on what to decode and what to train: Efficient training with hoi split decoders and specific target guided denoising. arXiv preprint arXiv:2307.02291, 2023. 6",
+            "[4] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 5",
+            "[5] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 5",
+            "[6] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2, 3",
+            "[7] Hanan Gani, Shariq Farooq Bhat, Muzammal Naseer, Salman Khan, and Peter Wonka. Lm blueprint: Enabling text-to-image generation with complex and detailed prompts. arXiv preprint arXiv:2310.10640, 2023. 2, 3",
+            "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6",
+            "[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 5",
+            "[10] Jiu Tian Hoe, Xudong Jiang, Chee Seng Chan, Yap-Peng Tan, and Weipeng Hu. Interactdiffusion: Interaction control in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6180-6189, 2024. 1, 2, 3, 6, 8",
+            "[11] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In Similarity-based pattern recognition: third international workshop, SIMBAD 2015, Copenhagen, Denmark, October 12-14, 2015. Proceedings 3, pages 84-92. Springer, 2015. 4",
+            "[12] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan,"
+        ],
+        "bbox": [
+            93,
+            186,
+            482,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 3",
+            "[13] Xu Jia, Takashi Isobe, Xiaomin Li, Qinghe Wang, Jing Mu, Dong Zhou, Huchuan Lu, Lu Tian, Ashish Sirasao, Emad Barsoum, et al. Customizing text-to-image generation with inverted interaction. In ACM Multimedia 2024, 2024. 3",
+            "[14] Jian-Yu Jiang-Lin, Kang-Yang Huang, Ling Lo, Yi-Ning Huang, Terence Lin, Jhih-Ciang Wu, Hong-Han Shuai, and Wen-Huang Cheng. Record: Reasoning and correcting diffusion for hoi generation. arXiv preprint arXiv:2407.17911, 2024.3",
+            "[15] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 1, 2, 3, 6, 8",
+            "[16] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2, 3",
+            "[17] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2025. 6, 8",
+            "[18] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7817-7826, 2024. 6",
+            "[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6",
+            "[20] Jinxiu Liu and Qi Liu. R3cd: Scene graph to image generation with relation-aware compositional contrastive control diffusion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3657-3665, 2024. 3",
+            "[21] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6",
+            "[22] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15579-15591, 2023. 2, 3",
+            "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4, 6",
+            "[24] N Reimers. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 2, 6, 8"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "8049",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 8",
+            "[26] Guibao Shen, Luozhou Wang, Jiantao Lin, Wenhang Ge, Chaozhe Zhang, Xin Tao, Yuan Zhang, Pengfei Wan, Zhongyuan Wang, Guangyong Chen, et al. Sg-adapter: Enhancing text-to-image generation with scene graph guidance. arXiv preprint arXiv:2405.15321, 2024. 3",
+            "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6",
+            "[28] Shengqiong Wu, Hao Fei, Hanwang Zhang, and Tat-Seng Chua. Imagine that! abstract-to-intricate text-to-image synthesis with scene graph hallucination diffusion. Advances in Neural Information Processing Systems, 36, 2024. 3",
+            "[29] Yinwei Wu, Xingyi Yang, and Xinchao Wang. Relation rectification in diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7685-7694, 2024. 3",
+            "[30] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 2, 3",
+            "[31] Bicheng Xu, Qi Yan, Renjie Liao, Lele Wang, and Leonid Sigal. Joint generative modeling of scene graphs and images via diffusion models. arXiv preprint arXiv:2401.01130, 2024. 3",
+            "[32] Ling Yang, Zhilin Huang, Yang Song, Shenda Hong, Guohao Li, Wentao Zhang, Bin Cui, Bernard Ghanem, and Ming-Hsuan Yang. Diffusion-based scene graph to image generation with masked contrastive pre-training. arXiv preprint arXiv:2211.11138, 2022. 3",
+            "[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3",
+            "[34] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2, 3"
+        ],
+        "bbox": [
+            91,
+            92,
+            485,
+            726
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "8050",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2025/VerbDiff_ Text-Only Diffusion Models with Enhanced Interaction Awareness/ad57dfe8-0121-4da2-a044-841b75f6e661_model.json

@@ -0,0 +1,1793 @@
+[
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.107,
+                0.003,
+                0.182,
+                0.044
+            ],
+            "angle": 0,
+            "content": "CVF"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.237,
+                0.001,
+                0.812,
+                0.047
+            ],
+            "angle": 0,
+            "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.113,
+                0.131,
+                0.885,
+                0.152
+            ],
+            "angle": 0,
+            "content": "VerbDiff: Text-Only Diffusion Models with Enhanced Interaction Awareness"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.168,
+                0.181,
+                0.831,
+                0.217
+            ],
+            "angle": 0,
+            "content": "SeungJu Cha Kwanyoung Lee Ye-Chan Kim Hyunwoo Oh Dong-Jin Kim Hanyang University, South Korea"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.234,
+                0.219,
+                0.756,
+                0.235
+            ],
+            "angle": 0,
+            "content": "{sju9020, mobled37, dpcksd178, komji, djdkim}@hanyang.ac.kr"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.108,
+                0.275,
+                0.891,
+                0.576
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.223,
+                0.578,
+                0.776,
+                0.593
+            ],
+            "angle": 0,
+            "content": "\"A basketball player jumping and throwing a basketball and a man blocking the ball\""
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.089,
+                0.614,
+                0.908,
+                0.657
+            ],
+            "angle": 0,
+            "content": "Figure 1. Generated samples illustrating multiple human-object interactions. Each color represents distinct humans, objects, and interaction words. GLIGEN [15] and InteractDiffusion [10] use grounding boxes as additional conditions, whereas Stable Diffusion [25] and VerbDiff rely solely on text."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.248,
+                0.669,
+                0.329,
+                0.685
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.705,
+                0.486,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Recent large-scale text-to-image diffusion models generate photorealistic images but often struggle to accurately depict interactions between humans and objects due to their limited ability to differentiate various interaction words. In this work, we propose VerbDiff to address the challenge of capturing nuanced interactions within text-to-image diffusion models. VerbDiff is a novel text-to-image generation model that weakens the bias between interaction words and objects, enhancing the understanding of interactions. Specifically, we disentangle various interaction words from frequency-based anchor words and leverage localized interaction regions from generated images to help the model better capture semantics in distinctive words without extra"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.671,
+                0.909,
+                0.763
+            ],
+            "angle": 0,
+            "content": "conditions. Our approach enables the model to accurately understand the intended interaction between humans and objects, producing high-quality images with accurate interactions aligned with specified verbs. Extensive experiments on the HICO-DET dataset demonstrate the effectiveness of our method compared to previous approaches."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.798,
+                0.646,
+                0.813
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.826,
+                0.907,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Recently, text-to-image diffusion models have demonstrated the ability to produce photorealistic images from natural texts [25]. However, these models sometimes struggle to accurately depict interactions between humans and objects [10]. For example, they often fail to differentiate"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.514,
+                0.957
+            ],
+            "angle": 0,
+            "content": "8041"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.097,
+                0.092,
+                0.481,
+                0.304
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.09,
+                0.315,
+                0.485,
+                0.386
+            ],
+            "angle": 0,
+            "content": "Figure 2. Examples of interactions from InteractDiffusion [10], showing a lack of understanding of interaction words. The model relies on precise bounding boxes rather than understanding interaction words to exhibit accurate interactions, as shown in the results for the \"red\" and \"blue\" boxes."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.414,
+                0.484,
+                0.476
+            ],
+            "angle": 0,
+            "content": "between semantically distinct prompts, such as \"A person walking a bicycle\" and \"A person riding a bicycle,\" particularly in capturing the semantics in verbs that are crucial for accurately depicting the intended interaction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.476,
+                0.484,
+                0.612
+            ],
+            "angle": 0,
+            "content": "We hypothesize that this problem arises from the strong object bias of CLIP [23], which fails to adequately capture semantic differences between verbs [22]. Specifically, CLIP [23] often focuses on objects in prompts, overlooking semantic distinctions in interaction verbs. To enhance prompt interpretation ability of models, previous methods have leveraged additional modules such as large language models (LLMs) [6, 7, 16] or extra bounding boxes [10, 15, 30, 34] to provide explicit relational visual information."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.613,
+                0.484,
+                0.794
+            ],
+            "angle": 0,
+            "content": "For example, InteractDiffusion [10] introduces controllable image generation using Human-Object Interaction (HOI) information with bounding boxes for human, relation, and object \\(<\\mathrm{h}, \\mathrm{r}, \\mathrm{o}>\\) triplets. Nevertheless, as shown in Fig. 2, the model still struggles with accurate interactions even with these additional conditions. For instance, given the prompt \"walking, bicycle,\" the model fails to depict the intended interaction when semantically distinctive interaction words share a similar bounding box (e.g. walking and riding). This suggests that the model lacks an understanding of interaction semantics and relies heavily on precise bounding boxes, which are labor-intensive to provide."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.796,
+                0.484,
+                0.903
+            ],
+            "angle": 0,
+            "content": "To improve the model's interaction understandability without additional conditions, we propose VerbDiff, a novel text-to-image generation method that better distinguishes interaction words via several objectives. First, we apply Relation Disentanglement Guidance (RDG) using frequency-based anchor texts from the dataset for each human-object pair to reduce interaction bias in generated images. We"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.092,
+                0.907,
+                0.259
+            ],
+            "angle": 0,
+            "content": "observe that the exhibited interactions in generated images have bias, which follows frequent verbs corresponding to each human-object pair in the data distribution. To address this, we align interaction features between real and generated images while separating those of generated images from anchor text features, enabling a more effective distinction of interaction words. Additionally, we introduce a balancing factor to address the extreme long-tail distribution of interaction words across human-object pairs, adjusting interaction modification during training. This allows the model to learn diverse interactions in a balanced way."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.27,
+                0.908,
+                0.662
+            ],
+            "angle": 0,
+            "content": "Secondly, to enforce the model to better focus on the interaction regions between humans and objects, we introduce Interaction Direction Guidance (IDG). This approach emphasizes fine-grained semantic distinctions in localized interactions. We design an Interaction Region (IR) module to capture specific interaction areas in generated images, leveraging cross-attention maps in text-only diffusion models. The IR module extracts interaction regions using the centroids of cross-attention maps associated with \\(<\\mathrm{h}\\), r, o> tokens. We obtain biased interaction features from these regions and apply interaction direction guidance. This guidance effectively modifies interaction regions within images, guiding Stable Diffusion (SD) to concentrate more accurately on these regions while capturing nuanced interaction details. We validate our method's effectiveness on the HICO-DET dataset [2]. Our approach accurately captures semantic differences between interaction words and generates high-quality images with accurate interactions compared to SD [25], InteractDiffusion [10], and other prior methods. Additionally, we introduce a new metric using Sentence BERT [24] for image generation tasks. This metric offers improved capability in discerning subtle and nuanced meanings of interaction words compared to CLIP. By focusing on semantic distinctions in interactions, this metric enables more precise evaluation, complementing CLIP similarity. Our contributions are as follows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.675,
+                0.909,
+                0.901
+            ],
+            "angle": 0,
+            "content": "(1) We introduce VerbDiff, a novel text-to-image model that enhances the semantic interaction understanding by distinguishing interaction verbs and generating images with accurate human-object interactions without any additional conditions. (2) To reduce generating images with biased interactions, we apply relation disentanglement guidance using frequency-based anchor text extracted from real data distributions. (3) Our proposed IR module captures detailed interaction regions using cross-attention map centroids, providing specified-interaction region guidance without explicit bounding boxes. (4) We implement a Sentence-BERT metric to effectively capture semantic differences between interaction sentences, addressing CLIP's limitations. Extensive experiments demonstrate that our approach outperforms previous methods across various settings."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "8042"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.092,
+                0.09,
+                0.234,
+                0.105
+            ],
+            "angle": 0,
+            "content": "2. Related Work"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.115,
+                0.26,
+                0.131
+            ],
+            "angle": 0,
+            "content": "2.1. Diffusion Models"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.139,
+                0.483,
+                0.26
+            ],
+            "angle": 0,
+            "content": "Recently, diffusion models have achieved state-of-the-art performance in text-to-image generation [25] and controllable image generation [10, 15, 33]. While Stable Diffusion (SD) [25] integrates conditioning embeddings into cross-attention layers for text-to-image generation, it often struggles to fully align generated images with text instructions. This arises from inherent limitations of texts to convey exact visual information and strong object bias of CLIP [22]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.262,
+                0.484,
+                0.442
+            ],
+            "angle": 0,
+            "content": "To address these, recent research has focused on controllable image generation using bounding boxes [15, 30, 34]. For example, GLIGEN [15] extends SD to incorporate grounding inputs such as bounding boxes, image prompts, key points, and spatially aligned conditions. LayoutDiffusion [34] enables multiple bounding box inputs for layout-specific generation. Although these methods better generate images, they require additional conditions beyond text input to achieve the intended image outputs. In contrast, our proposed method achieves comparable results without additional grounding information, providing a more efficient alternative for precise and flexible generation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.454,
+                0.285,
+                0.469
+            ],
+            "angle": 0,
+            "content": "2.2. Relation Generation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.478,
+                0.483,
+                0.629
+            ],
+            "angle": 0,
+            "content": "Relation generation between objects has been explored leveraging LLMs [6, 7, 16] and scene graph [20, 26, 28, 29, 31, 32]. While LLM-based methods can generate accurate object layouts, they primarily focus on compositional relations rather than interactions [6, 7, 16]. Scene graph-based approaches have addressed these limitations by combining scene graph encoders to enhance prompt understandability [20, 26, 28, 29, 31, 32]. However, they require additional training to extract scene graphs from prompts and struggle to capture subtle semantic differences in interactions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.63,
+                0.483,
+                0.902
+            ],
+            "angle": 0,
+            "content": "To depict accurate interactions during generation, researchers have focused on human-object interactions (HOI) [10, 12-14]. Some methods use inversion-based frameworks to capture interaction semantics but require optimization for each interaction word [12, 13]. More generalized approaches leverage additional conditions such as human poses [14] or bounding boxes [10]. Specifically, InteractDiffusion [10] leverages the HOI information with bounding boxes to generate interactions. However, this method still relies on additional information rather than semantic differences between prompts. Our work is similar to InteractDiffusion [10] by leveraging HOI information but differs from focusing on semantic differences between interaction words to enhance interaction understanding in SD. This approach aims to improve the generation of nuanced interactions without relying on additional conditions. It could potentially lead to more accurate and diverse depictions of human-object interactions in synthesized images."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.09,
+                0.605,
+                0.105
+            ],
+            "angle": 0,
+            "content": "3. Method"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.115,
+                0.907,
+                0.358
+            ],
+            "angle": 0,
+            "content": "Our goal is to generate images with accurate interactions by enhancing the model understandability of the interaction word without additional conditions. Our proposed method, VerbDiff as illustrated in Fig. 3 contains two main parts: (1) Relation disentanglement guidance leverages frequency-based anchor texts from human and object pairs along with input texts and real and generated image features. This guidance helps visually align the generated interaction features with real ones and distinguish the semantic differences between interaction words. (2) Interaction direction guidance that helps the model concentrate more on the interaction region while modifying the image features. To capture the interaction region in arbitrarily generated images, we further design an interaction region module (IR module) to extract a centroid-based interaction region from the cross-attention map corresponding to \\( < \\mathrm{h},\\mathrm{r},\\mathrm{o} > \\) tokens."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.366,
+                0.644,
+                0.383
+            ],
+            "angle": 0,
+            "content": "3.1. Preliminary"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.389,
+                0.907,
+                0.556
+            ],
+            "angle": 0,
+            "content": "SD Training Objective. Text-to-image generation models, such as Stable Diffusion (SD) [25], show photo-realistic images that align with given text prompts through an efficient two-stage training scheme. The first stage learns to map the input images \\(I\\) into latent space \\(z\\). The second stage learns the denoising UNet to gradually predict the real noise in the diffusion forward process. Starting from the original latent image \\(z_0\\), the diffusion model gradually adds noise \\(\\epsilon\\) to produce noisy sample \\(z_t\\) through uniformly sampled time steps \\(t\\). Then, with text condition \\(T\\), the UNet predicts the noise \\(\\epsilon_{\\theta}\\) for each time step to reconstruct \\(z_t\\) to \\(z_0\\) as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.528,
+                0.565,
+                0.907,
+                0.593
+            ],
+            "angle": 0,
+            "content": "\\[\n\\min  _ {\\theta} \\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {z, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| (\\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, T)) \\| _ {2} ^ {2} \\right]. \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.602,
+                0.906,
+                0.664
+            ],
+            "angle": 0,
+            "content": "Cross-Attention Layer. When given text, the CLIP [23] text encoder produces text embedding, and the UNet in SD takes embedding and computes cross-attention between encoded text \\( e \\) and intermediate image features \\( \\phi(z_{t}) \\):"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.623,
+                0.673,
+                0.907,
+                0.71
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {A} = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d}}\\right), \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.718,
+                0.906,
+                0.824
+            ],
+            "angle": 0,
+            "content": "where query \\(Q = W_{Q} \\cdot \\phi(z_{t})\\) and key \\(K = W_{K} \\cdot e\\) are computed with projection matrix \\(W_{Q}\\), \\(W_{K}\\) and \\(d\\) is the dimension of \\(Q\\). These cross-attention maps \\(\\mathcal{A}\\) enable the text-to-image diffusion models to depict texts into image features. In addition, each \\(\\mathcal{A}_{j}\\), a cross-attention map corresponding to \\(j^{th}\\) token in prompts, represents the attention weight on intermediate image features in various resolutions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.833,
+                0.828,
+                0.849
+            ],
+            "angle": 0,
+            "content": "3.2. Relation Disentanglement Guidance"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.856,
+                0.906,
+                0.901
+            ],
+            "angle": 0,
+            "content": "To depict accurate interaction in generated images without additional conditions, we first focus on the semantic difference between interaction words, especially the verbs. As the"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "8043"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.099,
+                0.089,
+                0.67,
+                0.399
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.682,
+                0.089,
+                0.902,
+                0.4
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.089,
+                0.41,
+                0.908,
+                0.468
+            ],
+            "angle": 0,
+            "content": "Figure 3. Pipeline of VerbDiff. VerbDiff uses Relation Disentanglment Guidance (Sec. 3.2) that separates the interaction features from the anchor text for each human-object pair. Additionally, it contains the IR module (Sec. 3.3) (right) which extracts localized interaction regions from generated images without explicit bounding boxes and Interaction Direction Guidance (Sec. 3.3) that guides the model to focus more on the fine-grained interaction regions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.493,
+                0.482,
+                0.584
+            ],
+            "angle": 0,
+            "content": "SD is known to have limitations in capturing the semantic nuances in interaction words, we apply relation disentangle-ment guidances. The goal of this guidance is to alleviate the relation bias within human-object pairs and improve interaction comprehension through anchor text, input text, and real and generated images."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.585,
+                0.483,
+                0.706
+            ],
+            "angle": 0,
+            "content": "First, given the ground-truth text \\( T^{gt} = \\) \"A photo of a \\( \\{h\\} \\{r^{gt}\\} a / an\\{o\\} \\)\" (e.g. \"A photo of a person holding a backpack\"), SD generates images \\( I^{gen} \\) during training. Although \\( T^{gt} \\) specifies the target interaction, \\( I^{gen} \\) often contains inaccurate and biased interactions that fail to meet verbs between each human and object pair. We define a sentence with accurate interaction verbs corresponding to inaccurate interaction image \\( I^{gen} \\) as \\( T^{anc} \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.707,
+                0.483,
+                0.858
+            ],
+            "angle": 0,
+            "content": "In particular, we observe biased interaction images tend to depict verbs that appear more frequently in the data along human-object pair. For instance, when generating images with the object \"backpack,\" the images commonly show a person wearing a backpack, which is the most frequent verb in the dataset. Based on this observation, we hypothesize that the biased interaction images typically align with frequent verbs in the real data distribution. Therefore, we define \\( r^{anc} \\) as the verb with the highest sample count for each human-object pair as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.203,
+                0.872,
+                0.483,
+                0.898
+            ],
+            "angle": 0,
+            "content": "\\[\nr ^ {a n c} = \\underset {r \\in R _ {o}} {\\arg \\max } \\mathcal {C} (r | o), \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.493,
+                0.907,
+                0.539
+            ],
+            "angle": 0,
+            "content": "where \\( R_{o} \\) denotes the set of verbs related to each human-object pair in dataset \\( \\mathcal{D} \\) and \\( \\mathcal{C}(\\cdot) \\) represents the number of samples corresponding to each verb \\( r \\) when given \\( o \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.54,
+                0.907,
+                0.646
+            ],
+            "angle": 0,
+            "content": "Given the anchor verb \\( r^{anc} \\), to reduce the generation of biased interaction images, we encourage SD to create images that with distant semantics from \\( T^{anc} = \"A photo of a \\{h\\} \\{r^{anc}\\} a/an \\{o\\}\" \\) (e.g. \"wearing a backpack\"). Specifically, we guide the \\( I^{gen} \\) closer to the accurate interaction sentence \\( T^{gt} \\) while pushing away from the sentence with anchor interaction \\( T^{anc} \\) by leveraging triplet loss [11]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.646,
+                0.909,
+                0.782
+            ],
+            "angle": 0,
+            "content": "We extract normalized feature \\( f^{gen} \\), \\( e^{gt} \\), and \\( e^{anc} \\) through CLIP [23] encoder \\( E_{T} \\) and \\( E_{I} \\) from \\( I^{gen} \\), \\( T^{gt} \\), and \\( T^{anc} \\), respectively. Then, we set \\( e^{gt} \\) as positive and \\( e^{anc} \\) as negative when given anchor \\( f^{gen} \\) and compute the cosine similarity with the generated image features, \\( f^{gen} \\), and each text feature, \\( e^{gt} \\) and \\( e^{anc} \\). Finally, we introduce a triplet loss to maximize the cosine similarity between \\( f^{gen} \\) and \\( e^{gt} \\) larger than the similarity between \\( f^{gen} \\) and \\( e^{anc} \\) with a margin of \\( m \\) as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.513,
+                0.794,
+                0.908,
+                0.827
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\text {t r i p l e}} = \\max  \\left(0, m + \\operatorname {s i m} \\left(f ^ {g e n}, e ^ {g t}\\right) - \\operatorname {s i m} \\left(f ^ {g e n}, e ^ {a n c}\\right)\\right). \\tag {4}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.84,
+                0.909,
+                0.901
+            ],
+            "angle": 0,
+            "content": "In addition to distinguishing the semantic difference between verbs, to generate accurate interaction at the image level as well, we add image aligning loss to make the model follow the interaction in \\( I^{gt} \\). However, \\( I^{gt} \\) often contains"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "8044"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.092,
+                0.485,
+                0.273
+            ],
+            "angle": 0,
+            "content": "multiple humans, objects, and interactions within a single image while we only use a single human, object, and interaction triplet as input during image generation. This input discrepancy makes it difficult for the model to focus solely on the relevant interaction specified by the input prompts, causing confusion when depicting the specified interaction words. To focus on the part in the image corresponding to the interaction label, we extract mask region \\(\\mathcal{M}\\) combining the human and object bounding boxes in a dataset and obtain \\(I_{\\mathcal{M}}^{gt} = I^{gt} \\odot \\mathcal{M}\\). Then, we extract \\(f_{\\mathcal{M}}^{gt}\\) through image encoder and minimize the cosine similarity between \\(f_{\\mathcal{M}}^{gt}\\) and \\(f^{gen}\\) as below:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.201,
+                0.284,
+                0.484,
+                0.322
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = 1 - \\frac {f _ {\\mathcal {M}} ^ {g t} \\cdot f ^ {g e n}}{| f _ {\\mathcal {M}} ^ {g t} | | f ^ {g e n} |}. \\tag {5}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.329,
+                0.484,
+                0.465
+            ],
+            "angle": 0,
+            "content": "Adaptive Interaction Modification. When applying \\(\\mathcal{L}_{\\mathrm{triple}}\\) and \\(\\mathcal{L}_{\\mathrm{align}}\\) to capture the semantic differences between interaction words, we find that the modification extent applied to the generated images differs across interaction words. As each interaction text \\(T^{gt}\\) has widely varying sample counts, we also focus on balancing the modification extent across interaction words. Specifically, we adaptively scale the loss considering the number of samples following [4]. We apply an effective number, which is defined as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.226,
+                0.473,
+                0.483,
+                0.505
+            ],
+            "angle": 0,
+            "content": "\\[\n\\alpha (k) = \\frac {1 - \\beta^ {n _ {k}}}{1 - \\beta}, \\tag {6}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.514,
+                0.484,
+                0.544
+            ],
+            "angle": 0,
+            "content": "where \\(\\beta = (N - 1) / N\\), \\(n_k\\) is the number of samples in class \\(k\\), and \\(\\mathbf{N}\\) is the total number of samples in the dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.545,
+                0.484,
+                0.589
+            ],
+            "angle": 0,
+            "content": "Relation Disentanglement Guidance. Finally, we multiply \\(\\alpha\\) on \\(\\mathcal{L}_{\\mathrm{triple}}\\) and \\(\\mathcal{L}_{\\mathrm{align}}\\), defining relation disentanglement guidance as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.19,
+                0.601,
+                0.483,
+                0.619
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\mathrm {R D G}} = \\alpha \\cdot \\left(\\mathcal {L} _ {\\text {t r i p l e}} + \\mathcal {L} _ {\\text {a l i g n}}\\right). \\tag {7}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.627,
+                0.484,
+                0.672
+            ],
+            "angle": 0,
+            "content": "This guidance induces SD to generate images with accurate interactions for each verb and to distinguish the semantic difference between interactions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.682,
+                0.482,
+                0.697
+            ],
+            "angle": 0,
+            "content": "3.3. IR Module & Interaction Direction Guidance"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.704,
+                0.483,
+                0.855
+            ],
+            "angle": 0,
+            "content": "Relation Disentanglement Guidance (RDG) effectively enables the model to better distinguish and represent each interaction word, helping to exhibit the intended interactions at the image level. However, the model sometimes shows images that misalign with human expectations, especially in the local regions where human-object interactions occur. We hypothesize that focusing on finer, localized interaction regions helps the model improve the detailed interaction expressions, thereby allowing the model to better capture more region-specific interaction information."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.856,
+                0.484,
+                0.901
+            ],
+            "angle": 0,
+            "content": "To focus more on specific interaction regions, we apply interaction direction guidance (IDG), leveraging interaction regions extracted from the interaction region module with"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.092,
+                0.906,
+                0.152
+            ],
+            "angle": 0,
+            "content": "cross-attention maps corresponding to the \\(<\\mathrm{h}, \\mathrm{r}, \\mathrm{o}>\\) tokens in generated images. IDG enhances the model's ability to depict region-level details, guiding the depicted interactions to align more perceptually close to human intentions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.153,
+                0.907,
+                0.32
+            ],
+            "angle": 0,
+            "content": "Interaction Region Module. To achieve a more detailed interaction expression, we first extract interaction regions from both real and generated images. Although real images contain explicit human and object bounding boxes in the dataset, it is challenging to extract specific interaction regions without additional conditions from generated images. Recently, [9] demonstrated that the cross-attention maps capture the human or object's existence corresponding to each token in prompts. We leverage the cross-attention maps \\(\\mathcal{A}_h\\), \\(\\mathcal{A}_r\\) and \\(\\mathcal{A}_o\\) corresponding to \\(<\\mathrm{h}, \\mathrm{r}, \\mathrm{o}>\\) token to extract interaction region between human and objects."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.321,
+                0.906,
+                0.41
+            ],
+            "angle": 0,
+            "content": "However, it is nontrivial to directly extract the specific interaction region from a continuous attention map. We apply the centroid extraction mechanism proposed in [5] to compute specific points within attention maps. Specifically, we calculate \\( c_{h} \\), \\( c_{r} \\) and \\( c_{o} \\) from each cross-attention maps as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.609,
+                0.423,
+                0.906,
+                0.46
+            ],
+            "angle": 0,
+            "content": "\\[\nc = \\frac {1}{\\sum_ {h , w} \\mathcal {A}} \\left[ \\begin{array}{l} \\sum_ {h, w} w \\cdot \\mathcal {A} \\\\ \\sum_ {h, w} h \\cdot \\mathcal {A} \\end{array} \\right], \\tag {8}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.472,
+                0.906,
+                0.61
+            ],
+            "angle": 0,
+            "content": "where \\( h \\) and \\( w \\) denote the height and width in the attention map \\( \\mathcal{A} \\). Then, we define the interaction center \\( c_{rel} \\) as the centroid of a triangle defined by \\( c_h \\), \\( c_r \\), and \\( c_o \\). Finally, we extract interaction region \\( B_{rel}^{gen} \\) by leveraging the distance between human and object centers with \\( B_{rel}^{gen} = c_{rel} \\pm \\|c_h - c_o\\|_2^2 \\). Additionally, we extract \\( B_{rel}^{gt} \\) in the same manner, leveraging the human and object bounding boxes provided in the dataset. We treat \\( c_{rel} \\) as the midpoint between the human and object centers from the given boxes \\( B^{gt} \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.611,
+                0.907,
+                0.854
+            ],
+            "angle": 0,
+            "content": "Interaction Direction Guidance. To generate images with more realistic interactions, we design guidance that aligns feature differences at the image level with those at the interaction region. With interaction region \\( B_{rel}^{gt} \\) and \\( B_{rel}^{gen} \\), we obtain interaction region image by masking the real and generated images: \\( I_{rel}^{g} = B_{rel}^{g} \\odot I^{g} \\), where \\( g \\in \\{gt, gen\\} \\). We then encode both interaction region images through a CLIP image encoder and extract interaction region features \\( f_{rel}^{gt} \\) and \\( f_{rel}^{gen} \\). Leveraging interaction region features, we design the interaction direction guidance to focus the model on modifying the image feature, especially within the interaction region. Specifically, we calculate the direction between \\( f_{rel}^{gt} \\) and \\( f_{rel}^{gen} \\), referred to as biased relation feature \\( f_{rel}^{bias} = f_{rel}^{gt} - f_{rel}^{gen} \\), to align the direction of biased interaction features with that of the real images. The interaction direction guidance is as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.588,
+                0.868,
+                0.906,
+                0.907
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\mathrm {I D G}} = 1 - \\frac {\\left(f _ {\\mathcal {M}} ^ {g t} - f ^ {g e n}\\right) \\cdot \\left(f _ {r e l} ^ {b i a s}\\right)}{\\left| f _ {\\mathcal {M}} ^ {g t} - f ^ {g e n} \\right| \\left| f _ {r e l} ^ {b i a s} \\right|}. \\tag {9}\n\\]"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "8045"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.091,
+                0.245,
+                0.108
+            ],
+            "angle": 0,
+            "content": "3.4. Training Phase"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.114,
+                0.484,
+                0.28
+            ],
+            "angle": 0,
+            "content": "We train only the cross-attention layer in SD to capture the semantically distinct relation words, as the cross-attention reflects the existence of each word token [18]. We adopt the reconstruction loss (Eq. 1) to train T2I models in addition to relation disentanglement guidance and interaction direction guidance. However, when an image contains multiple humans and objects, it can lead to multiple interactions that do not match a single target interaction verb. To accurately separate the human and object corresponding to interaction words, we apply mask \\(\\mathcal{M}\\) when calculating the reconstruction loss as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.097,
+                0.286,
+                0.484,
+                0.327
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\text {r e c}} = \\mathbb {E} _ {z, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| (\\epsilon \\odot \\mathcal {M} - \\epsilon_ {\\theta} (z _ {t}, t, T) \\odot \\mathcal {M}) \\| _ {2} ^ {2} \\right]. \\tag {10}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.336,
+                0.484,
+                0.381
+            ],
+            "angle": 0,
+            "content": "Finally, we combine reconstruction loss with disentanglement guidance and interaction direction guidance leveraging \\(\\lambda_1, \\lambda_2\\), and \\(\\lambda_3\\) as below:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.132,
+                0.392,
+                0.483,
+                0.409
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\cdot \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {2} \\cdot \\mathcal {L} _ {\\text {R D G}} + \\lambda_ {3} \\cdot \\mathcal {L} _ {\\text {I D G}}. \\tag {11}\n\\]"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.419,
+                0.224,
+                0.436
+            ],
+            "angle": 0,
+            "content": "4. Experiments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.444,
+                0.484,
+                0.566
+            ],
+            "angle": 0,
+            "content": "We train our model on SD v1.4 at a resolution of \\(512 \\times 512\\). Using the Adam optimizer with a learning rate of \\(4 \\times 10^{-6}\\), we trained for a single epoch over 17 hours with a batch size of 12 on 3 Nvidia A6000 GPUs. For image generation during training, we utilize the DDIM scheduler [27] with 30 sampling steps, and during inference, we increase sampling steps to 50. Finally, \\(\\lambda_{1}, \\lambda_{2}\\), and \\(\\lambda_{3}\\) are set to 1.0, 10, and 0.8, respectively, and the margin \\(m\\) is set to 0.2."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.574,
+                0.187,
+                0.589
+            ],
+            "angle": 0,
+            "content": "4.1. Dataset"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.597,
+                0.483,
+                0.658
+            ],
+            "angle": 0,
+            "content": "We conduct our experiments on the HICO-DET dataset [2], which consists of 47,776 images (38,118 for training and 9,658 for testing). The dataset provides annotations for 600 HOI categories using 80 object categories and 117 verbs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.658,
+                0.484,
+                0.794
+            ],
+            "angle": 0,
+            "content": "To train our model, we realign the training set according to the 600 HOI text descriptions and modify annotations to match each image with its corresponding text descriptions. This realignment yields 69,404 images across 600 prompts. We exclude text classes containing \"and\" (e.g. \"A photo of a person and an airplane\") to avoid ambiguity from various interactions within the same class, resulting in a final training set of 61,114 images across 501 prompts. More details about the dataset are in the supplementary materials."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.802,
+                0.187,
+                0.817
+            ],
+            "angle": 0,
+            "content": "4.2. Metrics"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.825,
+                0.484,
+                0.902
+            ],
+            "angle": 0,
+            "content": "CLIP & Sentence-BERT Similarity. We assess the model understandability of interaction words by calculating text-to-text similarity scores using both CLIP [23] and SentenceBERT (S-BERT) [24] that we newly adopt to image generation task. After generating captions with InstructBLIP [19],"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.092,
+                0.905,
+                0.152
+            ],
+            "angle": 0,
+            "content": "we compute cosine similarity against ground-truth texts. We also measure the similarity between the input prompts and generated images through CLIP. Details about the S-BERT metric are in the supplementary materials."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.153,
+                0.906,
+                0.334
+            ],
+            "angle": 0,
+            "content": "HOI Classification Accuracy. To evaluate the robustness of the model understanding interaction words, we measure interaction accuracy with generated images using the SOV-STG [3] HOI detection model pre-trained on HICODET [2]. We report scores on two different weights, SOV-STG-S and SOV-STG-Swin-L, with different backbones (e.g. ResNet [8] and Swin-L [21]). Accuracy is reported in two main settings: Known Object (KO), which measures verb accuracy assuming the objects are correct, and Default (Def.), which assesses both verb and object correctness, with each setting further divided into Full (all classes) and Rare (classes with fewer than 10 samples)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.335,
+                0.907,
+                0.547
+            ],
+            "angle": 0,
+            "content": "VQA-Score. We additionally use VQAScore [17], a large image-text alignment benchmark for generative models, to evaluate comprehensive interaction alignment scores. This score measures the model's ability to understand more complex compositional prompts, such as spatial and action relationships. For assessing the accuracy of depicting interaction between humans and objects in images, we use two main metrics: the Image-to-Text (I2T) alignment score and the VQA score. The I2T alignment score improves similarity measurement between complex prompts containing interaction words and generated images by incorporating a large language model. The VQA score measures the probability that a VQA model will answer \"yes\" to questions containing interaction words paired with generated images."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.558,
+                0.697,
+                0.574
+            ],
+            "angle": 0,
+            "content": "4.3. Qualitative Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.582,
+                0.906,
+                0.688
+            ],
+            "angle": 0,
+            "content": "Fig. 4 shows a comparison of generated images focusing on the interactions between humans and objects across our method and other models. Compared to previous approaches, our results exhibit more detailed interactions, closely resembling ground-truth images and producing high-quality outputs comparable to images from DALL-E 3 [1], a large-scale text-to-image generation model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.689,
+                0.906,
+                0.795
+            ],
+            "angle": 0,
+            "content": "For example, in Fig. 4 (a)-(f), GLIGEN [15] mainly focuses on generating humans and objects, failing to depict the accurate interactions related to the interaction words. While InteractDiffusion [10] improves interaction representation, its results show inaccurate or ambiguous interactions that lack both accurate object rendering (e) or fine-grained interaction details (a)-(d) and (f)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.796,
+                0.906,
+                0.901
+            ],
+            "angle": 0,
+            "content": "In particular, in Fig.4 (g), our model generates a visually enhanced interaction that closely resembles the ground truth interaction region, showing fluid coming out from a bottle. In the multi-humans with one interaction scenario (f), our results depict people holding foods similar to the image from Dall-E 3, while other models only show people sitting at the table. This demonstrates that our model fur"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "8046"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.137,
+                0.09,
+                0.866,
+                0.593
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.09,
+                0.602,
+                0.907,
+                0.675
+            ],
+            "angle": 0,
+            "content": "Figure 4. Interaction comparison in generated images with other models. We generate images using a fixed template: “A {H} {R} a/an {O}”. The top row displays the input human, interaction word, and object. Green and blue boxes represent additional grounding boxes used during image generation for each human and object, respectively. Our results produce images with more accurate interactions than other models, closely resembling ground-truth interactions while maintaining high image quality comparable to large-scale generative models like DALL-E 3 [1]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.699,
+                0.483,
+                0.804
+            ],
+            "angle": 0,
+            "content": "ther concentrates on detailed interaction regions and distinguishes subtle differences between interaction words with similar spatial arrangements (e.g. \"sitting at\" and \"eating at\"). Overall, our model demonstrates a similar level of interaction understandability as a large-scale model such as Dall-E 3 and more accurately exhibits the intended interactions, closely matching with the ground-truth interactions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.811,
+                0.483,
+                0.902
+            ],
+            "angle": 0,
+            "content": "We further evaluate our model with complex prompts with multiple interactions to test the ability to distinguish interaction words. As shown in Fig. 1, when given multiple interactions, our model accurately depicts each one, while other models struggle. For example, GLIGEN places objects correctly but fails to represent interactions like \"walk"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.699,
+                0.907,
+                0.807
+            ],
+            "angle": 0,
+            "content": "ing,\" \"jumping,\" and \"throwing.\" InteractDiffusion captures some interactions but misses key objects, like an umbrella. In contrast, our model generates accurate images that reflect the intended interactions. This shows that our model successfully captures the semantic difference between interaction words, generating more detailed interactions. More qualitative results are in the supplementary materials."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.817,
+                0.708,
+                0.832
+            ],
+            "angle": 0,
+            "content": "4.4.Quantitative Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.84,
+                0.907,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Tab. 1 compares similarity scores between VerbDiff and previous methods. Our model achieves the highest scores across nearly all evaluation settings. In the text-to-text (T2T) similarity evaluation, VerbDiff outperforms existing"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "8047"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.095,
+                0.089,
+                0.483,
+                0.197
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Models</td><td colspan=\"2\">CLIP</td><td>S-BERT</td><td colspan=\"2\">VQA-Score</td></tr><tr><td>T2T</td><td>T2I</td><td>T2T</td><td>I2T</td><td>VQA</td></tr><tr><td>HICO-DET</td><td>0.604</td><td>0.233</td><td>0.446</td><td>0.755</td><td>0.754</td></tr><tr><td>SD [25]</td><td>0.725</td><td>0.251</td><td>0.620</td><td>0.769</td><td>0.765</td></tr><tr><td>GLIGEN [15]</td><td>0.683</td><td>0.238</td><td>0.554</td><td>0.679</td><td>0.674</td></tr><tr><td>InteractDiffusion [10]</td><td>0.703</td><td>0.232</td><td>0.575</td><td>0.728</td><td>0.734</td></tr><tr><td>VerbDiff (Ours)</td><td>0.733</td><td>0.242</td><td>0.633</td><td>0.771</td><td>0.766</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.09,
+                0.206,
+                0.483,
+                0.25
+            ],
+            "angle": 0,
+            "content": "Table 1. Similarity comparison between VerbDiff and other models. We evaluate scores on CLIP, S-BERT [24] and a large vision-language alignment benchmark VQA-Score [17]."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.094,
+                0.264,
+                0.485,
+                0.372
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"3\">Models</td><td colspan=\"4\">SOV-STG-S (Acc ↑)</td><td colspan=\"4\">SOV-STG-Swin-L (Acc ↑)</td></tr><tr><td colspan=\"2\">Def.</td><td colspan=\"2\">KO.</td><td colspan=\"2\">Def.</td><td colspan=\"2\">KO.</td></tr><tr><td>Full</td><td>Rare</td><td>Full</td><td>Rare</td><td>Full</td><td>Rare</td><td>Full</td><td>Rare</td></tr><tr><td>HICO-DET</td><td>26.52</td><td>6.78</td><td>28.68</td><td>7.29</td><td>29.98</td><td>12.66</td><td>31.16</td><td>13.43</td></tr><tr><td>SD [25]</td><td>16.09</td><td>4.59</td><td>18.22</td><td>4.85</td><td>20.08</td><td>8.07</td><td>21.69</td><td>8.66</td></tr><tr><td>GLIGEN [15]</td><td>15.88</td><td>4.85</td><td>17.91</td><td>5.24</td><td>17.83</td><td>7.00</td><td>19.35</td><td>7.57</td></tr><tr><td>InteractDiffusion [10]</td><td>19.67</td><td>7.00</td><td>21.31</td><td>7.69</td><td>23.53</td><td>10.27</td><td>24.86</td><td>11.18</td></tr><tr><td>VerbDiff (Ours)</td><td>22.59</td><td>7.62</td><td>24.79</td><td>7.83</td><td>27.05</td><td>12.60</td><td>28.43</td><td>13.18</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.09,
+                0.382,
+                0.483,
+                0.41
+            ],
+            "angle": 0,
+            "content": "Table 2. HOI accuracy comparison between VerbDiff and previous methods. Def. and KO. refer to Default and Known Object."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.439,
+                0.483,
+                0.649
+            ],
+            "angle": 0,
+            "content": "methods, and the difference becomes even more significant in the S-BERT evaluation. This highlights that the S-BERT metric better captures the semantic differences within interaction verbs. Although the text-to-image (T2I) score is lower than SD, our model achieves higher scores than the models that leverage additional bounding boxes, demonstrating the ability of our model to understand interactions. To assess a more comprehensive interaction understanding of VerbDiff, we also report the VQA-Score, which evaluates how well generative models understand more complex compositional prompts, such as interactions. Although our method shows a slight improvement in the I2T and VQA scores compared to SD, it scores significantly higher than InteractDiffusion."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.652,
+                0.483,
+                0.834
+            ],
+            "angle": 0,
+            "content": "Tab. 2 presents the HOI accuracy scores, where VerbDiff achieves the highest accuracy across all settings compared to previous methods. In particular, although different backbones are used for computing accuracy, VerbDiff consistently shows the highest accuracy across all settings. This demonstrates that our method has robust interaction word understanding and produces images with precise interactions. Overall, VerbDiff consistently outperforms other models across nearly all evaluation metrics, demonstrating its strong ability to comprehend interaction words and generate high-quality images with accurate human-object interactions, even without extra conditions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.846,
+                0.245,
+                0.863
+            ],
+            "angle": 0,
+            "content": "4.5. Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.871,
+                0.483,
+                0.903
+            ],
+            "angle": 0,
+            "content": "We show the effectiveness of RDG and IDG in Tab. 3 and Fig. 5. As shown, both RDG (w/ \\(\\mathcal{L}_{\\text {triple }}\\) and \\(\\mathcal{L}_{\\text {align }}\\)) (b)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.517,
+                0.091,
+                0.906,
+                0.179
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.513,
+                0.19,
+                0.907,
+                0.232
+            ],
+            "angle": 0,
+            "content": "Figure 5. Image interaction comparison with and without guidance. IDG focuses on the specific localized interaction region (red box) in the image."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.521,
+                0.247,
+                0.905,
+                0.33
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>VerbDiff</td><td>Lrec</td><td>Ltriple</td><td>Lalign</td><td>LIDG</td><td>CLTP2T</td><td>S-BERT T2T</td><td>HOI Def.</td><td>Acc KO.</td></tr><tr><td>(a)</td><td>✓</td><td></td><td></td><td></td><td>0.691</td><td>0.582</td><td>19.38</td><td>20.89</td></tr><tr><td>(b)</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>0.700</td><td>0.589</td><td>20.32</td><td>21.87</td></tr><tr><td>(c)</td><td>✓</td><td></td><td></td><td>✓</td><td>0.699</td><td>0.588</td><td>20.21</td><td>21.67</td></tr><tr><td>(d)</td><td>✓</td><td>✓</td><td></td><td>✓</td><td>0.710</td><td>0.610</td><td>23.39</td><td>24.51</td></tr><tr><td>All (Ours)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>0.733</td><td>0.633</td><td>27.05</td><td>28.43</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.513,
+                0.34,
+                0.907,
+                0.396
+            ],
+            "angle": 0,
+            "content": "Table 3. Ablation of the VerbDiff guidance settings. We score the similarity and HOI accuracy on the Full setting. We use the SOV-STG-Swin-L model. Combining all the proposed loss functions shows the best performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.424,
+                0.907,
+                0.621
+            ],
+            "angle": 0,
+            "content": "and IDG (c) enhance the similarity score and HOI accuracy. Specifically, (b) scores slightly higher than (c) demonstrating that the model RDG distinguishes the semantics in interaction words. While scores improve significantly when using IDG and RDG without \\(\\mathcal{L}_{\\text {align }}\\), the model with all guidance terms achieves the highest overall performance. In Fig. 5, we compare images generated with and without each type of guidance. Notably, the enlarged interaction region (red box) shows that the model with IDG produces detailed representations (e.g. human hands), unlike models with only RDG. This indicates that IDG effectively captures fine-grained details in localized interaction regions. More ablations are in the supplementary materials."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.635,
+                0.634,
+                0.651
+            ],
+            "angle": 0,
+            "content": "5. Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.661,
+                0.907,
+                0.872
+            ],
+            "angle": 0,
+            "content": "We propose VerbDiff, a novel text-to-image (T2I) diffusion models that address the interaction misunderstanding problem in SD without additional conditions. Although previous methods leverage additional conditions to help the model understand the interaction between humans and objects, they still rely on precise instructions and lack an understanding of the semantic differences between interaction verbs. Our model successfully captures the semantic meanings inherent in interaction words and generates high-quality images with accurate interactions. Extensive experiments demonstrate the effectiveness of our method in enhancing the ability to understand interaction words of T2I models, achieving better interaction comprehension compared to the previous approaches."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.873,
+                0.906,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Acknowledgments. This was supported by the Institute of Information & Communications Technology Plan-"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "8048"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.092,
+                0.486,
+                0.136
+            ],
+            "angle": 0,
+            "content": "ning & Evaluation (IITP) grant funded by the Korean government(MSIT) (No.RS-2020-II201373, Artificial Intelligence Graduate School Program(Hanyang University))."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.093,
+                0.163,
+                0.188,
+                0.179
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.101,
+                0.188,
+                0.482,
+                0.258
+            ],
+            "angle": 0,
+            "content": "[1] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.26,
+                0.483,
+                0.316
+            ],
+            "angle": 0,
+            "content": "[2] Yu-Wei Chao, Yunfan Liu, Xieyang Liu, Huayi Zeng, and Jia Deng. Learning to detect human-object interactions. In 2018 IEEE winter conference on applications of computer vision (wacv), pages 381-389. IEEE, 2018. 2, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.318,
+                0.483,
+                0.373
+            ],
+            "angle": 0,
+            "content": "[3] Junwen Chen, Yingcheng Wang, and Keiji Yanai. Focusing on what to decode and what to train: Efficient training with hoi split decoders and specific target guided denoising. arXiv preprint arXiv:2307.02291, 2023. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.375,
+                0.483,
+                0.442
+            ],
+            "angle": 0,
+            "content": "[4] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.446,
+                0.483,
+                0.5
+            ],
+            "angle": 0,
+            "content": "[5] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.503,
+                0.483,
+                0.572
+            ],
+            "angle": 0,
+            "content": "[6] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.574,
+                0.483,
+                0.628
+            ],
+            "angle": 0,
+            "content": "[7] Hanan Gani, Shariq Farooq Bhat, Muzammal Naseer, Salman Khan, and Peter Wonka. Lm blueprint: Enabling text-to-image generation with complex and detailed prompts. arXiv preprint arXiv:2310.10640, 2023. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.631,
+                0.483,
+                0.686
+            ],
+            "angle": 0,
+            "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.688,
+                0.483,
+                0.742
+            ],
+            "angle": 0,
+            "content": "[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.746,
+                0.483,
+                0.814
+            ],
+            "angle": 0,
+            "content": "[10] Jiu Tian Hoe, Xudong Jiang, Chee Seng Chan, Yap-Peng Tan, and Weipeng Hu. Interactdiffusion: Interaction control in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6180-6189, 2024. 1, 2, 3, 6, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.817,
+                0.483,
+                0.884
+            ],
+            "angle": 0,
+            "content": "[11] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In Similarity-based pattern recognition: third international workshop, SIMBAD 2015, Copenhagen, Denmark, October 12-14, 2015. Proceedings 3, pages 84-92. Springer, 2015. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.887,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[12] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan,"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.094,
+                0.188,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.545,
+                0.093,
+                0.905,
+                0.121
+            ],
+            "angle": 0,
+            "content": "and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.123,
+                0.905,
+                0.177
+            ],
+            "angle": 0,
+            "content": "[13] Xu Jia, Takashi Isobe, Xiaomin Li, Qinghe Wang, Jing Mu, Dong Zhou, Huchuan Lu, Lu Tian, Ashish Sirasao, Emad Barsoum, et al. Customizing text-to-image generation with inverted interaction. In ACM Multimedia 2024, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.179,
+                0.905,
+                0.247
+            ],
+            "angle": 0,
+            "content": "[14] Jian-Yu Jiang-Lin, Kang-Yang Huang, Ling Lo, Yi-Ning Huang, Terence Lin, Jhih-Ciang Wu, Hong-Han Shuai, and Wen-Huang Cheng. Record: Reasoning and correcting diffusion for hoi generation. arXiv preprint arXiv:2407.17911, 2024.3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.25,
+                0.905,
+                0.331
+            ],
+            "angle": 0,
+            "content": "[15] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 1, 2, 3, 6, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.334,
+                0.905,
+                0.389
+            ],
+            "angle": 0,
+            "content": "[16] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.392,
+                0.905,
+                0.46
+            ],
+            "angle": 0,
+            "content": "[17] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2025. 6, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.463,
+                0.905,
+                0.532
+            ],
+            "angle": 0,
+            "content": "[18] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7817-7826, 2024. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.534,
+                0.905,
+                0.574
+            ],
+            "angle": 0,
+            "content": "[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.577,
+                0.905,
+                0.631
+            ],
+            "angle": 0,
+            "content": "[20] Jinxiu Liu and Qi Liu. R3cd: Scene graph to image generation with relation-aware compositional contrastive control diffusion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3657-3665, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.633,
+                0.905,
+                0.702
+            ],
+            "angle": 0,
+            "content": "[21] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.704,
+                0.905,
+                0.773
+            ],
+            "angle": 0,
+            "content": "[22] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15579-15591, 2023. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.776,
+                0.905,
+                0.857
+            ],
+            "angle": 0,
+            "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.86,
+                0.905,
+                0.9
+            ],
+            "angle": 0,
+            "content": "[24] N Reimers. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 2, 6, 8"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.517,
+                0.093,
+                0.905,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "8049"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.093,
+                0.486,
+                0.161
+            ],
+            "angle": 0,
+            "content": "[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.164,
+                0.483,
+                0.233
+            ],
+            "angle": 0,
+            "content": "[26] Guibao Shen, Luozhou Wang, Jiantao Lin, Wenhang Ge, Chaozhe Zhang, Xin Tao, Yuan Zhang, Pengfei Wan, Zhongyuan Wang, Guangyong Chen, et al. Sg-adapter: Enhancing text-to-image generation with scene graph guidance. arXiv preprint arXiv:2405.15321, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.234,
+                0.483,
+                0.274
+            ],
+            "angle": 0,
+            "content": "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.277,
+                0.483,
+                0.332
+            ],
+            "angle": 0,
+            "content": "[28] Shengqiong Wu, Hao Fei, Hanwang Zhang, and Tat-Seng Chua. Imagine that! abstract-to-intricate text-to-image synthesis with scene graph hallucination diffusion. Advances in Neural Information Processing Systems, 36, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.334,
+                0.483,
+                0.389
+            ],
+            "angle": 0,
+            "content": "[29] Yinwei Wu, Xingyi Yang, and Xinchao Wang. Relation rectification in diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7685-7694, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.391,
+                0.483,
+                0.471
+            ],
+            "angle": 0,
+            "content": "[30] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.474,
+                0.483,
+                0.528
+            ],
+            "angle": 0,
+            "content": "[31] Bicheng Xu, Qi Yan, Renjie Liao, Lele Wang, and Leonid Sigal. Joint generative modeling of scene graphs and images via diffusion models. arXiv preprint arXiv:2401.01130, 2024. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.531,
+                0.483,
+                0.599
+            ],
+            "angle": 0,
+            "content": "[32] Ling Yang, Zhilin Huang, Yang Song, Shenda Hong, Guohao Li, Wentao Zhang, Bin Cui, Bernard Ghanem, and Ming-Hsuan Yang. Diffusion-based scene graph to image generation with masked contrastive pre-training. arXiv preprint arXiv:2211.11138, 2022. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.602,
+                0.483,
+                0.657
+            ],
+            "angle": 0,
+            "content": "[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.659,
+                0.483,
+                0.727
+            ],
+            "angle": 0,
+            "content": "[34] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2, 3"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.093,
+                0.093,
+                0.486,
+                0.727
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "8050"
+        }
+    ]
+]

2025/VerbDiff_ Text-Only Diffusion Models with Enhanced Interaction Awareness/ad57dfe8-0121-4da2-a044-841b75f6e661_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2577f847253bfcea2c92de91ab6672e664bf50bd2c9d11097b230148e8611484
+size 2196710

2025/VerbDiff_ Text-Only Diffusion Models with Enhanced Interaction Awareness/full.md

@@ -0,0 +1,275 @@
+# VerbDiff: Text-Only Diffusion Models with Enhanced Interaction Awareness
+
+SeungJu Cha Kwanyoung Lee Ye-Chan Kim Hyunwoo Oh Dong-Jin Kim Hanyang University, South Korea
+
+{sju9020, mobled37, dpcksd178, komji, djdkim}@hanyang.ac.kr
+
+![](images/6762b1d33f38b4c1d57bbdf425112cd776d156f09d0a48b026c728e205ac7464.jpg)  
+"A basketball player jumping and throwing a basketball and a man blocking the ball"  
+Figure 1. Generated samples illustrating multiple human-object interactions. Each color represents distinct humans, objects, and interaction words. GLIGEN [15] and InteractDiffusion [10] use grounding boxes as additional conditions, whereas Stable Diffusion [25] and VerbDiff rely solely on text.
+
+# Abstract
+
+Recent large-scale text-to-image diffusion models generate photorealistic images but often struggle to accurately depict interactions between humans and objects due to their limited ability to differentiate various interaction words. In this work, we propose VerbDiff to address the challenge of capturing nuanced interactions within text-to-image diffusion models. VerbDiff is a novel text-to-image generation model that weakens the bias between interaction words and objects, enhancing the understanding of interactions. Specifically, we disentangle various interaction words from frequency-based anchor words and leverage localized interaction regions from generated images to help the model better capture semantics in distinctive words without extra
+
+conditions. Our approach enables the model to accurately understand the intended interaction between humans and objects, producing high-quality images with accurate interactions aligned with specified verbs. Extensive experiments on the HICO-DET dataset demonstrate the effectiveness of our method compared to previous approaches.
+
+# 1. Introduction
+
+Recently, text-to-image diffusion models have demonstrated the ability to produce photorealistic images from natural texts [25]. However, these models sometimes struggle to accurately depict interactions between humans and objects [10]. For example, they often fail to differentiate
+
+![](images/0266c3ca1bcd8f7ab298f6f0e703e45274128c8435b0270bd8e5974da1596d53.jpg)  
+Figure 2. Examples of interactions from InteractDiffusion [10], showing a lack of understanding of interaction words. The model relies on precise bounding boxes rather than understanding interaction words to exhibit accurate interactions, as shown in the results for the "red" and "blue" boxes.
+
+between semantically distinct prompts, such as "A person walking a bicycle" and "A person riding a bicycle," particularly in capturing the semantics in verbs that are crucial for accurately depicting the intended interaction.
+
+We hypothesize that this problem arises from the strong object bias of CLIP [23], which fails to adequately capture semantic differences between verbs [22]. Specifically, CLIP [23] often focuses on objects in prompts, overlooking semantic distinctions in interaction verbs. To enhance prompt interpretation ability of models, previous methods have leveraged additional modules such as large language models (LLMs) [6, 7, 16] or extra bounding boxes [10, 15, 30, 34] to provide explicit relational visual information.
+
+For example, InteractDiffusion [10] introduces controllable image generation using Human-Object Interaction (HOI) information with bounding boxes for human, relation, and object  $<\mathrm{h}, \mathrm{r}, \mathrm{o}>$  triplets. Nevertheless, as shown in Fig. 2, the model still struggles with accurate interactions even with these additional conditions. For instance, given the prompt "walking, bicycle," the model fails to depict the intended interaction when semantically distinctive interaction words share a similar bounding box (e.g. walking and riding). This suggests that the model lacks an understanding of interaction semantics and relies heavily on precise bounding boxes, which are labor-intensive to provide.
+
+To improve the model's interaction understandability without additional conditions, we propose VerbDiff, a novel text-to-image generation method that better distinguishes interaction words via several objectives. First, we apply Relation Disentanglement Guidance (RDG) using frequency-based anchor texts from the dataset for each human-object pair to reduce interaction bias in generated images. We
+
+observe that the exhibited interactions in generated images have bias, which follows frequent verbs corresponding to each human-object pair in the data distribution. To address this, we align interaction features between real and generated images while separating those of generated images from anchor text features, enabling a more effective distinction of interaction words. Additionally, we introduce a balancing factor to address the extreme long-tail distribution of interaction words across human-object pairs, adjusting interaction modification during training. This allows the model to learn diverse interactions in a balanced way.
+
+Secondly, to enforce the model to better focus on the interaction regions between humans and objects, we introduce Interaction Direction Guidance (IDG). This approach emphasizes fine-grained semantic distinctions in localized interactions. We design an Interaction Region (IR) module to capture specific interaction areas in generated images, leveraging cross-attention maps in text-only diffusion models. The IR module extracts interaction regions using the centroids of cross-attention maps associated with  $<\mathrm{h}$ , r, o> tokens. We obtain biased interaction features from these regions and apply interaction direction guidance. This guidance effectively modifies interaction regions within images, guiding Stable Diffusion (SD) to concentrate more accurately on these regions while capturing nuanced interaction details. We validate our method's effectiveness on the HICO-DET dataset [2]. Our approach accurately captures semantic differences between interaction words and generates high-quality images with accurate interactions compared to SD [25], InteractDiffusion [10], and other prior methods. Additionally, we introduce a new metric using Sentence BERT [24] for image generation tasks. This metric offers improved capability in discerning subtle and nuanced meanings of interaction words compared to CLIP. By focusing on semantic distinctions in interactions, this metric enables more precise evaluation, complementing CLIP similarity. Our contributions are as follows:
+
+(1) We introduce VerbDiff, a novel text-to-image model that enhances the semantic interaction understanding by distinguishing interaction verbs and generating images with accurate human-object interactions without any additional conditions. (2) To reduce generating images with biased interactions, we apply relation disentanglement guidance using frequency-based anchor text extracted from real data distributions. (3) Our proposed IR module captures detailed interaction regions using cross-attention map centroids, providing specified-interaction region guidance without explicit bounding boxes. (4) We implement a Sentence-BERT metric to effectively capture semantic differences between interaction sentences, addressing CLIP's limitations. Extensive experiments demonstrate that our approach outperforms previous methods across various settings.
+
+# 2. Related Work
+
+# 2.1. Diffusion Models
+
+Recently, diffusion models have achieved state-of-the-art performance in text-to-image generation [25] and controllable image generation [10, 15, 33]. While Stable Diffusion (SD) [25] integrates conditioning embeddings into cross-attention layers for text-to-image generation, it often struggles to fully align generated images with text instructions. This arises from inherent limitations of texts to convey exact visual information and strong object bias of CLIP [22].
+
+To address these, recent research has focused on controllable image generation using bounding boxes [15, 30, 34]. For example, GLIGEN [15] extends SD to incorporate grounding inputs such as bounding boxes, image prompts, key points, and spatially aligned conditions. LayoutDiffusion [34] enables multiple bounding box inputs for layout-specific generation. Although these methods better generate images, they require additional conditions beyond text input to achieve the intended image outputs. In contrast, our proposed method achieves comparable results without additional grounding information, providing a more efficient alternative for precise and flexible generation.
+
+# 2.2. Relation Generation
+
+Relation generation between objects has been explored leveraging LLMs [6, 7, 16] and scene graph [20, 26, 28, 29, 31, 32]. While LLM-based methods can generate accurate object layouts, they primarily focus on compositional relations rather than interactions [6, 7, 16]. Scene graph-based approaches have addressed these limitations by combining scene graph encoders to enhance prompt understandability [20, 26, 28, 29, 31, 32]. However, they require additional training to extract scene graphs from prompts and struggle to capture subtle semantic differences in interactions.
+
+To depict accurate interactions during generation, researchers have focused on human-object interactions (HOI) [10, 12-14]. Some methods use inversion-based frameworks to capture interaction semantics but require optimization for each interaction word [12, 13]. More generalized approaches leverage additional conditions such as human poses [14] or bounding boxes [10]. Specifically, InteractDiffusion [10] leverages the HOI information with bounding boxes to generate interactions. However, this method still relies on additional information rather than semantic differences between prompts. Our work is similar to InteractDiffusion [10] by leveraging HOI information but differs from focusing on semantic differences between interaction words to enhance interaction understanding in SD. This approach aims to improve the generation of nuanced interactions without relying on additional conditions. It could potentially lead to more accurate and diverse depictions of human-object interactions in synthesized images.
+
+# 3. Method
+
+Our goal is to generate images with accurate interactions by enhancing the model understandability of the interaction word without additional conditions. Our proposed method, VerbDiff as illustrated in Fig. 3 contains two main parts: (1) Relation disentanglement guidance leverages frequency-based anchor texts from human and object pairs along with input texts and real and generated image features. This guidance helps visually align the generated interaction features with real ones and distinguish the semantic differences between interaction words. (2) Interaction direction guidance that helps the model concentrate more on the interaction region while modifying the image features. To capture the interaction region in arbitrarily generated images, we further design an interaction region module (IR module) to extract a centroid-based interaction region from the cross-attention map corresponding to  $< \mathrm{h},\mathrm{r},\mathrm{o} >$  tokens.
+
+# 3.1. Preliminary
+
+SD Training Objective. Text-to-image generation models, such as Stable Diffusion (SD) [25], show photo-realistic images that align with given text prompts through an efficient two-stage training scheme. The first stage learns to map the input images  $I$  into latent space  $z$ . The second stage learns the denoising UNet to gradually predict the real noise in the diffusion forward process. Starting from the original latent image  $z_0$ , the diffusion model gradually adds noise  $\epsilon$  to produce noisy sample  $z_t$  through uniformly sampled time steps  $t$ . Then, with text condition  $T$ , the UNet predicts the noise  $\epsilon_{\theta}$  for each time step to reconstruct  $z_t$  to  $z_0$  as follows:
+
+$$
+\min  _ {\theta} \mathcal {L} _ {\mathrm {L D M}} = \mathbb {E} _ {z, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| (\epsilon - \epsilon_ {\theta} (z _ {t}, t, T)) \| _ {2} ^ {2} \right]. \tag {1}
+$$
+
+Cross-Attention Layer. When given text, the CLIP [23] text encoder produces text embedding, and the UNet in SD takes embedding and computes cross-attention between encoded text  $e$  and intermediate image features  $\phi(z_{t})$ :
+
+$$
+\mathcal {A} = \operatorname {S o f t m a x} \left(\frac {Q K ^ {\top}}{\sqrt {d}}\right), \tag {2}
+$$
+
+where query  $Q = W_{Q} \cdot \phi(z_{t})$  and key  $K = W_{K} \cdot e$  are computed with projection matrix  $W_{Q}$ ,  $W_{K}$  and  $d$  is the dimension of  $Q$ . These cross-attention maps  $\mathcal{A}$  enable the text-to-image diffusion models to depict texts into image features. In addition, each  $\mathcal{A}_{j}$ , a cross-attention map corresponding to  $j^{th}$  token in prompts, represents the attention weight on intermediate image features in various resolutions.
+
+# 3.2. Relation Disentanglement Guidance
+
+To depict accurate interaction in generated images without additional conditions, we first focus on the semantic difference between interaction words, especially the verbs. As the
+
+![](images/1f78b73afcb0a51fc75ba1ffc3ba614dd9f9b589ba3cbf7de08546515d27a8ca.jpg)  
+Figure 3. Pipeline of VerbDiff. VerbDiff uses Relation Disentanglment Guidance (Sec. 3.2) that separates the interaction features from the anchor text for each human-object pair. Additionally, it contains the IR module (Sec. 3.3) (right) which extracts localized interaction regions from generated images without explicit bounding boxes and Interaction Direction Guidance (Sec. 3.3) that guides the model to focus more on the fine-grained interaction regions.
+
+![](images/402f7dc2c45e45c40a2f9962c5008383d634ad746dbbb00349cad9eb8b99b42c.jpg)
+
+SD is known to have limitations in capturing the semantic nuances in interaction words, we apply relation disentangle-ment guidances. The goal of this guidance is to alleviate the relation bias within human-object pairs and improve interaction comprehension through anchor text, input text, and real and generated images.
+
+First, given the ground-truth text  $T^{gt} =$  "A photo of a  $\{h\} \{r^{gt}\} a / an\{o\}$ " (e.g. "A photo of a person holding a backpack"), SD generates images  $I^{gen}$  during training. Although  $T^{gt}$  specifies the target interaction,  $I^{gen}$  often contains inaccurate and biased interactions that fail to meet verbs between each human and object pair. We define a sentence with accurate interaction verbs corresponding to inaccurate interaction image  $I^{gen}$  as  $T^{anc}$ .
+
+In particular, we observe biased interaction images tend to depict verbs that appear more frequently in the data along human-object pair. For instance, when generating images with the object "backpack," the images commonly show a person wearing a backpack, which is the most frequent verb in the dataset. Based on this observation, we hypothesize that the biased interaction images typically align with frequent verbs in the real data distribution. Therefore, we define  $r^{anc}$  as the verb with the highest sample count for each human-object pair as follows:
+
+$$
+r ^ {a n c} = \underset {r \in R _ {o}} {\arg \max } \mathcal {C} (r | o), \tag {3}
+$$
+
+where  $R_{o}$  denotes the set of verbs related to each human-object pair in dataset  $\mathcal{D}$  and  $\mathcal{C}(\cdot)$  represents the number of samples corresponding to each verb  $r$  when given  $o$ .
+
+Given the anchor verb  $r^{anc}$ , to reduce the generation of biased interaction images, we encourage SD to create images that with distant semantics from  $T^{anc} = "A photo of a \{h\} \{r^{anc}\} a/an \{o\}"$  (e.g. "wearing a backpack"). Specifically, we guide the  $I^{gen}$  closer to the accurate interaction sentence  $T^{gt}$  while pushing away from the sentence with anchor interaction  $T^{anc}$  by leveraging triplet loss [11].
+
+We extract normalized feature  $f^{gen}$ ,  $e^{gt}$ , and  $e^{anc}$  through CLIP [23] encoder  $E_{T}$  and  $E_{I}$  from  $I^{gen}$ ,  $T^{gt}$ , and  $T^{anc}$ , respectively. Then, we set  $e^{gt}$  as positive and  $e^{anc}$  as negative when given anchor  $f^{gen}$  and compute the cosine similarity with the generated image features,  $f^{gen}$ , and each text feature,  $e^{gt}$  and  $e^{anc}$ . Finally, we introduce a triplet loss to maximize the cosine similarity between  $f^{gen}$  and  $e^{gt}$  larger than the similarity between  $f^{gen}$  and  $e^{anc}$  with a margin of  $m$  as follows:
+
+$$
+\mathcal {L} _ {\text {t r i p l e}} = \max  \left(0, m + \operatorname {s i m} \left(f ^ {g e n}, e ^ {g t}\right) - \operatorname {s i m} \left(f ^ {g e n}, e ^ {a n c}\right)\right). \tag {4}
+$$
+
+In addition to distinguishing the semantic difference between verbs, to generate accurate interaction at the image level as well, we add image aligning loss to make the model follow the interaction in  $I^{gt}$ . However,  $I^{gt}$  often contains
+
+multiple humans, objects, and interactions within a single image while we only use a single human, object, and interaction triplet as input during image generation. This input discrepancy makes it difficult for the model to focus solely on the relevant interaction specified by the input prompts, causing confusion when depicting the specified interaction words. To focus on the part in the image corresponding to the interaction label, we extract mask region  $\mathcal{M}$  combining the human and object bounding boxes in a dataset and obtain  $I_{\mathcal{M}}^{gt} = I^{gt} \odot \mathcal{M}$ . Then, we extract  $f_{\mathcal{M}}^{gt}$  through image encoder and minimize the cosine similarity between  $f_{\mathcal{M}}^{gt}$  and  $f^{gen}$  as below:
+
+$$
+\mathcal {L} _ {\text {a l i g n}} = 1 - \frac {f _ {\mathcal {M}} ^ {g t} \cdot f ^ {g e n}}{| f _ {\mathcal {M}} ^ {g t} | | f ^ {g e n} |}. \tag {5}
+$$
+
+Adaptive Interaction Modification. When applying  $\mathcal{L}_{\mathrm{triple}}$  and  $\mathcal{L}_{\mathrm{align}}$  to capture the semantic differences between interaction words, we find that the modification extent applied to the generated images differs across interaction words. As each interaction text  $T^{gt}$  has widely varying sample counts, we also focus on balancing the modification extent across interaction words. Specifically, we adaptively scale the loss considering the number of samples following [4]. We apply an effective number, which is defined as:
+
+$$
+\alpha (k) = \frac {1 - \beta^ {n _ {k}}}{1 - \beta}, \tag {6}
+$$
+
+where  $\beta = (N - 1) / N$ ,  $n_k$  is the number of samples in class  $k$ , and  $\mathbf{N}$  is the total number of samples in the dataset.
+
+Relation Disentanglement Guidance. Finally, we multiply  $\alpha$  on  $\mathcal{L}_{\mathrm{triple}}$  and  $\mathcal{L}_{\mathrm{align}}$ , defining relation disentanglement guidance as follows:
+
+$$
+\mathcal {L} _ {\mathrm {R D G}} = \alpha \cdot \left(\mathcal {L} _ {\text {t r i p l e}} + \mathcal {L} _ {\text {a l i g n}}\right). \tag {7}
+$$
+
+This guidance induces SD to generate images with accurate interactions for each verb and to distinguish the semantic difference between interactions.
+
+# 3.3. IR Module & Interaction Direction Guidance
+
+Relation Disentanglement Guidance (RDG) effectively enables the model to better distinguish and represent each interaction word, helping to exhibit the intended interactions at the image level. However, the model sometimes shows images that misalign with human expectations, especially in the local regions where human-object interactions occur. We hypothesize that focusing on finer, localized interaction regions helps the model improve the detailed interaction expressions, thereby allowing the model to better capture more region-specific interaction information.
+
+To focus more on specific interaction regions, we apply interaction direction guidance (IDG), leveraging interaction regions extracted from the interaction region module with
+
+cross-attention maps corresponding to the  $<\mathrm{h}, \mathrm{r}, \mathrm{o}>$  tokens in generated images. IDG enhances the model's ability to depict region-level details, guiding the depicted interactions to align more perceptually close to human intentions.
+
+Interaction Region Module. To achieve a more detailed interaction expression, we first extract interaction regions from both real and generated images. Although real images contain explicit human and object bounding boxes in the dataset, it is challenging to extract specific interaction regions without additional conditions from generated images. Recently, [9] demonstrated that the cross-attention maps capture the human or object's existence corresponding to each token in prompts. We leverage the cross-attention maps  $\mathcal{A}_h$ ,  $\mathcal{A}_r$  and  $\mathcal{A}_o$  corresponding to  $<\mathrm{h}, \mathrm{r}, \mathrm{o}>$  token to extract interaction region between human and objects.
+
+However, it is nontrivial to directly extract the specific interaction region from a continuous attention map. We apply the centroid extraction mechanism proposed in [5] to compute specific points within attention maps. Specifically, we calculate  $c_{h}$ ,  $c_{r}$  and  $c_{o}$  from each cross-attention maps as follows:
+
+$$
+c = \frac {1}{\sum_ {h , w} \mathcal {A}} \left[ \begin{array}{l} \sum_ {h, w} w \cdot \mathcal {A} \\ \sum_ {h, w} h \cdot \mathcal {A} \end{array} \right], \tag {8}
+$$
+
+where  $h$  and  $w$  denote the height and width in the attention map  $\mathcal{A}$ . Then, we define the interaction center  $c_{rel}$  as the centroid of a triangle defined by  $c_h$ ,  $c_r$ , and  $c_o$ . Finally, we extract interaction region  $B_{rel}^{gen}$  by leveraging the distance between human and object centers with  $B_{rel}^{gen} = c_{rel} \pm \|c_h - c_o\|_2^2$ . Additionally, we extract  $B_{rel}^{gt}$  in the same manner, leveraging the human and object bounding boxes provided in the dataset. We treat  $c_{rel}$  as the midpoint between the human and object centers from the given boxes  $B^{gt}$ .
+
+Interaction Direction Guidance. To generate images with more realistic interactions, we design guidance that aligns feature differences at the image level with those at the interaction region. With interaction region  $B_{rel}^{gt}$  and  $B_{rel}^{gen}$ , we obtain interaction region image by masking the real and generated images:  $I_{rel}^{g} = B_{rel}^{g} \odot I^{g}$ , where  $g \in \{gt, gen\}$ . We then encode both interaction region images through a CLIP image encoder and extract interaction region features  $f_{rel}^{gt}$  and  $f_{rel}^{gen}$ . Leveraging interaction region features, we design the interaction direction guidance to focus the model on modifying the image feature, especially within the interaction region. Specifically, we calculate the direction between  $f_{rel}^{gt}$  and  $f_{rel}^{gen}$ , referred to as biased relation feature  $f_{rel}^{bias} = f_{rel}^{gt} - f_{rel}^{gen}$ , to align the direction of biased interaction features with that of the real images. The interaction direction guidance is as follows:
+
+$$
+\mathcal {L} _ {\mathrm {I D G}} = 1 - \frac {\left(f _ {\mathcal {M}} ^ {g t} - f ^ {g e n}\right) \cdot \left(f _ {r e l} ^ {b i a s}\right)}{\left| f _ {\mathcal {M}} ^ {g t} - f ^ {g e n} \right| \left| f _ {r e l} ^ {b i a s} \right|}. \tag {9}
+$$
+
+# 3.4. Training Phase
+
+We train only the cross-attention layer in SD to capture the semantically distinct relation words, as the cross-attention reflects the existence of each word token [18]. We adopt the reconstruction loss (Eq. 1) to train T2I models in addition to relation disentanglement guidance and interaction direction guidance. However, when an image contains multiple humans and objects, it can lead to multiple interactions that do not match a single target interaction verb. To accurately separate the human and object corresponding to interaction words, we apply mask  $\mathcal{M}$  when calculating the reconstruction loss as follows:
+
+$$
+\mathcal {L} _ {\text {r e c}} = \mathbb {E} _ {z, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| (\epsilon \odot \mathcal {M} - \epsilon_ {\theta} (z _ {t}, t, T) \odot \mathcal {M}) \| _ {2} ^ {2} \right]. \tag {10}
+$$
+
+Finally, we combine reconstruction loss with disentanglement guidance and interaction direction guidance leveraging  $\lambda_1, \lambda_2$ , and  $\lambda_3$  as below:
+
+$$
+\mathcal {L} _ {\text {t o t a l}} = \lambda_ {1} \cdot \mathcal {L} _ {\text {r e c}} + \lambda_ {2} \cdot \mathcal {L} _ {\text {R D G}} + \lambda_ {3} \cdot \mathcal {L} _ {\text {I D G}}. \tag {11}
+$$
+
+# 4. Experiments
+
+We train our model on SD v1.4 at a resolution of  $512 \times 512$ . Using the Adam optimizer with a learning rate of  $4 \times 10^{-6}$ , we trained for a single epoch over 17 hours with a batch size of 12 on 3 Nvidia A6000 GPUs. For image generation during training, we utilize the DDIM scheduler [27] with 30 sampling steps, and during inference, we increase sampling steps to 50. Finally,  $\lambda_{1}, \lambda_{2}$ , and  $\lambda_{3}$  are set to 1.0, 10, and 0.8, respectively, and the margin  $m$  is set to 0.2.
+
+# 4.1. Dataset
+
+We conduct our experiments on the HICO-DET dataset [2], which consists of 47,776 images (38,118 for training and 9,658 for testing). The dataset provides annotations for 600 HOI categories using 80 object categories and 117 verbs.
+
+To train our model, we realign the training set according to the 600 HOI text descriptions and modify annotations to match each image with its corresponding text descriptions. This realignment yields 69,404 images across 600 prompts. We exclude text classes containing "and" (e.g. "A photo of a person and an airplane") to avoid ambiguity from various interactions within the same class, resulting in a final training set of 61,114 images across 501 prompts. More details about the dataset are in the supplementary materials.
+
+# 4.2. Metrics
+
+CLIP & Sentence-BERT Similarity. We assess the model understandability of interaction words by calculating text-to-text similarity scores using both CLIP [23] and SentenceBERT (S-BERT) [24] that we newly adopt to image generation task. After generating captions with InstructBLIP [19],
+
+we compute cosine similarity against ground-truth texts. We also measure the similarity between the input prompts and generated images through CLIP. Details about the S-BERT metric are in the supplementary materials.
+
+HOI Classification Accuracy. To evaluate the robustness of the model understanding interaction words, we measure interaction accuracy with generated images using the SOV-STG [3] HOI detection model pre-trained on HICODET [2]. We report scores on two different weights, SOV-STG-S and SOV-STG-Swin-L, with different backbones (e.g. ResNet [8] and Swin-L [21]). Accuracy is reported in two main settings: Known Object (KO), which measures verb accuracy assuming the objects are correct, and Default (Def.), which assesses both verb and object correctness, with each setting further divided into Full (all classes) and Rare (classes with fewer than 10 samples).
+
+VQA-Score. We additionally use VQAScore [17], a large image-text alignment benchmark for generative models, to evaluate comprehensive interaction alignment scores. This score measures the model's ability to understand more complex compositional prompts, such as spatial and action relationships. For assessing the accuracy of depicting interaction between humans and objects in images, we use two main metrics: the Image-to-Text (I2T) alignment score and the VQA score. The I2T alignment score improves similarity measurement between complex prompts containing interaction words and generated images by incorporating a large language model. The VQA score measures the probability that a VQA model will answer "yes" to questions containing interaction words paired with generated images.
+
+# 4.3. Qualitative Results
+
+Fig. 4 shows a comparison of generated images focusing on the interactions between humans and objects across our method and other models. Compared to previous approaches, our results exhibit more detailed interactions, closely resembling ground-truth images and producing high-quality outputs comparable to images from DALL-E 3 [1], a large-scale text-to-image generation model.
+
+For example, in Fig. 4 (a)-(f), GLIGEN [15] mainly focuses on generating humans and objects, failing to depict the accurate interactions related to the interaction words. While InteractDiffusion [10] improves interaction representation, its results show inaccurate or ambiguous interactions that lack both accurate object rendering (e) or fine-grained interaction details (a)-(d) and (f).
+
+In particular, in Fig.4 (g), our model generates a visually enhanced interaction that closely resembles the ground truth interaction region, showing fluid coming out from a bottle. In the multi-humans with one interaction scenario (f), our results depict people holding foods similar to the image from Dall-E 3, while other models only show people sitting at the table. This demonstrates that our model fur
+
+![](images/ea8e80d0108754070695e5b12030661ea5f49784a7303a807c350731f7f80887.jpg)  
+Figure 4. Interaction comparison in generated images with other models. We generate images using a fixed template: “A {H} {R} a/an {O}”. The top row displays the input human, interaction word, and object. Green and blue boxes represent additional grounding boxes used during image generation for each human and object, respectively. Our results produce images with more accurate interactions than other models, closely resembling ground-truth interactions while maintaining high image quality comparable to large-scale generative models like DALL-E 3 [1].
+
+ther concentrates on detailed interaction regions and distinguishes subtle differences between interaction words with similar spatial arrangements (e.g. "sitting at" and "eating at"). Overall, our model demonstrates a similar level of interaction understandability as a large-scale model such as Dall-E 3 and more accurately exhibits the intended interactions, closely matching with the ground-truth interactions.
+
+We further evaluate our model with complex prompts with multiple interactions to test the ability to distinguish interaction words. As shown in Fig. 1, when given multiple interactions, our model accurately depicts each one, while other models struggle. For example, GLIGEN places objects correctly but fails to represent interactions like "walk
+
+ing," "jumping," and "throwing." InteractDiffusion captures some interactions but misses key objects, like an umbrella. In contrast, our model generates accurate images that reflect the intended interactions. This shows that our model successfully captures the semantic difference between interaction words, generating more detailed interactions. More qualitative results are in the supplementary materials.
+
+# 4.4.Quantitative Results
+
+Tab. 1 compares similarity scores between VerbDiff and previous methods. Our model achieves the highest scores across nearly all evaluation settings. In the text-to-text (T2T) similarity evaluation, VerbDiff outperforms existing
+
+<table><tr><td rowspan="2">Models</td><td colspan="2">CLIP</td><td>S-BERT</td><td colspan="2">VQA-Score</td></tr><tr><td>T2T</td><td>T2I</td><td>T2T</td><td>I2T</td><td>VQA</td></tr><tr><td>HICO-DET</td><td>0.604</td><td>0.233</td><td>0.446</td><td>0.755</td><td>0.754</td></tr><tr><td>SD [25]</td><td>0.725</td><td>0.251</td><td>0.620</td><td>0.769</td><td>0.765</td></tr><tr><td>GLIGEN [15]</td><td>0.683</td><td>0.238</td><td>0.554</td><td>0.679</td><td>0.674</td></tr><tr><td>InteractDiffusion [10]</td><td>0.703</td><td>0.232</td><td>0.575</td><td>0.728</td><td>0.734</td></tr><tr><td>VerbDiff (Ours)</td><td>0.733</td><td>0.242</td><td>0.633</td><td>0.771</td><td>0.766</td></tr></table>
+
+Table 1. Similarity comparison between VerbDiff and other models. We evaluate scores on CLIP, S-BERT [24] and a large vision-language alignment benchmark VQA-Score [17].  
+
+<table><tr><td rowspan="3">Models</td><td colspan="4">SOV-STG-S (Acc ↑)</td><td colspan="4">SOV-STG-Swin-L (Acc ↑)</td></tr><tr><td colspan="2">Def.</td><td colspan="2">KO.</td><td colspan="2">Def.</td><td colspan="2">KO.</td></tr><tr><td>Full</td><td>Rare</td><td>Full</td><td>Rare</td><td>Full</td><td>Rare</td><td>Full</td><td>Rare</td></tr><tr><td>HICO-DET</td><td>26.52</td><td>6.78</td><td>28.68</td><td>7.29</td><td>29.98</td><td>12.66</td><td>31.16</td><td>13.43</td></tr><tr><td>SD [25]</td><td>16.09</td><td>4.59</td><td>18.22</td><td>4.85</td><td>20.08</td><td>8.07</td><td>21.69</td><td>8.66</td></tr><tr><td>GLIGEN [15]</td><td>15.88</td><td>4.85</td><td>17.91</td><td>5.24</td><td>17.83</td><td>7.00</td><td>19.35</td><td>7.57</td></tr><tr><td>InteractDiffusion [10]</td><td>19.67</td><td>7.00</td><td>21.31</td><td>7.69</td><td>23.53</td><td>10.27</td><td>24.86</td><td>11.18</td></tr><tr><td>VerbDiff (Ours)</td><td>22.59</td><td>7.62</td><td>24.79</td><td>7.83</td><td>27.05</td><td>12.60</td><td>28.43</td><td>13.18</td></tr></table>
+
+methods, and the difference becomes even more significant in the S-BERT evaluation. This highlights that the S-BERT metric better captures the semantic differences within interaction verbs. Although the text-to-image (T2I) score is lower than SD, our model achieves higher scores than the models that leverage additional bounding boxes, demonstrating the ability of our model to understand interactions. To assess a more comprehensive interaction understanding of VerbDiff, we also report the VQA-Score, which evaluates how well generative models understand more complex compositional prompts, such as interactions. Although our method shows a slight improvement in the I2T and VQA scores compared to SD, it scores significantly higher than InteractDiffusion.
+
+Tab. 2 presents the HOI accuracy scores, where VerbDiff achieves the highest accuracy across all settings compared to previous methods. In particular, although different backbones are used for computing accuracy, VerbDiff consistently shows the highest accuracy across all settings. This demonstrates that our method has robust interaction word understanding and produces images with precise interactions. Overall, VerbDiff consistently outperforms other models across nearly all evaluation metrics, demonstrating its strong ability to comprehend interaction words and generate high-quality images with accurate human-object interactions, even without extra conditions.
+
+# 4.5. Ablation Study
+
+We show the effectiveness of RDG and IDG in Tab. 3 and Fig. 5. As shown, both RDG (w/  $\mathcal{L}_{\text {triple }}$  and  $\mathcal{L}_{\text {align }}$ ) (b)
+
+![](images/b4fde7740f8d83a829e41e22ad41dfce07c740751338bab4626517108a7f9e31.jpg)  
+Figure 5. Image interaction comparison with and without guidance. IDG focuses on the specific localized interaction region (red box) in the image.
+
+Table 2. HOI accuracy comparison between VerbDiff and previous methods. Def. and KO. refer to Default and Known Object.  
+
+<table><tr><td>VerbDiff</td><td>Lrec</td><td>Ltriple</td><td>Lalign</td><td>LIDG</td><td>CLTP2T</td><td>S-BERT T2T</td><td>HOI Def.</td><td>Acc KO.</td></tr><tr><td>(a)</td><td>✓</td><td></td><td></td><td></td><td>0.691</td><td>0.582</td><td>19.38</td><td>20.89</td></tr><tr><td>(b)</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>0.700</td><td>0.589</td><td>20.32</td><td>21.87</td></tr><tr><td>(c)</td><td>✓</td><td></td><td></td><td>✓</td><td>0.699</td><td>0.588</td><td>20.21</td><td>21.67</td></tr><tr><td>(d)</td><td>✓</td><td>✓</td><td></td><td>✓</td><td>0.710</td><td>0.610</td><td>23.39</td><td>24.51</td></tr><tr><td>All (Ours)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>0.733</td><td>0.633</td><td>27.05</td><td>28.43</td></tr></table>
+
+Table 3. Ablation of the VerbDiff guidance settings. We score the similarity and HOI accuracy on the Full setting. We use the SOV-STG-Swin-L model. Combining all the proposed loss functions shows the best performance.
+
+and IDG (c) enhance the similarity score and HOI accuracy. Specifically, (b) scores slightly higher than (c) demonstrating that the model RDG distinguishes the semantics in interaction words. While scores improve significantly when using IDG and RDG without  $\mathcal{L}_{\text {align }}$ , the model with all guidance terms achieves the highest overall performance. In Fig. 5, we compare images generated with and without each type of guidance. Notably, the enlarged interaction region (red box) shows that the model with IDG produces detailed representations (e.g. human hands), unlike models with only RDG. This indicates that IDG effectively captures fine-grained details in localized interaction regions. More ablations are in the supplementary materials.
+
+# 5. Conclusion
+
+We propose VerbDiff, a novel text-to-image (T2I) diffusion models that address the interaction misunderstanding problem in SD without additional conditions. Although previous methods leverage additional conditions to help the model understand the interaction between humans and objects, they still rely on precise instructions and lack an understanding of the semantic differences between interaction verbs. Our model successfully captures the semantic meanings inherent in interaction words and generates high-quality images with accurate interactions. Extensive experiments demonstrate the effectiveness of our method in enhancing the ability to understand interaction words of T2I models, achieving better interaction comprehension compared to the previous approaches.
+
+Acknowledgments. This was supported by the Institute of Information & Communications Technology Plan-
+
+ning & Evaluation (IITP) grant funded by the Korean government(MSIT) (No.RS-2020-II201373, Artificial Intelligence Graduate School Program(Hanyang University)).
+
+# References
+
+[1] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 6, 7  
+[2] Yu-Wei Chao, Yunfan Liu, Xieyang Liu, Huayi Zeng, and Jia Deng. Learning to detect human-object interactions. In 2018 IEEE winter conference on applications of computer vision (wacv), pages 381-389. IEEE, 2018. 2, 6  
+[3] Junwen Chen, Yingcheng Wang, and Keiji Yanai. Focusing on what to decode and what to train: Efficient training with hoi split decoders and specific target guided denoising. arXiv preprint arXiv:2307.02291, 2023. 6  
+[4] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 5  
+[5] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 5  
+[6] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2, 3  
+[7] Hanan Gani, Shariq Farooq Bhat, Muzammal Naseer, Salman Khan, and Peter Wonka. Lm blueprint: Enabling text-to-image generation with complex and detailed prompts. arXiv preprint arXiv:2310.10640, 2023. 2, 3  
+[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6  
+[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 5  
+[10] Jiu Tian Hoe, Xudong Jiang, Chee Seng Chan, Yap-Peng Tan, and Weipeng Hu. Interactdiffusion: Interaction control in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6180-6189, 2024. 1, 2, 3, 6, 8  
+[11] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In Similarity-based pattern recognition: third international workshop, SIMBAD 2015, Copenhagen, Denmark, October 12-14, 2015. Proceedings 3, pages 84-92. Springer, 2015. 4  
+[12] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan,
+
+and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 3  
+[13] Xu Jia, Takashi Isobe, Xiaomin Li, Qinghe Wang, Jing Mu, Dong Zhou, Huchuan Lu, Lu Tian, Ashish Sirasao, Emad Barsoum, et al. Customizing text-to-image generation with inverted interaction. In ACM Multimedia 2024, 2024. 3  
+[14] Jian-Yu Jiang-Lin, Kang-Yang Huang, Ling Lo, Yi-Ning Huang, Terence Lin, Jhih-Ciang Wu, Hong-Han Shuai, and Wen-Huang Cheng. Record: Reasoning and correcting diffusion for hoi generation. arXiv preprint arXiv:2407.17911, 2024.3  
+[15] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 1, 2, 3, 6, 8  
+[16] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2, 3  
+[17] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2025. 6, 8  
+[18] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7817-7826, 2024. 6  
+[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6  
+[20] Jinxiu Liu and Qi Liu. R3cd: Scene graph to image generation with relation-aware compositional contrastive control diffusion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3657-3665, 2024. 3  
+[21] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6  
+[22] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15579-15591, 2023. 2, 3  
+[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4, 6  
+[24] N Reimers. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 2, 6, 8
+
+[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 8  
+[26] Guibao Shen, Luozhou Wang, Jiantao Lin, Wenhang Ge, Chaozhe Zhang, Xin Tao, Yuan Zhang, Pengfei Wan, Zhongyuan Wang, Guangyong Chen, et al. Sg-adapter: Enhancing text-to-image generation with scene graph guidance. arXiv preprint arXiv:2405.15321, 2024. 3  
+[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6  
+[28] Shengqiong Wu, Hao Fei, Hanwang Zhang, and Tat-Seng Chua. Imagine that! abstract-to-intricate text-to-image synthesis with scene graph hallucination diffusion. Advances in Neural Information Processing Systems, 36, 2024. 3  
+[29] Yinwei Wu, Xingyi Yang, and Xinchao Wang. Relation rectification in diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7685-7694, 2024. 3  
+[30] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 2, 3  
+[31] Bicheng Xu, Qi Yan, Renjie Liao, Lele Wang, and Leonid Sigal. Joint generative modeling of scene graphs and images via diffusion models. arXiv preprint arXiv:2401.01130, 2024. 3  
+[32] Ling Yang, Zhilin Huang, Yang Song, Shenda Hong, Guohao Li, Wentao Zhang, Bin Cui, Bernard Ghanem, and Ming-Hsuan Yang. Diffusion-based scene graph to image generation with masked contrastive pre-training. arXiv preprint arXiv:2211.11138, 2022. 3  
+[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3  
+[34] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2, 3

2025/VerbDiff_ Text-Only Diffusion Models with Enhanced Interaction Awareness/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a594e59da9e7f8ecb8de77dccfc5ee021a8fd3ee00f4c3590ff285cd4ce3be08
+size 853987

2025/ViCaS_ A Dataset for Combining Holistic and Pixel-level Video Understanding using Captions with Grounded Segmentation/ae891128-2b76-44ec-8a3e-060beba26fb9_content_list.json

@@ -0,0 +1,1864 @@
+[
+    {
+        "type": "text",
+        "text": "ViCaS: A Dataset for Combining Holistic and Pixel-level Video Understanding using Captions with Grounded Segmentation",
+        "text_level": 1,
+        "bbox": [
+            104,
+            128,
+            893,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ali Athar Xueqing Deng Liang-Chieh Chen",
+        "bbox": [
+            308,
+            203,
+            686,
+            220
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ByteDance Inc.",
+        "bbox": [
+            434,
+            227,
+            558,
+            244
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "https://ali2500.github.io/vicas-project/",
+        "bbox": [
+            259,
+            252,
+            733,
+            270
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{ali.athar}@bytedance.com",
+        "bbox": [
+            387,
+            277,
+            609,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            327,
+            326,
+            343
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent advances in multimodal large language models (MLLMs) have expanded research in video understanding, primarily focusing on high-level tasks such as video captioning and question-answering. Meanwhile, a smaller body of work addresses dense, pixel-precise segmentation tasks, which typically involve category-guided or referral-based object segmentation. Although both directions are essential for developing models with human-level video comprehension, they have largely evolved separately, with distinct benchmarks and architectures. This paper aims to unify these efforts by introducing ViCaS, a new dataset containing thousands of challenging videos, each annotated with detailed, human-written captions and temporally consistent, pixel-accurate masks for multiple objects with phrase grounding. Our benchmark evaluates models on both holistic/high-level understanding and language-guided, pixel-precise segmentation. We also present carefully validated evaluation measures and propose an effective model architecture that can tackle our benchmark.",
+        "bbox": [
+            88,
+            359,
+            483,
+            647
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            89,
+            676,
+            220,
+            694
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The emergence of open-source large language models (LLMs) [9, 32, 96], coupled with progress in aligning vision and language feature spaces [55, 58, 83], has enabled significant research into vision-language models that can jointly reason over visual and language inputs. Earlier advances in image-language models [8, 56, 66] have spurred development of multimodal LLMs capable of reasoning over other modalities such as videos [63] and 3D point-clouds [128].",
+        "bbox": [
+            89,
+            703,
+            482,
+            824
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In the field of video-language models, research has mainly focused on holistic video understanding tasks such as video captioning and question-answering [63, 67, 72, 73, 113, 127], with a particular emphasis on generating temporally dense text output for long video sequences [18,",
+        "bbox": [
+            89,
+            825,
+            483,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/ffa5ccd61bc4bc1ff6154a9799cf8562893278fe35d62d898b90124f30365264.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            325,
+            903,
+            465
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/01806b4f24f654f99ad9dbea969466b1fd1246951fe515fa781b9f6ad06beb3b.jpg",
+        "image_caption": [
+            "Figure 1. ViCaS Dataset/Benchmark. Our dataset contains detailed video captions with phrase-level grounding for accurate object segmentation masks. The benchmark comprises two tasks to evaluate holistic and pixel-level video understanding, respectively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            516,
+            469,
+            905,
+            616
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "62, 92, 118, 132]. By contrast, comparatively less attention has been given to predicting spatially dense outputs — such as bounding boxes or segmentation masks — that can localize key objects and actors in videos based on text prompts, even though this has important practical applications, e.g., autonomous robots and video editing. Although recent language-driven benchmarks [29, 47, 89, 117] in the video segmentation community have made progress in this area, these datasets are largely object-centric and lack evaluation of high-level, holistic video understanding.",
+        "bbox": [
+            511,
+            715,
+            906,
+            868
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we seek to bridge the gap between benchmarks that evaluate holistic video understanding and those",
+        "bbox": [
+            511,
+            869,
+            906,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "19023",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "focused on pixel-level localization. We introduce ViCaS, a new video dataset and benchmark containing thousands of videos with detailed, human-annotated captions. In these captions, words and phrases corresponding to key objects are grounded in human-drawn, pixel-precise segmentation masks that span the entire video duration. To the best of our knowledge, this is the first video dataset to offer human-labeled annotations of this kind.",
+        "bbox": [
+            89,
+            90,
+            480,
+            210
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our benchmark comprises two tasks: (1) Video Captioning, which requires describing the video events and objects in detail, and (2) our newly proposed Language-Guided Video Instance Segmentation (LG-VIS) task, which requires predicting temporally-consistent segmentation masks for multiple objects based on a text prompt. An example of our dataset annotations and tasks is given in Fig. 1. To effectively evaluate Video Captioning, we conduct a comprehensive user study to validate the evaluation measures used for open-ended text similarity by comparing several existing approaches. Finally, we introduce VideoLLaVA-Seg, an end-to-end architecture designed to effectively tackle our benchmark by integrating insights from recent vision-language models [63, 67] and prompt-based video segmentation approaches [53, 86]. In summary, our contributions are as follows:",
+        "bbox": [
+            91,
+            212,
+            483,
+            450
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We introduce a large, human-annotated video dataset with detailed text captions with phrase-level grounding for objects, accompanied by pixel-precise segmentation masks. This dataset enables the evaluation of both holistic and pixel-level video understanding.",
+            "- We propose accurate, reproducible evaluation measures for open-ended text similarity which are verified by a comprehensive user study.",
+            "- We present Video-LLaVA-Seg, an effective, end-to-end trained architecture that can tackle our benchmark."
+        ],
+        "bbox": [
+            89,
+            454,
+            482,
+            603
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            89,
+            617,
+            232,
+            632
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Although video understanding is a well-researched topic in computer vision, researchers have traditionally approached holistic and pixel-level video understanding as separate streams, each with its own datasets and benchmarks. Below, we review related work in these areas.",
+        "bbox": [
+            89,
+            642,
+            482,
+            717
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Holistic Video Understanding",
+        "text_level": 1,
+        "bbox": [
+            89,
+            726,
+            354,
+            742
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Video Classification. This is one of the earliest tasks in video understanding, popularized by activity recognition datasets [52, 93, 125], and later expanded by larger datasets [37, 38, 91], e.g., Kinetics [46]. Early deep learning approaches employed 3D CNNs [39, 45, 97-99] to capture spatio-temporal information. With the popularization of transformers [100], attention-based architectures [4, 68, 79] emerged, offering improved performance. Although these datasets and architectures have greatly advanced video understanding, they are limited from a language perspective,",
+        "bbox": [
+            89,
+            750,
+            482,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "typically assigning a single, predefined label to each video.",
+        "bbox": [
+            511,
+            90,
+            901,
+            107
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Video Captioning and Question-Answering. Alongside classification, language-oriented tasks such as video captioning and question-answering (Q/A) have gained research attention. Early datasets like MSVD [17], MSR-VTT [111], and TGIF-QA [61] laid the groundwork for video captioning and were later adapted as Q/A benchmarks [110]. Initial approaches [31, 112] combined CNNs for visual reasoning with RNNs for text generation. As transformers gained popularity, architectures like VideoBERT [94] and UniVL [71] were among the first to unify vision and language by learning shared representations for both modalities.",
+        "bbox": [
+            511,
+            111,
+            903,
+            277
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Multimodal Large Language Models (MLLMs). The recent popularization of LLMs [9, 24, 32, 96] has enabled research on multimodal models which extend LLMs to process visual inputs such as images [14, 19, 28, 66, 124] and videos [18, 62, 63, 72, 73, 92]. This research has been supported by large-scale video captioning datasets [11, 20, 116] and multi-task video understanding benchmarks [25, 57].",
+        "bbox": [
+            511,
+            284,
+            903,
+            388
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Overall, the datasets and models discussed above focus primarily on high-level, holistic video understanding and do not address finegrained, pixel-level localization.",
+        "bbox": [
+            511,
+            391,
+            903,
+            435
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Pixel-level Video Understanding",
+        "text_level": 1,
+        "bbox": [
+            511,
+            446,
+            795,
+            463
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Object Tracking and Segmentation. Object localization and tracking is a deeply studied problem, even prior to the deep learning era [2, 78]. This research has been propelled by several datasets, with early efforts focused on bounding-box-level object tracking [26, 27, 34, 36, 44, 51, 122]. As model architectures advanced, benchmarks for pixel-precise video segmentation emerged, covering tasks such as object segmentation based on predefined categories [7, 48, 74, 75, 81, 82, 102, 108, 119] or segmenting specific objects given their first-frame ground-truth masks [30, 80, 115] or points [7, 43, 133]. Popular approaches have evolved from tracking-by-detection methods [16, 69, 95, 102] to end-to-end trainable transformer-based [6, 40, 42, 49, 90, 107, 109, 130], and object-level attention-based architectures [5, 21–23, 76]. Despite substantial progress, these works focus solely on tracking and segmentation, without addressing holistic video understanding.",
+        "bbox": [
+            511,
+            470,
+            906,
+            728
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Language-Guided Segmentation. Recent progress in vision-language models has motivated multiple language-guided video segmentation datasets [103]. These include \"open-vocabulary\" datasets [104, 105] that cover a large set of object classes, and referral-based video segmentation benchmarks [29, 47, 89], which require segmenting objects based on text prompts. Although more language-oriented, these datasets are still object-centric and do not involve predicting a high-level description of the entire video.",
+        "bbox": [
+            511,
+            734,
+            903,
+            869
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Some approaches [29, 41, 123] for these tasks use text and image backbones to encode the text prompt and video",
+        "bbox": [
+            511,
+            869,
+            903,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "19024",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/416e52fa60cb6ae45bcff030bd73826d136a89c88ef4484587a51b52ecc45305.jpg",
+        "image_caption": [],
+        "image_footnote": [
+            "A woman in blue rides her skateboard down a ramp, but falls and ends up lying on her side. Another woman in red riding her skateboard rushes over to assist her. The two skateboards continue to roll around on the ground."
+        ],
+        "bbox": [
+            258,
+            90,
+            418,
+            162
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/1bda4780b43798693ca2ecf1eec7a31060344a31aeae03fc89cac974b1e09037.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            419,
+            90,
+            578,
+            162
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/afcbd858785c20a9063e186a807fb019cfc8b97d3870ecd049bd91806056cee7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            578,
+            90,
+            738,
+            162
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/7a94198028c4ebd5796f5586b05ce2cd9ba14914c933b9577562869a5e0f10be.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            738,
+            90,
+            898,
+            162
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/97db2441fb90d6ec149392e139a5c1356a09f186835b7489a19c1c60f4f52160.jpg",
+        "image_caption": [],
+        "image_footnote": [
+            "A white cat sniffs a purple cup on the desk in an indoor room. It then knocks the purple cup onto the floor, which causes the three yellow balls inside it to fall out."
+        ],
+        "bbox": [
+            259,
+            165,
+            418,
+            236
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/89142621276ef55c419a72a6cc2d179f7ad63dea2f4b9a5ab3bd64d31822bc1a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            419,
+            165,
+            578,
+            236
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/5377d021e765b0e0e650ef864104cc63bb6668d70a56183f54721f1e35cf48f3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            578,
+            165,
+            738,
+            236
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/23ebfc4728149226244b2e956ed3d58011818a8d0579883988386b31a5d3cb73.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            738,
+            165,
+            898,
+            236
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/8f716cfb673f40ec9b3b777deea91e5c530b40150c6be86ee95f3cdf3ecdefca.jpg",
+        "image_caption": [
+            "Figure 2. ViCaS Examples. Our dataset showcases diverse scenes with a variety of objects and video events, along with detailed captions. Phrases referring to multiple objects are written with multiple colors, e.g., \"three yellow balls\" in row 2 references three different objects."
+        ],
+        "image_footnote": [
+            "A black car is driving in front on a two-way road. A white car drives by on the left. Then, the car in front speeds up while a second white car drives by and a silver car very dangerously overtakes it."
+        ],
+        "bbox": [
+            259,
+            239,
+            416,
+            310
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/b88e1d750fcdae64be1bb0423ea29dc4d95dde1ac7ac33279001606171cc3d42.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            419,
+            239,
+            578,
+            310
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/96da968836f806f7ae69b732638746869eb2980bf830e6f972c70aabc19c8706.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            578,
+            239,
+            738,
+            310
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/811e1b84077f377c3dd62e84092e9204549a831f95142b31e9a4ad4334e8c709.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            738,
+            239,
+            898,
+            310
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "frames, followed by a transformer decoder and segmentation head to generate masks. Meanwhile, recent LLM-based approaches [10, 117] use a special SEG vocabulary token in conjunction with a segmentation network to predict the target masks. These approaches typically utilize several image-level and video segmentation datasets for training, and utilize other models [13, 62] for dynamic frame selection or temporal mask propagation.",
+        "bbox": [
+            88,
+            378,
+            482,
+            498
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In contrast to these two categories of works, we propose a unified dataset that provides the annotations as well as benchmark tasks to evaluate both holistic/high-level, and pixel-level video understanding. Additionally, we propose a baseline architecture that is end-to-end trainable, and can be effectively trained for segmentation using only our dataset, thereby making it easy to setup and extend.",
+        "bbox": [
+            88,
+            500,
+            483,
+            604
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. ViCaS Dataset",
+        "text_level": 1,
+        "bbox": [
+            89,
+            617,
+            236,
+            633
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our dataset is designed to evaluate both holistic as well as pixel-level video understanding. To this end, we annotate 20,416 videos with detailed captions in which words/phrases referencing salient objects are grounded with temporally consistent segmentation masks. All annotations are done by professional human annotators.",
+        "bbox": [
+            89,
+            642,
+            483,
+            733
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Video Source",
+        "text_level": 1,
+        "bbox": [
+            89,
+            741,
+            230,
+            756
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For our objective, it is crucial to annotate videos which contain meaningful events that can only be explained through effective temporal reasoning. At the same time, to effectively evaluate pixel-level understanding, the videos should feature challenging scenes with multiple moving objects.",
+        "bbox": [
+            89,
+            763,
+            482,
+            840
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To meet these requirements, we annotate videos from three sources: (1) Oops dataset [33], a collection of 'fail videos' from the internet, (2) Unidentified Video Object (UVO) [105], which is a video segmentation dataset, and",
+        "bbox": [
+            89,
+            840,
+            483,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "(3) Kinetics-700 [46], a popular human action recognition dataset. These sources are suitable for our dataset since they contain in-the-wild videos with diverse objects and backgrounds. Moreover, they contain videos with multiple objects undergoing appearance and shape changes, and fast motion. This combination of attributes makes the videos challenging both in captioning and in segmentation.",
+        "bbox": [
+            511,
+            378,
+            903,
+            484
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We annotate 20,416 videos with durations ranging from 4 to 30 seconds (distribution illustrated in Fig. 3c). Note that we use only the raw videos from these Oops dataset and disregard their annotations. More details on video selection are given in supplementary.",
+        "bbox": [
+            511,
+            484,
+            905,
+            560
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Annotation Process",
+        "text_level": 1,
+        "bbox": [
+            513,
+            570,
+            697,
+            584
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The annotation process consists of two main steps: first, a detailed caption is written (Step 1a) in which objects selected for segmentation masks are identified and marked using a specific syntax (Step 1b). In the second step, a different annotator reviews the video and its text, and draws the corresponding segmentation masks (Step 2). The process is illustrated in Fig. 4. Further details are given below.",
+        "bbox": [
+            511,
+            594,
+            905,
+            699
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Step 1a: Text Captions. Professional annotators fluent in English were tasked with writing a detailed caption covering the events in each video. To ensure consistency in the level of detail and writing style, they received comprehensive guidelines with several examples. Annotators were advised to think as if they have  $\\sim 30$  seconds to convey the video content in as much detail as possible over an audio call. They were instructed to describe the 'fail' event [33], the actions and movements of objects, as well as the background scene and any additional, interesting elements. Since our dataset is video-centric, annotators were explicitly guided to include temporal information, with the instruction: \"Focus on details that require watching the en",
+        "bbox": [
+            511,
+            704,
+            906,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "19025",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/a34eb619ccc6e6324ac2d826eafc7183be22c4d87589cf47ae38d910e89c88f1.jpg",
+        "image_caption": [
+            "(a)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            85,
+            349,
+            243
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/cc697eea83b4d4c7d603c44bafec5d153f840487ffbbb5148fb4cba58ecfbf43.jpg",
+        "image_caption": [
+            "(b)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            361,
+            89,
+            630,
+            241
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/4347307a66e839e745049a67d397f9dbb905dc9c8e33cf2c064ce518f5bb560c.jpg",
+        "image_caption": [
+            "(c)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            643,
+            90,
+            903,
+            241
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/eb287631c6cf06dfe5343ab52e50a2af3bad35bbce87d54b6dc087b00116e11c.jpg",
+        "image_caption": [
+            "Figure 3. ViCaS Dataset. (a): Word cloud for nouns in grounding phrases. (b), (c): Histograms for caption length and video duration, respectively. Since these distribution are long-tailed, the right-most bar labeled  $n+$  captures all values  $\\geq n$  to prevent visual distortion.",
+            "Figure 4. Annotation Process. Our annotation process consists of two steps. Step 1: A detailed video caption is written by human annotators. Salient objects are identified and marked using a special syntax (highlighted in blue). Step 2: Segmentation masks are drawn for the salient objects throughout the video."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            320,
+            483,
+            473
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "tire video and cannot be inferred from a few frames.\"",
+        "bbox": [
+            89,
+            585,
+            442,
+            599
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Step 1b: Phrase Grounding. Alongside writing the text caption, annotators were instructed to identify salient objects/actors in the video, and mark the corresponding phrases or words using a specific syntax, as illustrated in Fig. 4. This syntax is programmatically parsable to extract object IDs that can be removed to produce a standard, human-readable caption. To streamline the mask-drawing process, annotators were advised to mark no more than 10 objects per video.",
+        "bbox": [
+            89,
+            604,
+            483,
+            742
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Step 2: Segmentation Masks. In the second step, a different annotator reviews the video and its caption, and uses a polygon tool to draw segmentation masks for each marked object. The annotators were instructed to read the text carefully and watch the video multiple times to ensure that correct masks were drawn for the each grounding phrase. To balance annotation accuracy and cost, we annotate masks at 1 frame-per-second (fps) and then use an off-the-shelf SAM2 [86] model to propagate these masks to adjacent frames, producing temporally dense mask annotations at 30",
+        "bbox": [
+            89,
+            750,
+            483,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "fps. A similar approach has been successfully applied in other video segmentation datasets [7, 105].",
+        "bbox": [
+            511,
+            324,
+            906,
+            356
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Annotation Statistics and Quality Control. For Step 1, we employed 24 annotators and 4 quality inspectors who reviewed/corrected the text captions and language grounding. On average, each video required 9 minutes for annotation and quality checking. For Step 2, we hired 40 annotators and 5 quality inspectors, with each video frame requiring an average of 7 minutes to annotate.",
+        "bbox": [
+            511,
+            359,
+            908,
+            467
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Benchmark Design and Evaluation Tasks",
+        "text_level": 1,
+        "bbox": [
+            511,
+            479,
+            864,
+            494
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Leveraging our annotations, we create a benchmark to evaluate both high-level video understanding and pixel-precise segmentation. Consequently, it includes two tasks:",
+        "bbox": [
+            511,
+            503,
+            906,
+            549
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1) Video Captioning. In this task, the model is expected to produce an open-ended text summary that explains the events in the video, including descriptions of salient objects and background elements.",
+            "2) Language-Guided Video Instance Segmentation (LG-VIS). For this task, the model receives a text prompt describing a specific set of objects, e.g., 'Where is the person who walks from left to right?', and is expected to predict segmentation masks for the corresponding objects. For cases where multiple objects are referenced, e.g., 'Where is the group of kittens that are playing?', a separate mask is required for each object. This is in contrast to Referral Video Object Segmentation [29, 47, 89], where only a single, binary mask covering all target objects is required."
+        ],
+        "bbox": [
+            511,
+            554,
+            906,
+            772
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "From a vision-language perspective, both tasks are forms of Video Question-Answering. For Video Captioning, the question asks for a video description, with the answer provided as open-ended text. For LG-VIS, the question is a 'Where is?' query regarding specific objects, and the answer is a video-length segmentation mask for each object.",
+        "bbox": [
+            511,
+            773,
+            905,
+            864
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Converting Grounded Captions to LG-VIS Prompts. As outlined in Sec. 3.2, our annotations contain phrase ground-",
+        "bbox": [
+            511,
+            869,
+            905,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "19026",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/1fe8c66398dc21fb4e4677444b86c5857f9e1ce0ca97c0f5b571b2a71547eca2.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Type</td><td>Dataset Name</td><td>Year</td><td>Human \nAnn.</td><td>Videos</td><td>Duration \n(sec)</td><td>Caption \n(words)</td><td>Object \nClasses</td><td>Object \nTracks</td><td>Object \nMasks</td></tr><tr><td rowspan=\"8\">Captioning and Q/A</td><td>MSVD [17]</td><td>2011</td><td>✓</td><td>2K</td><td>9.7</td><td>8.7</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MSR-VTT [110]</td><td>2016</td><td>✓</td><td>10K</td><td>15.0</td><td>9.3</td><td>-</td><td>-</td><td>-</td></tr><tr><td>ActivityNet [125]</td><td>2017</td><td>✓</td><td>20K / 100K</td><td>36.0</td><td>13.5</td><td>-</td><td>-</td><td>-</td></tr><tr><td>YouCook2 [131]</td><td>2018</td><td>✓</td><td>2K / 14K</td><td>19.6</td><td>8.8</td><td>-</td><td>-</td><td>-</td></tr><tr><td>VATEX [106]</td><td>2019</td><td>✓</td><td>41K</td><td>~10</td><td>15.2</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DiDeMo [3]</td><td>2017</td><td>✓</td><td>10K / 27K</td><td>6.9</td><td>8.0</td><td>-</td><td>-</td><td>-</td></tr><tr><td>WebVid10M [11]</td><td>2021</td><td>X</td><td>10M</td><td>17.2</td><td>14.2</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Panda70M [20]</td><td>2024</td><td>X</td><td>3.8M / 70.8M</td><td>8.5</td><td>13.2</td><td>-</td><td>-</td><td>-</td></tr><tr><td rowspan=\"5\">Referral VOS</td><td>A2D Sentence [35]</td><td>2018</td><td>✓</td><td>3,782</td><td>4.9</td><td>-</td><td>-</td><td>4,825</td><td>58K</td></tr><tr><td>DAVIS-17 RVOS [47]</td><td>2019</td><td>✓</td><td>90</td><td>2.4</td><td>-</td><td>78</td><td>205</td><td>13.5K</td></tr><tr><td>Refer-Youtu-VOS [89]</td><td>2020</td><td>✓</td><td>3,978</td><td>5.5</td><td>-</td><td>94</td><td>7,451</td><td>131K</td></tr><tr><td>MeViS [29]</td><td>2023</td><td>✓</td><td>2,006</td><td>13.2</td><td>-</td><td>-</td><td>8,171</td><td>443K</td></tr><tr><td>ReVOS [117]</td><td>2024</td><td>✓</td><td>1,042</td><td>8.6</td><td>-</td><td>-</td><td>9,084</td><td>273K</td></tr><tr><td rowspan=\"5\">Class-guided Segmentation</td><td>YouTube-VIS [119]</td><td>2019</td><td>✓</td><td>2,883</td><td>5.6</td><td>-</td><td>40</td><td>4,883</td><td>131K</td></tr><tr><td>OVIS [81]</td><td>2021</td><td>✓</td><td>901</td><td>12.8</td><td>-</td><td>25</td><td>5,223</td><td>296K</td></tr><tr><td>UVO [105]</td><td>2021</td><td>✓</td><td>11,228</td><td>9.6</td><td>-</td><td>80</td><td>104,898</td><td>593K</td></tr><tr><td>BURST [7]</td><td>2023</td><td>✓</td><td>2,914</td><td>35.7</td><td>-</td><td>482</td><td>16,089</td><td>600K</td></tr><tr><td>LV-VIS [104]</td><td>2023</td><td>✓</td><td>4,828</td><td>4.6</td><td>-</td><td>1,196</td><td>25,588</td><td>544K</td></tr><tr><td></td><td>ViCaS (Ours)</td><td>2025</td><td>✓</td><td>20,416</td><td>9.1</td><td>39.0</td><td>11,492</td><td>65,588</td><td>18.3M</td></tr></table>",
+        "bbox": [
+            91,
+            88,
+            903,
+            414
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Dataset Comparison. In the 'Videos' column, entries formatted as 'X / Y' indicate a total of X videos with Y labeled sub-clips, with the average sub-clip duration being reported.",
+        "bbox": [
+            89,
+            425,
+            906,
+            455
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "ing for labeled objects. We use GPT-4 [1] to convert these grounded captions into 'Where is?' style questions, which serve as text prompts for the LG-VIS task. Further details about this process are provided in supplementary.",
+        "bbox": [
+            89,
+            479,
+            482,
+            542
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Dataset Splits. We split the 20,416 videos into train, validation, and test sets with 14,516 videos for training, and 2,950 each for validation and testing. Statistics for each set are given in supplementary.",
+        "bbox": [
+            89,
+            545,
+            483,
+            608
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4. Comparison with Related Datasets",
+        "text_level": 1,
+        "bbox": [
+            89,
+            614,
+            393,
+            632
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1 provides a quantitative comparison between ViCaS and existing datasets, showing that ViCaS uniquely combines text captions and pixel-precise segmentation masks.",
+        "bbox": [
+            89,
+            638,
+            482,
+            685
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Captioning and Q/A. For text descriptions, ViCaS prioritizes quality over quantity, offering detailed, high-quality, human-written captions with phrase grounding for objects. Our captions average 39 words (distribution illustrated in Fig. 3b), significantly longer than those in other datasets. While ViCaS annotates more videos than early datasets like MSVD [17] and MSR-VTT [111], later datasets with temporally dense captions, such DiDeMo [3], and YouCook2 [131], include more labeled video segments. Recently, large-scale datasets such as WebVid10M [11] and Panda70M [20] have emerged, containing millions of automatically captioned videos sourced from the internet. While useful for (pre-)training large models, these datasets offer shorter, less detailed descriptions and have neither segmen",
+        "bbox": [
+            89,
+            689,
+            482,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "tation masks nor phrase grounding.",
+        "bbox": [
+            511,
+            479,
+            746,
+            494
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Referral VOS. Compared to Referral Video Object Segmentation (VOS) datasets, ViCaS stands out by providing holistic, video-level captions in addition to object-level language expressions. With 20,416 annotated videos, ViCaS is significantly larger than the second-largest ReferYouTubeVOS [89], which contains only 3,978 videos. Although MeViS [29] features longer videos, averaging 13.2s compared to our 9.1s, ViCaS offers significantly more annotated objects (65,588) compared to MeViS (8,171).",
+        "bbox": [
+            511,
+            500,
+            906,
+            638
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Class-Guided Segmentation. Compared to these datasets, ViCaS provides more nuanced language labels, going beyond single-word category labels typical in Video Instance Segmentation (VIS) [81, 119] datasets. Using NLTK [15], we found that our object-grounding phrases contains 11,492 unique nouns/noun- phrases (illustrated in Fig. 3a), which is significantly higher than the 1,196 object categories in LV-VIS [104]. Furthermore, ViCaS contains more videos than the others, and also contains more labeled object tracks than all other datasets except UVO [105]. Lastly, ViCaS stands out by providing detailed text captions in addition to the segmentation masks.",
+        "bbox": [
+            511,
+            642,
+            908,
+            824
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Evaluation Measures",
+        "text_level": 1,
+        "bbox": [
+            511,
+            839,
+            714,
+            854
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As outlined in Sec. 3.3, our benchmark comprises two tasks: Video Captioning and LG-VIS. Evaluating the video caption quality is challenging since it requires computing",
+        "bbox": [
+            511,
+            854,
+            906,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "19027",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/f1931e90c2c87affc1e68df7ff1d98d5591f5381bd627d7e27bab1f64093b87e.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\" colspan=\"2\">Evaluation Method</td><td colspan=\"3\">Part 1</td><td>Part 2</td></tr><tr><td>r</td><td>rmax</td><td>ravg</td><td>ΔPN</td></tr><tr><td rowspan=\"3\">LLM</td><td>GPT4 [1]</td><td>58.9</td><td>62.1</td><td>67.3</td><td>37.6</td></tr><tr><td>Llama3-70B [32]</td><td>59.3</td><td>64.3</td><td>65.6</td><td>29.2</td></tr><tr><td>Llama3-8B [32]</td><td>31.3</td><td>-</td><td>45.9</td><td>2.4</td></tr><tr><td rowspan=\"6\">Embedding Sim.</td><td>Angle Lm2-7B [60]</td><td>50.5</td><td>50.0</td><td>50.3</td><td>3.3</td></tr><tr><td>STSB Roberta-L [87]</td><td>49.5</td><td>48.3</td><td>49.1</td><td>7.5</td></tr><tr><td>Angle Lm2-13B [60]</td><td>45.6</td><td>47.4</td><td>46.4</td><td>4.5</td></tr><tr><td>Angle UAE-L [60]</td><td>44.5</td><td>42.7</td><td>40.5</td><td>0.9</td></tr><tr><td>MPNet-B [87]</td><td>41.5</td><td>35.4</td><td>34.7</td><td>0.4</td></tr><tr><td>BERTScore [129]</td><td>22.1</td><td>22.9</td><td>26.2</td><td>-2.4</td></tr><tr><td rowspan=\"4\">Word mat.</td><td>METEOR [12]</td><td>22.3</td><td>23.9</td><td>22.8</td><td>-16.4</td></tr><tr><td>BLEU4 [77]</td><td>19.9</td><td>15.9</td><td>15.4</td><td>-27.6</td></tr><tr><td>ROUGE-L [64]</td><td>18.1</td><td>18.2</td><td>13.2</td><td>-23.7</td></tr><tr><td>CIDEr [101]</td><td>2.8</td><td>0.1</td><td>-0.8</td><td>-21.2</td></tr></table>",
+        "bbox": [
+            91,
+            88,
+            483,
+            303
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Caption Scoring User Study. Results for Part 1 (Correlation with Human Scores) and Part 2 (Robustness to Hard Positives/Negatives).  $r$ ,  $r_{max}$  and  $r_{avg}$  denote Pearson correlation coefficients.  $\\Delta_{PN}$  is the absolute difference in predicted scores.",
+        "bbox": [
+            89,
+            313,
+            482,
+            369
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "the similarity between two open-ended text passages—the human-written ground truth and the model's prediction. To address this, we first conduct a comprehensive user study for scoring open-ended text similarity in Sec. 4.1, before introducing the selected evaluation measures in Sec. 4.2.",
+        "bbox": [
+            89,
+            395,
+            483,
+            470
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. User Study for Video Caption Scoring",
+        "text_level": 1,
+        "bbox": [
+            89,
+            479,
+            419,
+            496
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To decide evaluation measures for video captioning, we carried out a two-part user study comparing several scoring methods which are categorized into three groups: (1) classical text similarity metrics, e.g., METEOR [12] and BLEU4 [77], which rely on word/phrase matching; (2) embedding similarity-based measures, e.g., BERTScore [129] and Angle [60], which summarize the ground-truth and prediction as an embedding using a language model, followed by computing the cosine similarity between the two; and (3) recent LLM-based scoring methods [73], where an LLM is provided with the ground-truth and predicted captions along with instructions for assessing similarity, and it outputs a similarity score directly as text.",
+        "bbox": [
+            88,
+            502,
+            482,
+            699
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Part 1: Correlation with Human Scores. In this part, 20 participants rated the accuracy of various model-predicted video captions on a scale of 0 to 10, with a minimum interval of 0.5. Similar to the evaluation methods, participants could only read the ground-truth caption and could not watch the video itself. They represented 10 different nationalities and had either native, or professional English proficiency. A total of 131 video captions were evaluated, with each sample scored by two different individuals.",
+        "bbox": [
+            88,
+            703,
+            482,
+            839
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To filter samples with diverging human scores, we calculated the difference  $\\Delta$  between the two assigned scores for each sample, discarding any with  $\\Delta >\\alpha$  (i.e., outlier removal). We used multiple thresholds  $\\{\\alpha_{i}\\}$  ranging from",
+        "bbox": [
+            89,
+            839,
+            483,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "0.5 to 2.5 in steps of 0.5. For each  $\\alpha_{i}$  and each method, we calculated the Pearson correlation coefficient  $r_i$  between averaged human scores and the predicted scores. The final correlation score  $r$  for each method was computed as the mean of the set  $\\{r_i\\}$ , and is reported in Table 2.",
+        "bbox": [
+            511,
+            90,
+            903,
+            167
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Results indicate that LLM-based scoring methods performed best, with Llama3-70B [32] achieving  $r = 59.3\\%$ , closely followed by GPT-4 [1] with  $58.9\\%$ . Model size significantly impacted performance, as evidenced by Llama3-8B's lower  $r = 31.3\\%$ . Among embedding-based methods, AngI E Llama2-7B [60] performed best at  $50.5\\%$ , while classical word/phrase matching metrics lagged behind, with METEOR [12] achieving only  $r = 22.3\\%$ .",
+        "bbox": [
+            511,
+            167,
+            906,
+            289
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Pooling Scores Across Multiple Ground Truths. A common strategy to enhance robustness in text similarity evaluation is to compare predictions with multiple, synonymous ground-truths. Using GPT-4 [1], we generated four rephrasings for each human-written caption, resulting in five ground-truth variants per video. For each video, we calculated the similarity between the prediction and all ground-truths, followed by applying either average or maximum pooling to obtain the final score. We report the resulting correlations as  $r_{\\mathrm{avg}}$  and  $r_{\\mathrm{max}}$  in Table 2. As shown in the table, LLM-based scoring methods benefited most from average pooling, with Llama3-70B improving from  $59.3\\%$  to  $65.6\\%$  and GPT-4 from  $58.9\\%$  to  $67.3\\%$ . Embedding-based methods generally saw minimal improvements for both types of pooling, with some, like MPNet-B and AngI E-UAE-L, experiencing slight declines. Word/phrase matching metrics similarly showed no substantial improvement.",
+        "bbox": [
+            511,
+            294,
+            906,
+            551
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Part 2: Robustness to Hard Positives/Negatives. Here, participants watched videos from our dataset and read the ground-truth captions (GT), and were then asked to write a hard-positive (HP) and a hard-negative (HN) caption for each video. For the HP, they rephrased the caption using different wording and style while retaining all original information. For the HN, they altered critical details while retaining most wording and sentence structure. Each of the 10 participants processed 5 videos, yielding 50 samples.",
+        "bbox": [
+            511,
+            556,
+            905,
+            694
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We computed the similarity between GT-HP and GT-HN pairs for each video using different methods. The scores were scaled to  $[0,100]^1$  and the difference  $\\Delta_{\\mathrm{PN}}$  between the GT-HP and GT-HN pairs was calculated, and reported in Table 2. Ideally,  $\\Delta_{\\mathrm{PN}}$  should be high, indicating that methods assign higher similarity to GT-HP pairs and lower similarity to GT-HN pairs. GPT-4 achieved the highest  $\\Delta_{\\mathrm{PN}}$  (37.6), followed by Llama3-70B (29.2). Most embedding-based models showed low  $\\Delta_{\\mathrm{PN}}$ , indicating similar scores for GT-HP and GT-HN pairs, while word/phrase matching metrics performed worst, with negative  $\\Delta_{\\mathrm{PN}}$  values suggesting higher scores for GT-HN pairs than GT-HP pairs.",
+        "bbox": [
+            511,
+            694,
+            905,
+            876
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "1 Except CIDEr, which is not a normalized metric",
+        "bbox": [
+            531,
+            886,
+            795,
+            898
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "19028",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/dc7dc5b271fa6a7b642f0e57e240abe60e7e094931b63cd6abc586f0a73761f1.jpg",
+        "image_caption": [
+            "Figure 5. Video-LLaVA-Seg Architecture. The vision backbone and projection MLP encode the input video frames into a set of features  $\\mathcal{F}_{sf}$  which are concatenated with the text embeddings and input to the LLM. For Video Captioning, the output  $\\mathbf{F}_o$  is decoded into text. For LG-VIS, the <SEG> token in  $\\mathbf{F}_o$  is applied to the mask decoder along with multi-scale features from the segmentation backbone."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            89,
+            903,
+            262
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2. Selected Evaluation Measures",
+        "text_level": 1,
+        "bbox": [
+            89,
+            342,
+            354,
+            356
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Video Captioning. Although GPT-4 achieved the highest  $r_{\\mathrm{avg}}$  and  $\\Delta_{\\mathrm{PN}}$  in our study, it is proprietary and subject to updates, which complicates reproducibility. Moreover, its runtime can be slow due to token limits, and it incurs monetary costs which effectively makes the benchmark 'pay-to-evaluate'. To support open-source, reproducible research, we accept a slight performance trade-off and select Llama3-70B as the evaluation measure for our video captioning task. Since  $r_{\\mathrm{avg}}$  is noticeably higher than  $r$  for Llama3-70B (Table 2), we generate 4 reworded variants of each ground-truth caption, and compute the final accuracy by averaging over the prediction-GT scores for each video. This score lies in [0, 5] and is abbreviated as CA (Caption Accuracy).",
+        "bbox": [
+            88,
+            359,
+            483,
+            556
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Language-Guided Video Instance Segmentation (LG-VIS). Unlike text similarity, segmentation mask accuracy is a well-studied problem with several feasible metrics [59, 70, 119]. We select Track mean Average Precision (mAP) as our primary metric since it is widely used by video segmentation benchmarks [81, 105, 119] and accommodates multi-object prediction.",
+        "bbox": [
+            88,
+            561,
+            483,
+            667
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5. Video-LLaVA-Seg Model",
+        "text_level": 1,
+        "bbox": [
+            89,
+            672,
+            326,
+            690
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We propose an effective baseline called Video-LLaVA-Seg that can tackle both Video Captioning and LG-VIS with a single, end-to-end trained model.",
+        "bbox": [
+            89,
+            698,
+            483,
+            743
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1. Architecture",
+        "text_level": 1,
+        "bbox": [
+            89,
+            751,
+            225,
+            766
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Video-LLaVA-Seg extends the popular LLaVA [65-67] architecture with segmentation ability. It is illustrated in Fig. 5 and comprises 3 main parts: (1) a multi-modal LLM, (2), a vision backbone, and (3) a segmentation network.",
+        "bbox": [
+            89,
+            773,
+            482,
+            835
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Vision Features. We input  $T$  uniformly sampled frames from the video to the vision backbone, yielding a set of video features  $\\mathbf{F}_v\\in \\mathbb{R}^{T\\times H\\times W\\times C}$ . Next, we adopt the workflow proposed by Xu et al. [114] and split  $\\mathbf{F}_v$  into two",
+        "bbox": [
+            89,
+            839,
+            483,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "sets of features, namely 'slow' features  $\\mathbf{F}_v^s\\in \\mathbb{R}^{T_s\\times H\\times W\\times C}$  and fast features'  $\\mathbf{F}_v^f\\in \\mathbb{R}^{T\\times H_f\\times W_f\\times C}$ . Fast features are generated by aggressively downsampling the spatial dimensions in  $\\mathbf{F}_v$  using adaptive average pooling, i.e.  $H_{f} <   <   H$  and  $W_{f} <   <   W$ . Meanwhile, slow features are obtained by uniformly sampling  $T_{s} <   T$  frame features from  $\\mathbf{F}_v$ . These two sets of features are then flattened and concatenated to yield the set of 'slow-fast' video features  $\\mathbf{F}_{sf}\\in \\mathbb{R}^{N_v\\times C}$  where  $N_{v} = (T\\times H_{f}\\times W_{f}) + (T_{s}\\times H\\times W)$ . The idea is to retain finegrained information along both temporal and spatial dimensions: the slow features are temporally dense but spatially condensed, whereas the fast features are spatially dense but temporally sparse. This reduces the number of video tokens  $N_{v}$  input to the multimodal LLM.",
+        "bbox": [
+            511,
+            340,
+            906,
+            555
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Multi-modal LLM. The slow-fast video features are then applied to a projection MLP, concatenated with the text prompt embeddings, and then input to the LLM, which outputs a set of embeddings  $\\mathbf{F}_o$ . For the Video Captioning task,  $\\mathbf{F}_o$  is decoded into text. For LG-VIS, the model is trained to output a special <SEG> token which is used by the segmentation network to regress the mask. Note that when the text prompt references multiple objects, the model generates a separate <SEG> token for each target object.",
+        "bbox": [
+            511,
+            564,
+            908,
+            700
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Segmentation Network. The video frames are encoded at high resolution by a segmentation backbone to yield multi-scale features which, together with <SEG> tokens, are input to a mask decoder to obtain the final segmentation masks. The mask decoder architecture is borrowed from SAM2 [86] and consists of multiple layers of bi-directional cross-attention between the <SEG> tokens and the segmentation backbone features. Finally, the dot-product between the two is computed to obtain the mask logits. We refer to Ravi et al. [86] for more details. Our approach for mask prediction using <SEG> tokens generated by an MLLM is inspired from recent image segmentation [53, 85, 120, 126] and video segmentation works [10, 117].",
+        "bbox": [
+            511,
+            704,
+            908,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "19029",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Comparison to Existing Methods. Video-LLaVA-Seg is a compact baseline model that is easy to setup and extend, unlike existing LLM-based video segmentation approaches [10, 117]: VISA [117] involves inference with a pretrained Llama-Vid [62] to select key frames for their own model, and both VISA and VideoLISA [10] rely on pretrained models [13, 21] to perform/improve temporal mask propagation. Moreover, these approaches tackle only segmentation tasks whereas our model is trained to tackle both captioning and segmentation.",
+        "bbox": [
+            89,
+            90,
+            480,
+            242
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2. Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            89,
+            250,
+            307,
+            265
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We employ an AM-RADIO [84] ViT-H [50] model as the vision backbone and Llama3-8B [32] as the LLM. The segmentation backbone is a Hiera-Small [88] network which, together with the mask decoder, is initialized with weights from SAM2 [86]. We sample  $T = 32$  frames and use  $T_{s} = 8$  slow frames. We train our model in three stages: (1) a vision-language alignment stage to optimize the projection MLP, (2) a finetuning step to optimize the MLLM, projection MLP and vision backbone for video captioning, and finally (3) the entire model with the segmentation network is fintuned for both tasks. For the first two stages, we utilize a total of 3.5M video caption samples from WebVid10M [11] and Panda70M [20]. In stage 3, we finetune the model on the ViCaS training set for both tasks. Training takes a total of  $\\sim 4$  days on 32 A100 GPUs. Further implementation details can be found in supplementary.",
+        "bbox": [
+            89,
+            273,
+            482,
+            515
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3. Ablation Studies",
+        "text_level": 1,
+        "bbox": [
+            89,
+            523,
+            254,
+            537
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We report ablations in Table 3 and discuss them below.",
+        "bbox": [
+            89,
+            546,
+            452,
+            561
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/c7e9bb15cfbe7ad73f944c08be9fb3e4fff1b4b583ace92ca3bb718e91ac9cb3.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>#</td><td>Vis Enc.</td><td>Ts</td><td>SF-Pool.</td><td>Cap.</td><td>Seg.</td><td>CA</td><td>mAP</td></tr><tr><td>1</td><td>AM-RAD.</td><td>1</td><td>X</td><td>✓</td><td>✓</td><td>2.1</td><td>15.9</td></tr><tr><td>2</td><td>AM-RAD.</td><td>2</td><td>X</td><td>✓</td><td>✓</td><td>2.4</td><td>17.0</td></tr><tr><td>3</td><td>AM-RAD.</td><td>4</td><td>X</td><td>✓</td><td>✓</td><td>2.8</td><td>18.8</td></tr><tr><td>4</td><td>AM-RAD.</td><td>8</td><td>✓</td><td>✓</td><td>X</td><td>3.0</td><td>-</td></tr><tr><td>5</td><td>AM-RAD.</td><td>8</td><td>✓</td><td>X</td><td>✓</td><td>-</td><td>18.5</td></tr><tr><td>6</td><td>CLIP</td><td>8</td><td>X</td><td>✓</td><td>✓</td><td>2.9</td><td>19.9</td></tr><tr><td></td><td>AM-RAD.</td><td>8</td><td>✓</td><td>✓</td><td>✓</td><td>3.0</td><td>20.5</td></tr></table>",
+        "bbox": [
+            91,
+            571,
+            483,
+            708
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Input Frames. In rows 1-3, we train our model using  $T_{s} = \\{1,2,4\\}$  frames instead of  $T_{s} = 8$  in the final setting. We see that using fewer frames strongly reduces performance, especially for captioning, showing that our dataset/benchmark is video-centric and requires temporal context to effectively tackle.",
+        "bbox": [
+            89,
+            758,
+            482,
+            851
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Task Synergy. For rows 4 and 5, we train the model for either one of our benchmark tasks. We see that the model trained only for LG-VIS achieves  $18.5\\mathrm{mAP}$  which is",
+        "bbox": [
+            89,
+            854,
+            483,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "much worse than the 20.5 achieved by the multi-task model. Meanwhile, the Caption Accuracy remains unchanged at 3.0. This indicates that pixel-level segmentation benefits greatly from holistic video understanding.",
+        "bbox": [
+            511,
+            90,
+            903,
+            152
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "LLaVA Baseline. Row 6 shows results for a LLaVA-NeXT baseline trained with our recipe which uses a CLIP vision encoder and no Slow-Fast Pooling. This achieves  $19.9\\mathrm{mAP}$  and  $2.9\\mathrm{CA}$ , both of which are worse than our final setting.",
+        "bbox": [
+            511,
+            156,
+            906,
+            218
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4. Benchmark Results",
+        "text_level": 1,
+        "bbox": [
+            513,
+            224,
+            700,
+            239
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/4c0a9c9bb4364a047e8aad9e6dfb1c55508ef60fe654f2b8815b541f889a38e9.jpg",
+        "table_caption": [
+            "Table 3. Ablation experiments on the ViCaS validation set. CA: Caption Accuracy. SF-Pool: Slow-Fast Pooling."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>CA</td><td>mAP</td><td>\\( AP_{50} \\)</td><td>\\( AP_{75} \\)</td><td>\\( AP_{90} \\)</td></tr><tr><td>LLaVA-OV [54] (ZS)</td><td>2.9</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MiniCPM-o 2.6 (ZS)</td><td>3.0</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>LMPM [29]</td><td>-</td><td>8.4</td><td>17.1</td><td>7.7</td><td>1.8</td></tr><tr><td>DsHmp [41]</td><td>-</td><td>13.3</td><td>27.6</td><td>11.6</td><td>2.9</td></tr><tr><td>VideoLISA [10]</td><td>-</td><td>10.7</td><td>23.8</td><td>8.5</td><td>1.8</td></tr><tr><td>Video-LLaVA-Seg</td><td>3.0</td><td>20.5</td><td>38.6</td><td>19.0</td><td>5.1</td></tr></table>",
+        "bbox": [
+            514,
+            248,
+            903,
+            354
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Benchmark Results on our validation set. Refer to supplementary for test set results. CA: Caption Accuracy",
+        "bbox": [
+            511,
+            361,
+            903,
+            390
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4 compares Video-LLaVA-Seg to other methods on the validation set. For video captioning, we evaluate off-the-shelf LLaVA-OneVision [54] and MiniCPM-o[121] models which achieve 2.9 and 3.0 CA scores, respectively. This is the same as Video-LLaVA-Seg, but these models are trained on significantly more data and cannot tackle LG-VIS. For LG-VIS, we evaluate multiple existing Referral-VOS approaches. LMPM is an earlier transformer-based approach which only achieves  $8.4\\mathrm{mAP}$ . DsHmp [41] performs better with  $13.3\\mathrm{mAP}$  since the architecture decouples static and motion cues to improve video segmentation performance. Finally, VideoLISA [10] is a recent LLM-based approach which only achieves  $10.7\\mathrm{mAP}$  when finetuned on ViCaS. By contrast, Video-LLaVA-Seg achieves  $20.5\\mathrm{mAP}$  while tackling both captioning and LG-VIS. Furthermore, unlike existing Referral-VOS approaches, it can predict multiple segmentation masks for a single prompt.",
+        "bbox": [
+            511,
+            398,
+            906,
+            656
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            511,
+            662,
+            633,
+            678
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We introduce ViCaS, a first-of-its-kind, human-annotated dataset that provides detailed captions for videos along with phrase-grounded segmentation masks for salient objects. Our associated benchmark comprises two tasks: (1) Video Captioning, which evaluates high-level, holistic understanding, and (2) Language-Guided Video Instance Segmentation (LG-VIS), which evaluates pixel-level understanding. We propose evaluation measures for open-ended caption accuracy which are based on open-source models, and are experimentally verified through a comprehensive user study. Furthermore, we propose Video-LLaVA-Seg, a compact baseline which effectively tackles both tasks.",
+        "bbox": [
+            511,
+            683,
+            906,
+            866
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements. We thank the participants of the user study and all the annotation personnel for their effort.",
+        "bbox": [
+            511,
+            869,
+            903,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "19030",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            91,
+            89,
+            187,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. Arxiv, 2023. 5, 6",
+            "[2] Mykhaylo Andriluka, Stefan Roth, and Bernt Schiele. People-tracking-by-detection and people-detection-by-tracking. In CVPR, 2008. 2",
+            "[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 5",
+            "[4] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 2",
+            "[5] Ali Athar, Jonathon Luiten, Alexander Hermans, Deva Ramanan, and Bastian Leibe. Hodor: High-level object descriptors for object re-segmentation in video learned from static images. In CVPR, 2022. 2",
+            "[6] Ali Athar, Alexander Hermans, Jonathon Luiten, Deva Ramanan, and Bastian Leibe. Tarvis: A unified architecture for target-based video segmentation. In CVPR, 2023. 2",
+            "[7] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In WACV, 2023. 2, 4, 5",
+            "[8] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An open-source framework for training large autoregressive vision-language models. Arxiv, 2023. 1",
+            "[9] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. Arxiv, 2023. 1, 2",
+            "[10] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. NeurIPS, 2024. 3, 7, 8",
+            "[11] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2, 5, 8",
+            "[12] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL Workshop, 2005. 6",
+            "[13] Maksym Bekuzarov, Ariana Bermudez, Joon-Young Lee, and Hao Li. Xmem++: Production-level video segmentation from few annotated frames. *Arxiv*, 2023. 3, 8",
+            "[14] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. Arxiv, 2024. 2",
+            "[15] Steven Bird, Ewan Klein, and Edward Loper. Natural language processing with Python: analyzing text with the natural language toolkit. \"O'Reilly Media, Inc.\", 2009. 5"
+        ],
+        "bbox": [
+            99,
+            114,
+            483,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[16] Guillem Brasó and Laura Leal-Taixe. Learning a neural solver for multiple object tracking. In CVPR, 2020. 2",
+            "[17] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, 2011. 2, 5",
+            "[18] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In CVPR, 2024. 1, 2",
+            "[19] Jieneng Chen, Qihang Yu, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. Vitamin: Designing scalable vision models in the vision-language era. In CVPR, 2024. 2",
+            "[20] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In CVPR, 2024. 2, 5, 8",
+            "[21] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 2, 8",
+            "[22] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. NeurIPS, 2021.",
+            "[23] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Joon-Young Lee, and Alexander Schwing. Putting the object back into video object segmentation. In CVPR, 2024. 2",
+            "[24] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with  $90\\%$  * chatgpt quality, 2023. 2",
+            "[25] Daniel Cores, Michael Dorkenwald, Manuel Mucientes, Cees GM Snoek, and Yuki M Asano. Tvbench: Redesigning video-language evaluation. Arxiv, 2024. 2",
+            "[26] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In ECCV, 2020. 2",
+            "[27] Patrick Dendorfer, Aljosa Osep, Anton Milan, Konrad Schindler, Daniel Cremers, Ian Reid, Stefan Roth, and Laura Leal-Taixe. Motchallenge: A benchmark for single-camera multiple target tracking. IJCV, 2021. 2",
+            "[28] Xueqing Deng, Qihang Yu, Peng Wang, Xiaohui Shen, and Liang-Chieh Chen. Coconut: Modernizing coco segmentation. In CVPR, 2024. 2",
+            "[29] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In ICCV, 2023. 1, 2, 4, 5, 8",
+            "[30] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. Mose: A new dataset for video object segmentation in complex scenes. In ICCV, 2023. 2",
+            "[31] Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and description. In CVPR, 2015. 2"
+        ],
+        "bbox": [
+            522,
+            92,
+            905,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "19031",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[32] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. Arxiv, 2024. 1, 2, 6, 8",
+            "[33] Dave Epstein, Boyuan Chen, and Carl Vondrick. Oops! predicting unintentional action in video. In CVPR, 2020. 3",
+            "[34] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 2",
+            "[35] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In CVPR, 2018. 5",
+            "[36] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2",
+            "[37] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 2",
+            "[38] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 2",
+            "[39] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In CVPR, 2018. 2",
+            "[40] Ju He, Qihang Yu, Inkyu Shin, Xueqing Deng, Alan Yuille, Xiaohui Shen, and Liang-Chieh Chen. Maxtron: Mask transformer with trajectory attention for video panoptic segmentation. Arxiv, 2023. 2",
+            "[41] Shuting He and Henghui Ding. Decoupling static and hierarchical motion perception for referring video segmentation. In CVPR, 2024. 2, 8",
+            "[42] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. NeurIPS, 2022. 2",
+            "[43] Namdar Homayounfar, Justin Liang, Wei-Chiu Ma, and Raquel Urtasun. Videoclick: Video object segmentation with a single click. Arxiv, 2021. 2",
+            "[44] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE TPAMI, 2021. 2",
+            "[45] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014. 2",
+            "[46] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. Arxiv, 2017. 2, 3",
+            "[47] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In ECCV, 2019. 1, 2, 4, 5"
+        ],
+        "bbox": [
+            99,
+            92,
+            482,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[48] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Video panoptic segmentation. In CVPR, 2020. 2",
+            "[49] Dahun Kim, Jun Xie, Huiyu Wang, Siyuan Qiao, Qihang Yu, Hong-Seok Kim, Hartwig Adam, In So Kweon, and Liang-Chieh Chen. Tubeformer-deeplab: Video mask transformer. In CVPR, 2022. 2",
+            "[50] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 8",
+            "[51] Matej Kristan, Jiri Matas, Aleš Leonardis, Tomas Vojir, Roman Pflugfelder, Gustavo Fernandez, Georg Nebehay, Fatih Porikli, and Luka Čehovin. A novel performance evaluation methodology for single-target trackers. IEEE TPAMI, 2016. 2",
+            "[52] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 2",
+            "[53] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In CVPR, 2024. 2, 7",
+            "[54] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. Arxiv, 2024. 8",
+            "[55] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 1",
+            "[56] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1",
+            "[57] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In CVPR, 2024. 2",
+            "[58] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. Arxiv, 2019. 1",
+            "[59] Siyuan Li, Martin Danelljan, Henghui Ding, Thomas E. Huang, and Fisher Yu. Tracking every thing in the wild. In ECCV, 2022. 7",
+            "[60] Xianming Li and Jing Li. Angle-optimized text embeddings. Arxiv, 2023. 6",
+            "[61] Yuncheng Li, Yale Song, Liangliang Cao, Joel Tetreault, Larry Goldberg, Alejandro Jaimes, and Jiebo Luo. Tgif: A new dataset and benchmark on animated gif description. In CVPR, 2016. 2",
+            "[62] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In ECCV, 2024. 1, 2, 3, 8",
+            "[63] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. In EMNLP, 2024. 1, 2"
+        ],
+        "bbox": [
+            524,
+            92,
+            903,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "19032",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[64] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In ACL Workshop, 2004. 6",
+            "[65] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. *Arxiv*, 2023. 7",
+            "[66] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 1, 2",
+            "[67] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. 1, 2, 7",
+            "[68] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, 2022. 2",
+            "[69] Jonathon Luiten, Paul Voigtlaender, and Bastian Leibe. Premvos: Proposal-generation, refinement and merging for video object segmentation. In ACCV, 2018. 2",
+            "[70] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. IJCV, 2020. 7",
+            "[71] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. Arxiv, 2020. 2",
+            "[72] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. Arxiv, 2023. 1, 2",
+            "[73] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 1, 2, 6",
+            "[74] Jieru Mei, Alex Zihao Zhu, Xinchen Yan, Hang Yan, Siyuan Qiao, Liang-Chieh Chen, and Henrik Kretzschmar. Waymo open dataset: Panoramic video panoptic segmentation. In ECCV, 2022. 2",
+            "[75] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In CVPR, 2022. 2",
+            "[76] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In ICCV, 2019. 2",
+            "[77] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 6",
+            "[78] Nikos Paragios and Rachid Deriche. Geodesic active contours and level sets for the detection and tracking of moving objects. IEEE TPAMI, 2000. 2",
+            "[79] AJ Piergiovanni, Weicheng Kuo, and Anelia Angelova. Rethinking video vits: Sparse video tubes for joint image and video learning. In CVPR, 2023. 2",
+            "[80] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. *Arxiv*, 2017. 2"
+        ],
+        "bbox": [
+            99,
+            92,
+            482,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[81] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. IJCV, 2022. 2, 5, 7",
+            "[82] Siyuan Qiao, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Vip-deeplab: Learning visual perception with depth-aware video panoptic segmentation. In CVPR, 2021. 2",
+            "[83] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1",
+            "[84] Mike Ranzinger, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Am-radio: Agglomerative vision foundation model reduce all domains into one. In CVPR, 2024. 8",
+            "[85] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In CVPR, 2024. 7",
+            "[86] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. *Arxiv*, 2024. 2, 4, 7, 8",
+            "[87] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In EMNLP, 2019. 6",
+            "[88] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In ICML, 2023. 8",
+            "[89] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, 2020. 1, 2, 4, 5",
+            "[90] Inkyu Shin, Dahun Kim, Qihang Yu, Jun Xie, Hong-Seok Kim, Bradley Green, In So Kweon, Kuk-Jin Yoon, and Liang-Chieh Chen. Video-kmax: A simple unified approach for online and near-online video panoptic segmentation. In WACV, 2024. 2",
+            "[91] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, 2016. 2",
+            "[92] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In CVPR, 2024. 1, 2",
+            "[93] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. Arxiv, 2012. 2",
+            "[94] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, 2019. 2"
+        ],
+        "bbox": [
+            522,
+            92,
+            903,
+            900
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "19033",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[95] Siyu Tang, Mykhaylo Andriluka, Bjoern Andres, and Bernt Schiele. Multiple people tracking by lifted multicut and person re-identification. In CVPR, 2017. 2",
+            "[96] Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Riviere, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. Arxiv, 2024. 1, 2",
+            "[97] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 2",
+            "[98] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018.",
+            "[99] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, 2019. 2",
+            "[100] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 2",
+            "[101] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, 2015. 6",
+            "[102] Paul Voigtlaender, Michael Krause, Aljosa Osep, Jonathon Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger, and Bastian Leibe. Mots: Multi-object tracking and segmentation. In CVPR, 2019. 2",
+            "[103] Paul Voigtlaender, Soravit Changpinyo, Jordi Pont-Tuset, Radu Soricut, and Vittorio Ferrari. Connecting vision and language with video localized narratives. In CVPR, 2023. 2",
+            "[104] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In ICCV, 2023. 2, 5",
+            "[105] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, open-world segmentation. In ICCV, 2021. 2, 3, 4, 5, 7",
+            "[106] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In ICCV, 2019. 5",
+            "[107] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2",
+            "[108] Mark Weber, Jun Xie, Maxwell Collins, Yukun Zhu, Paul Voigtlaender, Hartwig Adam, Bradley Green, Andreas Geiger, Bastian Leibe, Daniel Cremers, Aljosa Osep, Laura Leal-Taixe, and Liang-Chieh Chen. Step: Segmenting and tracking every pixel. NeurIPS, 2021. 2",
+            "[109] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 2",
+            "[110] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answer"
+        ],
+        "bbox": [
+            93,
+            92,
+            480,
+            900
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "ing via gradually refined attention over appearance and motion. In ACM MM, 2017. 2, 5",
+            "[111] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 2, 5",
+            "[112] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In ICML, 2015. 2",
+            "[113] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. *Arxiv*, 2024. 1",
+            "[114] Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. Arxiv, 2024. 7",
+            "[115] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. Arxiv, 2018. 2",
+            "[116] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 2",
+            "[117] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. In ECCV, 2024. 1, 3, 5, 7, 8",
+            "[118] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, 2023. 1",
+            "[119] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 5, 7",
+            "[120] Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. An improved baseline for reasoning segmentation with large language model. *Arxiv*, 2023. 7",
+            "[121] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. Arxiv, 2024. 8",
+            "[122] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2",
+            "[123] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. NeurIPS, 2023. 2",
+            "[124] Qihang Yu, Xiaohui Shen, and Liang-Chieh Chen. Towards open-ended visual recognition with large language models. In ECCV, 2024. 2",
+            "[125] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            900
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "19034",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "understanding complex web videos via question answering. In AAAI, 2019. 2, 5",
+            "[126] Yuqian Yuan, Wentong Li, Jian Liu, Dongqi Tang, Xinjie Luo, Chi Qin, Lei Zhang, and Jianke Zhu. Osprey: Pixel understanding with visual instruction tuning. In CVPR, 2024. 7",
+            "[127] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. Arxiv, 2023. 1",
+            "[128] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In CVPR, 2022. 1",
+            "[129] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In ICLR, 2020. 6",
+            "[130] Tao Zhang, Xingye Tian, Yikang Zhou, Shunping Ji, Xuebo Wang, Xin Tao, Yuan Zhang, Pengfei Wan, Zhongyuan Wang, and Yu Wu. Dvis++: Improved decoupled framework for universal video segmentation. Arxiv, 2023. 2",
+            "[131] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 5",
+            "[132] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In CVPR, 2024. 1",
+            "[133] Idil Esen Zulfikar, Sabarinath Mahadevan, Paul Voigtlaender, and Bastian Leibe. Point-vos: Pointing up video object segmentation. In CVPR, 2024. 2"
+        ],
+        "bbox": [
+            91,
+            92,
+            482,
+            517
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "19035",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 12
+    }
+]