Add Batch 48ba8781-9f3e-400b-b71e-44fbd479eca8 data

.gitattributes

@@ -2343,3 +2343,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/Towards[[:space:]]Transformer-Based[[:space:]]Aligned[[:space:]]Generation[[:space:]]with[[:space:]]Self-Coherence[[:space:]]Guidance/0430b3b2-c194-4129-b2be-acaa49b5720f_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/Towards[[:space:]]Unbiased[[:space:]]and[[:space:]]Robust[[:space:]]Spatio-Temporal[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation[[:space:]]and[[:space:]]Anticipation/61a4ff36-4fd8-43f1-a4cb-46547e183676_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/Towards[[:space:]]Understanding[[:space:]]How[[:space:]]Knowledge[[:space:]]Evolves[[:space:]]in[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models/feb88a82-d0b8-411b-8ce3-184786042b8f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Towards[[:space:]]Understanding[[:space:]]and[[:space:]]Quantifying[[:space:]]Uncertainty[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Generation/c3ad3744-8ee2-4117-a4a1-7d23c58e3fe4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Towards[[:space:]]Universal[[:space:]]AI-Generated[[:space:]]Image[[:space:]]Detection[[:space:]]by[[:space:]]Variational[[:space:]]Information[[:space:]]Bottleneck[[:space:]]Network/c23bce64-f04c-4e24-b700-3d6473e34043_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Towards[[:space:]]Universal[[:space:]]Dataset[[:space:]]Distillation[[:space:]]via[[:space:]]Task-Driven[[:space:]]Diffusion/cd8b4c9e-1814-423f-bdad-2b631dc6a75a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Towards[[:space:]]Universal[[:space:]]Soccer[[:space:]]Video[[:space:]]Understanding/bd5f526c-bc30-45cc-bbcd-e151dfc21f1a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Towards[[:space:]]Visual[[:space:]]Discrimination[[:space:]]and[[:space:]]Reasoning[[:space:]]of[[:space:]]Real-World[[:space:]]Physical[[:space:]]Dynamics_[[:space:]]Physics-Grounded[[:space:]]Anomaly[[:space:]]Detection/e496d86c-0d1a-42d1-b098-7f552a70c4ff_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Towards[[:space:]]Zero-Shot[[:space:]]Anomaly[[:space:]]Detection[[:space:]]and[[:space:]]Reasoning[[:space:]]with[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models/9698d3f1-26f6-4c2e-aef0-60a71f074ba9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Towards[[:space:]]a[[:space:]]Universal[[:space:]]Synthetic[[:space:]]Video[[:space:]]Detector_[[:space:]]From[[:space:]]Face[[:space:]]or[[:space:]]Background[[:space:]]Manipulations[[:space:]]to[[:space:]]Fully[[:space:]]AI-Generated[[:space:]]Content/49360ca2-47e3-4c7b-81b1-6a87621996c4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Tra-MoE_[[:space:]]Learning[[:space:]]Trajectory[[:space:]]Prediction[[:space:]]Model[[:space:]]from[[:space:]]Multiple[[:space:]]Domains[[:space:]]for[[:space:]]Adaptive[[:space:]]Policy[[:space:]]Conditioning/d41e3854-0a29-4aee-85fc-0fcf3b667c89_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TraF-Align_[[:space:]]Trajectory-aware[[:space:]]Feature[[:space:]]Alignment[[:space:]]for[[:space:]]Asynchronous[[:space:]]Multi-agent[[:space:]]Perception/0f12efc0-153e-4cda-95dd-b16491feb48e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Track[[:space:]]Any[[:space:]]Anomalous[[:space:]]Object_A[[:space:]]Granular[[:space:]]Video[[:space:]]Anomaly[[:space:]]Detection[[:space:]]Pipeline/1ddd1714-9e44-4427-b96d-74620c052e99_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Track4Gen_[[:space:]]Teaching[[:space:]]Video[[:space:]]Diffusion[[:space:]]Models[[:space:]]to[[:space:]]Track[[:space:]]Points[[:space:]]Improves[[:space:]]Video[[:space:]]Generation/52b14d4e-8da8-4b5e-a32c-278d64dd36e6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Tracktention_[[:space:]]Leveraging[[:space:]]Point[[:space:]]Tracking[[:space:]]to[[:space:]]Attend[[:space:]]Videos[[:space:]]Faster[[:space:]]and[[:space:]]Better/55c83ffe-0183-42a7-a82c-e5f5408552e5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Training[[:space:]]Data[[:space:]]Provenance[[:space:]]Verification_[[:space:]]Did[[:space:]]Your[[:space:]]Model[[:space:]]Use[[:space:]]Synthetic[[:space:]]Data[[:space:]]from[[:space:]]My[[:space:]]Generative[[:space:]]Model[[:space:]]for[[:space:]]Training_/95a43e45-ca4b-40d2-9bb6-79d7e388dedb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Training-free[[:space:]]Dense-Aligned[[:space:]]Diffusion[[:space:]]Guidance[[:space:]]for[[:space:]]Modular[[:space:]]Conditional[[:space:]]Image[[:space:]]Synthesis/237b334f-3be2-4e40-8ac1-466a51400496_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Training-free[[:space:]]Neural[[:space:]]Architecture[[:space:]]Search[[:space:]]through[[:space:]]Variance[[:space:]]of[[:space:]]Knowledge[[:space:]]of[[:space:]]Deep[[:space:]]Network[[:space:]]Weights/740054ef-86e4-4bac-b7b3-53d57b1f3a66_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Trajectory[[:space:]]Mamba_[[:space:]]Efficient[[:space:]]Attention-Mamba[[:space:]]Forecasting[[:space:]]Model[[:space:]]Based[[:space:]]on[[:space:]]Selective[[:space:]]SSM/14500b01-a80f-4244-a6f9-1c844f2302da_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TransPixeler_[[:space:]]Advancing[[:space:]]Text-to-Video[[:space:]]Generation[[:space:]]with[[:space:]]Transparency/a57aba62-4ac6-4f75-a668-a4f223b5c05f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Transfer[[:space:]]Your[[:space:]]Perspective_[[:space:]]Controllable[[:space:]]3D[[:space:]]Generation[[:space:]]from[[:space:]]Any[[:space:]]Viewpoint[[:space:]]in[[:space:]]a[[:space:]]Driving[[:space:]]Scene/0e711492-3ab2-4748-8147-c897ce7d76b8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Transformers[[:space:]]without[[:space:]]Normalization/05ce47c0-7126-423f-aeff-039d314c6fa4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Traversing[[:space:]]Distortion-Perception[[:space:]]Tradeoff[[:space:]]using[[:space:]]a[[:space:]]Single[[:space:]]Score-Based[[:space:]]Generative[[:space:]]Model/dcdd686b-eeb1-4343-a7e6-20b5cd329198_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TreeMeshGPT_[[:space:]]Artistic[[:space:]]Mesh[[:space:]]Generation[[:space:]]with[[:space:]]Autoregressive[[:space:]]Tree[[:space:]]Sequencing/cc43c3b2-b6bf-472a-9dd6-c74d4f03cee9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TriTex_[[:space:]]Learning[[:space:]]Texture[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Mesh[[:space:]]via[[:space:]]Triplane[[:space:]]Semantic[[:space:]]Features/f0b8cf58-1685-4de9-9df0-451922c97e91_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Tripartite[[:space:]]Weight-Space[[:space:]]Ensemble[[:space:]]for[[:space:]]Few-Shot[[:space:]]Class-Incremental[[:space:]]Learning/6894bed0-33d2-4729-8056-07cf7489b89f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Tuning[[:space:]]the[[:space:]]Frequencies_[[:space:]]Robust[[:space:]]Training[[:space:]]for[[:space:]]Sinusoidal[[:space:]]Neural[[:space:]]Networks/79e8b6d6-50a7-4215-bd2a-fb5524e17539_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Turbo3D_[[:space:]]Ultra-fast[[:space:]]Text-to-3D[[:space:]]Generation/bd4523a6-3f35-4dc6-9b48-90da3f36ebea_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TurboFill_[[:space:]]Adapting[[:space:]]Few-step[[:space:]]Text-to-image[[:space:]]Model[[:space:]]for[[:space:]]Fast[[:space:]]Image[[:space:]]Inpainting/6302d669-b0e3-4a35-adf2-a42f09f692e9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Twinner_[[:space:]]Shining[[:space:]]Light[[:space:]]on[[:space:]]Digital[[:space:]]Twins[[:space:]]in[[:space:]]a[[:space:]]Few[[:space:]]Snaps/0293a039-5500-41cc-90fa-fe4f8710c6e9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Two[[:space:]]by[[:space:]]Two_[[:space:]]Learning[[:space:]]Multi-Task[[:space:]]Pairwise[[:space:]]Objects[[:space:]]Assembly[[:space:]]for[[:space:]]Generalizable[[:space:]]Robot[[:space:]]Manipulation/57fa886e-a784-460f-94a7-de5f04b8ab96_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Two[[:space:]]is[[:space:]]Better[[:space:]]than[[:space:]]One_[[:space:]]Efficient[[:space:]]Ensemble[[:space:]]Defense[[:space:]]for[[:space:]]Robust[[:space:]]and[[:space:]]Compact[[:space:]]Models/f8d31d9f-8dff-45eb-9d02-43167ab5a77a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Type-R_[[:space:]]Automatically[[:space:]]Retouching[[:space:]]Typos[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Generation/6f176b19-a787-4b85-9ebe-081a3f4759be_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/U-Know-DiffPAN_[[:space:]]An[[:space:]]Uncertainty-aware[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]Diffusion[[:space:]]Framework[[:space:]]with[[:space:]]Details[[:space:]]Enhancement[[:space:]]for[[:space:]]PAN-Sharpening/ca20f1f2-4ee9-4b83-80ca-6e2c71410bad_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UA-Pose_[[:space:]]Uncertainty-Aware[[:space:]]6D[[:space:]]Object[[:space:]]Pose[[:space:]]Estimation[[:space:]]and[[:space:]]Online[[:space:]]Object[[:space:]]Completion[[:space:]]with[[:space:]]Partial[[:space:]]References/0d3743ff-3af5-445c-ae4c-ba5a33b9d7b4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UCM-VeID[[:space:]]V2_[[:space:]]A[[:space:]]Richer[[:space:]]Dataset[[:space:]]and[[:space:]]A[[:space:]]Pre-training[[:space:]]Method[[:space:]]for[[:space:]]UAV[[:space:]]Cross-Modality[[:space:]]Vehicle[[:space:]]Re-Identification/9b7d21db-8ee1-4718-8ddf-98653daa1c9e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UCOD-DPL_[[:space:]]Unsupervised[[:space:]]Camouflaged[[:space:]]Object[[:space:]]Detection[[:space:]]via[[:space:]]Dynamic[[:space:]]Pseudo-label[[:space:]]Learning/dec1b26d-4dd9-4c06-a461-c03d0e31f356_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UHD-processer_[[:space:]]Unified[[:space:]]UHD[[:space:]]Image[[:space:]]Restoration[[:space:]]with[[:space:]]Progressive[[:space:]]Frequency[[:space:]]Learning[[:space:]]and[[:space:]]Degradation-aware[[:space:]]Prompts/2e2768b6-2887-4bd5-935c-d0b880ebe319_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UIBDiffusion_[[:space:]]Universal[[:space:]]Imperceptible[[:space:]]Backdoor[[:space:]]Attack[[:space:]]for[[:space:]]Diffusion[[:space:]]Models/7f0574b2-e282-4c3e-a072-3d64d026983a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UMFN_[[:space:]]Unified[[:space:]]Multi-Domain[[:space:]]Face[[:space:]]Normalization[[:space:]]for[[:space:]]Joint[[:space:]]Cross-domain[[:space:]]Prototype[[:space:]]Learning[[:space:]]and[[:space:]]Heterogeneous[[:space:]]Face[[:space:]]Recognition/3e3f9780-5e1d-4d68-aee1-68dcb88c5368_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UMotion_[[:space:]]Uncertainty-driven[[:space:]]Human[[:space:]]Motion[[:space:]]Estimation[[:space:]]from[[:space:]]Inertial[[:space:]]and[[:space:]]Ultra-wideband[[:space:]]Units/20c112a8-f19a-4ffc-8e5b-a5d02b15caa1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UNEM_[[:space:]]UNrolled[[:space:]]Generalized[[:space:]]EM[[:space:]]for[[:space:]]Transductive[[:space:]]Few-Shot[[:space:]]Learning/592279b8-3f9e-4b39-91e5-638d8677cb8a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UNIALIGN_[[:space:]]Scaling[[:space:]]Multimodal[[:space:]]Alignment[[:space:]]within[[:space:]]One[[:space:]]Unified[[:space:]]Model/7e109366-fc0f-438c-9b91-2ba7f07e59e9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UNIC-Adapter_[[:space:]]Unified[[:space:]]Image-instruction[[:space:]]Adapter[[:space:]]with[[:space:]]Multi-modal[[:space:]]Transformer[[:space:]]for[[:space:]]Image[[:space:]]Generation/409eb6c5-d571-4968-a329-2c1ebf51e321_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UNICL-SAM_[[:space:]]Uncertainty-Driven[[:space:]]In-Context[[:space:]]Segmentation[[:space:]]with[[:space:]]Part[[:space:]]Prototype[[:space:]]Discovery/fc453b95-4cba-4bb9-95fb-6c8d34e4ccaa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UNOPose_[[:space:]]Unseen[[:space:]]Object[[:space:]]Pose[[:space:]]Estimation[[:space:]]with[[:space:]]an[[:space:]]Unposed[[:space:]]RGB-D[[:space:]]Reference[[:space:]]Image/94d3840c-f3fe-4b2b-9388-48ccee10c9d7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UPME_[[:space:]]An[[:space:]]Unsupervised[[:space:]]Peer[[:space:]]Review[[:space:]]Framework[[:space:]]for[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Evaluation/42fa34eb-b512-4300-b1c0-40e45c00118f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/URWKV_[[:space:]]Unified[[:space:]]RWKV[[:space:]]Model[[:space:]]with[[:space:]]Multi-state[[:space:]]Perspective[[:space:]]for[[:space:]]Low-light[[:space:]]Image[[:space:]]Restoration/e077664a-8838-4206-9246-0d60f57e032f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/USP-Gaussian_[[:space:]]Unifying[[:space:]]Spike-based[[:space:]]Image[[:space:]]Reconstruction,[[:space:]]Pose[[:space:]]Correction[[:space:]]and[[:space:]]Gaussian[[:space:]]Splatting/62c2f0f5-8ea0-4959-8e97-f59984bffdfe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UVGS_[[:space:]]Reimagining[[:space:]]Unstructured[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]using[[:space:]]UV[[:space:]]Mapping/77087a0b-d1cb-4230-a846-7760f7510d17_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UWAV_[[:space:]]Uncertainty-weighted[[:space:]]Weakly-supervised[[:space:]]Audio-Visual[[:space:]]Video[[:space:]]Parsing/82f7769f-b767-4619-96ea-539823053afb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UltraFusion_[[:space:]]Ultra[[:space:]]High[[:space:]]Dynamic[[:space:]]Imaging[[:space:]]using[[:space:]]Exposure[[:space:]]Fusion/c2c26ec1-c8af-49f8-87ba-b3c749531067_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UnCommon[[:space:]]Objects[[:space:]]in[[:space:]]3D/4d2bd976-41fd-45e5-bfc6-8a4578f73725_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unbiased[[:space:]]Video[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation[[:space:]]via[[:space:]]Visual[[:space:]]and[[:space:]]Semantic[[:space:]]Dual[[:space:]]Debiasing/168f707b-a7fa-4e7c-93c1-aac7eef770eb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unbiasing[[:space:]]through[[:space:]]Textual[[:space:]]Descriptions_[[:space:]]Mitigating[[:space:]]Representation[[:space:]]Bias[[:space:]]in[[:space:]]Video[[:space:]]Benchmarks/5edfcfce-ad23-41c6-9a43-d3cb9f8bcb90_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unboxed_[[:space:]]Geometrically[[:space:]]and[[:space:]]Temporally[[:space:]]Consistent[[:space:]]Video[[:space:]]Outpainting/3a973275-71b1-47b0-9c32-b162722059fe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Uncertain[[:space:]]Multimodal[[:space:]]Intention[[:space:]]and[[:space:]]Emotion[[:space:]]Understanding[[:space:]]in[[:space:]]the[[:space:]]Wild/50195041-da23-4b5c-8721-3858d2b20dc9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Uncertainty[[:space:]]Meets[[:space:]]Diversity_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Active[[:space:]]Learning[[:space:]]Framework[[:space:]]for[[:space:]]Indoor[[:space:]]3D[[:space:]]Object[[:space:]]Detection/d2e21437-1e7e-4dc1-839a-f40a2bf9489e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Uncertainty[[:space:]]Weighted[[:space:]]Gradients[[:space:]]for[[:space:]]Model[[:space:]]Calibration/792a362b-85b5-4092-b750-5040ac391b64_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Uncertainty-Instructed[[:space:]]Structure[[:space:]]Injection[[:space:]]for[[:space:]]Generalizable[[:space:]]HD[[:space:]]Map[[:space:]]Construction/308073a9-0b23-46de-a5fd-816ea2a49dc7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Uncertainty-guided[[:space:]]Perturbation[[:space:]]for[[:space:]]Image[[:space:]]Super-Resolution[[:space:]]Diffusion[[:space:]]Model/10430da4-a80e-4106-a1b5-f1210ade542e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Understanding[[:space:]]Fine-tuning[[:space:]]CLIP[[:space:]]for[[:space:]]Open-vocabulary[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]in[[:space:]]Hyperbolic[[:space:]]Space/4933d0fc-2175-4f82-8204-a2ef1c424606_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Understanding[[:space:]]Multi-Task[[:space:]]Activities[[:space:]]from[[:space:]]Single-Task[[:space:]]Videos/b633800a-1331-4862-80d6-5ccbeba54c12_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Understanding[[:space:]]Multi-layered[[:space:]]Transmission[[:space:]]Matrices/0c5698ec-c1a0-4efc-b1d6-15a1169745e5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Uni-Renderer_[[:space:]]Unifying[[:space:]]Rendering[[:space:]]and[[:space:]]Inverse[[:space:]]Rendering[[:space:]]Via[[:space:]]Dual[[:space:]]Stream[[:space:]]Diffusion/ef5cd413-b783-4606-bcb0-152603271134_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Uni4D_[[:space:]]Unifying[[:space:]]Visual[[:space:]]Foundation[[:space:]]Models[[:space:]]for[[:space:]]4D[[:space:]]Modeling[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Video/b0bfb21a-5579-4e48-8be6-e0a25d4d029f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UniAP_[[:space:]]Unifying[[:space:]]Inter-[[:space:]]and[[:space:]]Intra-Layer[[:space:]]Automatic[[:space:]]Parallelism[[:space:]]by[[:space:]]Mixed[[:space:]]Integer[[:space:]]Quadratic[[:space:]]Programming/b1ddd4d6-5b56-44ba-82ed-6323538fa782_origin.pdf filter=lfs diff=lfs merge=lfs -text

2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/c3ad3744-8ee2-4117-a4a1-7d23c58e3fe4_content_list.json

@@ -0,0 +1,1544 @@
+[
+    {
+        "type": "text",
+        "text": "Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation",
+        "text_level": 1,
+        "bbox": [
+            228,
+            130,
+            772,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Gianni Franchi, $^{1}$ * Nacim Belkhir, $^{2}$ * Dat Nguyen Trong, $^{1}$  Guoxuan Xia, $^{3}$  Andrea Pilzer $^{4}$   \nU2IS, ENSTA Paris, Institut Polytechnique de Paris,  $^{1}$ Mirai, $^{2}$  Imperial College London, $^{3}$  NVIDIA $^{4}$",
+        "bbox": [
+            258,
+            202,
+            738,
+            276
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/0df14e7c0fe8aff5150a47ddde96930d88820f306268a239773e70e5602c43fa.jpg",
+        "image_caption": [
+            "Figure 1. Examples of Applications for Uncertainty Quantification in Text-to-Image Generation. Text-to-image generation models may exhibit uncertainty, and that need to be quantified since it can provide insights into the model's training dataset, aiding in deepfake prevention, detecting model biases, and protecting copyrighted content from unauthorized generation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            308,
+            377,
+            597
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/e4cb42dfba865827dd7ac22e034ddfa0a402644ec557b371ebb37dce40429cc8.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            379,
+            308,
+            879,
+            597
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            660,
+            328,
+            676
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Uncertainty quantification in text-to-image (T2I) generative models is crucial for understanding model behavior and improving output reliability. In this paper, we are the first to quantify and evaluate the uncertainty of T2I models with respect to the prompt. Alongside adapting existing approaches designed to measure uncertainty in the image space, we also introduce Prompt-based Uncertainty Estimation for T2I models (PUNC), a novel method leveraging Large Vision-Language Models (LVLMs) to better address uncertainties arising from the semantics of the prompt and generated images. PUNC utilizes a LVLM to caption a gen",
+        "bbox": [
+            88,
+            694,
+            485,
+            861
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "erated image, and then compares the caption with the original prompt in the more semantically meaningful text space. PUNC also enables the disentanglement of both aleatoric and epistemic uncertainties via precision and recall, which image-space approaches are unable to do. Extensive experiments demonstrate that PUNC outperforms state-of-the-art uncertainty estimation techniques across various settings. Uncertainty quantification in text-to-image generation models can be used on various applications including bias detection, copyright protection, and OOD detection. We also introduce a comprehensive dataset of text prompts and generation pairs to foster further research in uncertainty quantification for generative models. Our findings illustrate that PUNC not only achieves competitive performance but also enables novel applications in evaluating",
+        "bbox": [
+            511,
+            662,
+            908,
+            888
+        ],
+        "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\nresponding author - gianni.franchi@ensta-paris.fr",
+        "bbox": [
+            109,
+            875,
+            370,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "+",
+        "bbox": [
+            473,
+            875,
+            482,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "8062",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "and improving the trustworthiness of text-to-image models. The code is available at https://github.com/ENSTA-U2IS-AI/Uncertainty_diffusion",
+        "bbox": [
+            89,
+            90,
+            480,
+            136
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            89,
+            165,
+            222,
+            181
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Generative models have revolutionized artificial intelligence, making significant strides in fields such as robotics [2], healthcare [28], and physics [33]. These models have elevated the performance of AI systems by enabling the generation of complex data representations. However, while generative models offer immense potential, they do not address all challenges, especially those related to robustness and uncertainty in Deep Neural Networks(DNNs)—issues that have become even more pronounced in generative settings.",
+        "bbox": [
+            89,
+            190,
+            482,
+            340
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we focus on the uncertainty quantification in deep generative models, in particular, we are the first to investigate modern text-to-image (T2I) models where the input is a natural language prompt [8, 18, 48-50]. These models have surged in popularity in recent years due to their capacity to generate high-quality images conveniently in a user-guided fashion, finding usage in a wide array of real-world applications. We are the first to investigate and propose a dedicated approach for quantifying uncertainty in text-to-image generation, aiming to fill a critical gap in generative AI research. The task of quantifying uncertainty in these models is not only complex but also crucial. Understanding the uncertainty in generative models can be invaluable for applications such as Out-of-Distribution (OOD) detection, trustworthy AI, and decision-making in high-stakes domains. To the best of our knowledge, no existing work has comprehensively addressed the challenge of uncertainty quantification in text-to-image generative models. However, given the semantic nature of the uncertainty we aim to measure, our work is closely related to recent studies on text-to-image alignment [37, 52, 67].",
+        "bbox": [
+            89,
+            342,
+            482,
+            657
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Existing research explores uncertainty estimation and OOD detection in the image/output space [4, 5, 23, 39]. For example, given a test image, can we determine whether or not it is OOD? On the other hand, the uncertainty of a generation with respect to conditioning (such as text) is underexplored. We argue that this is an important area to be addressed, as it represents a widespread real-world application of deep generative models. In this case, it remains an open question as to how should we define uncertainty, and given a definition, how should we quantify it? As illustrated in Fig. 1, focusing on uncertainty quantification in text-to-image generation could enable novel applications, allowing us to extract and interpret the knowledge embedded within models using language.",
+        "bbox": [
+            89,
+            659,
+            482,
+            869
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To answer these questions, we first argue that uncertainties should be rooted in the semantics of the image",
+        "bbox": [
+            89,
+            869,
+            482,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "and the prompt. We then adapt and evaluate a number of existing image-space approaches, as well as proposing a novel method for uncertainty quantification in T2I generative models, which we refer to as \"Prompt-based Uncertainty Estimation for T2I Generation\" (PUNC). Our approach leverages the growing capabilities of Large Vision-Language Models (LVLMs) to extract the semantics from generated images. LVLMs, which have been trained on vast amounts of text and visual data, can serve as powerful tools for interpreting the underlying meaning of text/image inputs and assessing how variations in phrasing or semantics impact the uncertainty of the generated outputs. Our hypothesis is that as LVLMs become increasingly adept at understanding complex relationships between text and images, they will allow us to better analyze and quantify uncertainty. Through this paper, we will demonstrate Prompt-based Uncertainty Estimation for T2I (PUNC) effectiveness in quantifying uncertainties in generative models. Showcasing its ability to decompose and interpret different types of uncertainties arising from the multimodal nature of the input, offering a new perspective on uncertainty quantification in generative AI.",
+        "bbox": [
+            511,
+            90,
+            903,
+            422
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To summarize, our contributions are as follows: (1) We introduce a new task aimed at quantifying uncertainty (with respect to the prompt) in text-to-image generation models, and we share a dataset of prompts to support future research on uncertainty in these models. (2) We both adapt existing image-space approaches to our tasks, as well as propose a novel straightforward method, PUNC, designed to better quantify semantic uncertainty in text-to-image generation. (3) Extensive experiments demonstrate that PUNC, despite its simplicity and computational efficiency, achieves competitive performance. (4) We highlight the practical utility of uncertainty quantification in text-to-image generation by showcasing additional applications where this approach could be beneficial.",
+        "bbox": [
+            511,
+            422,
+            903,
+            633
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            511,
+            648,
+            653,
+            664
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Deep learning models typically encounter two primary types of uncertainty: aleatoric and epistemic uncertainty [26]. Over the years, a variety of methods have been developed to quantify these uncertainties. Ensemble methods [16, 32, 34, 58, 61], for instance, have often achieved state-of-the-art performance by leveraging multiple models to assess prediction confidence. However, these methods tend to be computationally expensive due to the need to train and maintain several models. Another approach involves Bayesian methods [21], especially Bayesian Neural Networks (BNNs) [19, 27, 42, 57], which offer a theoretically grounded way to quantify uncertainty. Additionally, there are techniques aimed at directly predicting uncertainty within DNNs [10, 20, 30]. Despite their effectiveness, most of these methods are challenging to apply to text-to-image",
+        "bbox": [
+            511,
+            674,
+            903,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "8063",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/f22022d38cc577c40f1d07107d04862ce1c9139eb16d6a4dce45345b8c7c4a8d.jpg",
+        "image_caption": [
+            "Generation/Image-space Uncertainty (existing work)",
+            "Figure 2. Diagram illustrating generation/image-space uncertainty which is considered in previous work and condition/prompt-space uncertainty which is investigated in our work."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            93,
+            102,
+            480,
+            205
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/61e51754cec18673e6e57272b130161e06703729facb467446902ddcc7dc5863.jpg",
+        "image_caption": [
+            "Figure 3. Three generations from PixArt- $\\Sigma$  [7] illustrating uncertainty with regards to prompt semantics. For the Normal image, we used an ImageNet-inspired prompt; for the corrupted image, additional corruption was applied to the prompt (e.g. fish was perturbed to fis) increasing the aleatoric uncertainty; and for the out-of-distribution (OOD) case, the model was prompted to generate an image of the Prime Minister of Japan Kishida Fumio under similar conditions, where epistemic uncertainty was injected in the form of a semantic concept the model was not familiar with."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            290,
+            460,
+            386
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "generation, as they rely on ensembling predictions or accessing model uncertainty in ways that are difficult to implement for generative models without explicit access to intermediate predictions.",
+        "bbox": [
+            89,
+            559,
+            483,
+            619
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Generative models, as a subset of deep learning, also exhibit uncertainty. In the past, traditional generative models were employed for uncertainty quantification and used to help estimate the uncertainty within standard DNNs [6, 38, 56, 64]. With the recent popularity of diffusion models, research has begun exploring whether image-based diffusion models can detect out-of-distribution (OOD) samples [4, 23, 39] or quantify their own uncertainty [5, 6, 14]. Whilst these approaches align with our goals, they are limited to image-space tasks, where the input is an image rather than a textual prompt. This input difference fundamentally alters the nature of uncertainty in the model. Our approach, in contrast, specifically addresses text-to-image generation, a setting where prompt ambiguity and multimodal complexity introduce new dimensions of uncertainty. We demonstrate that quantifying uncertainty in this context can lead to valuable insights and applications across various downstream tasks.",
+        "bbox": [
+            89,
+            628,
+            483,
+            898
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Uncertainty in Image Generation Models",
+        "text_level": 1,
+        "bbox": [
+            511,
+            89,
+            880,
+            107
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Consider a deep generative model  $p_{\\theta}(\\boldsymbol{x})$  with parameters  $\\theta$  that models the data distribution  $p_{\\mathrm{data}}(\\boldsymbol{x})$  of  $\\boldsymbol{x} \\in \\mathbb{R}^D$ . We may be interested in measuring the uncertainty of our model  $p_{\\theta}$  with respect to some test datum  $\\boldsymbol{x}^*$ . For example, consider the case where we want to determine whether or not  $\\boldsymbol{x}^*$  was drawn from  $p_{\\mathrm{data}}(\\boldsymbol{x})$ , i.e. out-of-distribution (OOD) detection. This can be achieved by quantifying the uncertainty of  $p_{\\theta}$  that arises from a lack of knowledge of  $\\boldsymbol{x}^*$ , i.e. epistemic uncertainty [26]. A range of existing work that leverages diffusion-based generative models for OOD image detection [23, 39] fall into the above paradigm. From hereon we will refer to this as generation/image-space uncertainty. This is illustrated in Fig. 2 (top).",
+        "bbox": [
+            511,
+            114,
+            906,
+            311
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this work, however, we are primarily interested in a different scenario: text-to-image generation. Consider instead a conditional model  $p_{\\theta}(\\pmb{x}|\\pmb{c})$  that models the conditional data distribution  $p_{\\mathrm{data}}(\\pmb{x}|\\pmb{c})$ . This describes real-world text-to-image generation, where image  $\\pmb{x}$  is drawn from  $p_{\\theta}(\\pmb{x}|\\pmb{c}^*)$  conditioned on some test text prompt  $\\pmb{c}^*$ . In this case, we wish to quantify the uncertainty of  $p_{\\theta}$  not with respect to some test  $\\pmb{x}^*$  (e.g. image) but instead with respect to a test condition  $\\pmb{c}^*$ . We will hereon refer to this as condition/prompt-space uncertainty. This is illustrated in Fig. 2 (bottom).",
+        "bbox": [
+            511,
+            311,
+            905,
+            477
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We argue that in this case, uncertainties should be concerned with semantics, as intuitively, these are what inherently link the user's intentions, represented by the text prompt, and the generated output image. This is similar to recent discourse on uncertainty estimation for large language models [31]. If we consider the semantic concepts in the prompt and generated image, then we can define useful concepts of uncertainty (illustrated in Fig. 3):",
+        "bbox": [
+            511,
+            478,
+            905,
+            598
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Aleatoric uncertainty is irreducible uncertainty in the data distribution  $p_{\\mathrm{data}}(\\boldsymbol{x}|\\boldsymbol{c})$ . From the perspective of semantics, high aleatoric uncertainty would be where a variation of generated concepts may arise from a single prompt. For example, a spelling mistake where fish is mistyped as fis in the prompt may result in either a \"fish\" or a \"fist\" appearing in the image.",
+            "- Epistemic uncertainty should correspond to a model's lack of knowledge of semantic concepts found in a prompt. For example, a model trained on ImageNet will not know what the prime minister of Japan Kishida Fumio looks like, and thus will have high epistemic uncertainty for this semantic concept."
+        ],
+        "bbox": [
+            511,
+            598,
+            905,
+            792
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We note that broadly speaking uncertainty estimation for deep conditional models is well explored [20-22, 24, 26, 32, 34, 41, 59-61, 63, 66]. However, research in this area is generally concerned with problem settings where 1) the",
+        "bbox": [
+            511,
+            795,
+            905,
+            854
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>We note that although there is a discussion over how to define and quantify epistemic uncertainty [21, 26, 41, 45, 60], in this work, we stick to an intuitive albeit informal view based on a deep model's knowledge.",
+        "bbox": [
+            511,
+            862,
+            906,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "8064",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "models explicitly evaluate  $p_{\\theta}(\\pmb{x}|c)$  and 2) downstream applications of uncertainty relate directly to the numerical values of  $\\pmb{x}$  and  $p_{\\theta}(\\pmb{x}|c)$ . For example, a cross-entropy-trained image classifier explicitly predicts the probability mass  $P_{\\theta}(\\pmb{x}|c^{*})$  of label  $\\pmb{x}$  given test image  $c^{*}$ .  $P_{\\theta}(\\pmb{x}|c^{*})$  directly relates to the likelihood of classification error – which is an event that we may want to detect using an uncertainty measure, i.e. downstream misclassification detection [22, 24]. Thus, we are able to easily extract useful measures of uncertainty from the model  $P_{\\theta}$ , e.g. the maximum of the softmax.",
+        "bbox": [
+            89,
+            90,
+            480,
+            241
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "On the other hand, text-to-image generation models are not as clear-cut. They may not explicitly evaluate  $p_{\\theta}(\\pmb{x}|c^{*})$  (e.g. diffusion-based models [25, 54]), and even when they do (e.g. vector-quantised autoregressive models [17, 55]) it is hard to directly relate (latent) pixel values directly to our above definitions of uncertainty. For example, simple variations in contrast and brightness that may leave semantics completely unchanged could result in a very high variance in  $p_{\\theta}(\\pmb{x}|c^{*})$ . To give a simple example, inverting the black and white stripes on a zebra would lead to a large Euclidean distance in the image space whilst preserving the semantics of the image. In the next section, we present and discuss a range of potential methods to quantify the uncertainty of text-to-image models with respect to a given prompt.",
+        "bbox": [
+            89,
+            243,
+            482,
+            454
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4. Techniques to Quantify Uncertainty",
+        "text_level": 1,
+        "bbox": [
+            89,
+            469,
+            413,
+            488
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We quantify text-to-image uncertainty by adapting existing methods and proposing a novel approach named Prompt-based UNCertainty estimation (PUNC)(Fig. 4).",
+        "bbox": [
+            89,
+            496,
+            482,
+            542
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1. Adapting Existing Approaches",
+        "text_level": 1,
+        "bbox": [
+            89,
+            551,
+            362,
+            569
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "There are a number of existing methods for quantifying the image-space uncertainty of diffusion models. As diffusion-based models represent a majority of current text-to-image foundation models [8, 18, 48, 50], we choose to adapt these in order to quantify prompt-space uncertainty (Fig. 4). We first present a brief primer on diffusion models to the reader.",
+        "bbox": [
+            89,
+            575,
+            482,
+            667
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1.1. Diffusion Models",
+        "text_level": 1,
+        "bbox": [
+            89,
+            676,
+            256,
+            690
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Diffusion models [25] generate data by iteratively denoising pure noise. The forward process over time  $t \\in [0,1]$  for  $\\boldsymbol{x}_t \\in \\mathbb{R}^D$ , conditioned on a text prompt  $c$ , follows:",
+        "bbox": [
+            89,
+            695,
+            482,
+            742
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}, \\boldsymbol {c}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; a _ {t} \\boldsymbol {x} _ {0}, \\sigma_ {t} ^ {2} \\boldsymbol {I}\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            171,
+            753,
+            482,
+            771
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\pmb{x}_0$  is sampled from an unknown data distribution  $p(\\pmb{x}_0)$ , and  $a_{t},\\sigma_{t}$  define the noise schedule. This formulation allows diffusion models to progressively refine noisy inputs, enabling high-quality data generation through a learned denoising process. By leveraging stochastic differential equations (SDEs) or their deterministic counterparts,",
+        "bbox": [
+            89,
+            784,
+            482,
+            875
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "these models sample from complex distributions in a structured manner. Various solvers [29, 40, 53, 54, 62, 65, 69] can then be employed to efficiently generate realistic samples with a limited number of denoising steps.",
+        "bbox": [
+            511,
+            90,
+            903,
+            152
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Latent diffusion models (LDMs) [39] perform this process within a compressed latent space rather than directly on pixel space. By first encoding images into a more compact representation, LDMs reduce computational demands, enabling high-quality high-resolution generations. We will not consider the latent representation to keep the notation simple, even though experiments will use LDMs.",
+        "bbox": [
+            511,
+            160,
+            906,
+            267
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1.2. Time-step based approaches",
+        "text_level": 1,
+        "bbox": [
+            511,
+            299,
+            756,
+            314
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "DDPM-OOD [23]. This approach, developed for image-space OOD detection, involves noising a test sample  $\\boldsymbol{x}_0^*$  to different points in the forward process to get  $\\{\\boldsymbol{x}_{t_1}^*,\\boldsymbol{x}_{t_2}^*,\\dots \\}$  and denoising them, generating multiple reconstructions  $\\{\\hat{x}_{0_1},\\hat{x}_{0_2},\\ldots \\}$ . Similarity is then measured between each reconstruction and the original image. The idea is that test samples  $\\boldsymbol{x}_0^*$  for which the model has higher epistemic uncertainty will lead to reconstructions that differ from the test sample. However, in our case, we do not have a test image  $\\boldsymbol{x}_0^*$ . In order to adapt DDPM-OOD we propose measuring similarity between reconstructions at different  $t$  and a baseline generation,  $S(\\hat{x}_{0_i},\\hat{x}_0)$ , where  $\\hat{x}_0$  represents an image directly generated using the original test prompt  $c^*$  alone. Both aleatoric uncertainty (e.g. ambiguous prompt) and epistemic uncertainty (e.g. unfamiliar concepts in prompt) could lead to more divergent reconstructions.",
+        "bbox": [
+            511,
+            325,
+            906,
+            568
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "LMD [39]. Another similar approach for image-space OOD detection, known as Learned Manifold Denoising (LMD), instead explores alternative transforms on  $\\boldsymbol{x}_0^*$  to noising. The idea is to lift  $\\boldsymbol{x}_0^*$  from its original data manifold and then map it back using the trained diffusion model. If the sample is in-distribution, the mapping process should return it close to its original position on the data manifold. Conversely, if it is OOD, the mapped position will likely diverge significantly from the original. The corruption process typically involves masking portions of the image, with various mask patterns applied. Similar to DDPM-OOD, similarity is measured between the original image and the reconstructed version after applying the diffusion process. Similar to before, since we lack a testimage, we measure the similarity between reconstructed images and  $\\hat{\\boldsymbol{x}}_{0_i}$ , the image generated directly from the prompt without perturbations.",
+        "bbox": [
+            511,
+            569,
+            908,
+            810
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Similarity Metric. Various methods are used to evaluate similarity between images. Mean-squared error (MSE) is commonly applied between input and reconstruction, while some approaches favor the Learned Perceptual Image Patch Similarity (LPIPS) [70] metric. The choice of similarity metric is a crucial component of these techniques.",
+        "bbox": [
+            511,
+            810,
+            908,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "2We stick with this non-standard notation for the sake of consistency.",
+        "bbox": [
+            107,
+            886,
+            473,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "8065",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/028fd6b0c4c53b8460bc355b984d6d2584ea5339d9111611314b3422a07d66e0.jpg",
+        "image_caption": [
+            "Figure 4. Illustration showing the different baselines and PUNC. PUNC leverages a LVLM to describe generated images and assess similarity with the original prompt, providing a refined uncertainty score. In contrast, baseline methods employ traditional techniques such as noise injection, ensembling, or masking to quantify uncertainty, followed by image-based similarity scoring."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            93,
+            87,
+            906,
+            247
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1.3. Test-Time Ensembling Approaches",
+        "text_level": 1,
+        "bbox": [
+            89,
+            314,
+            382,
+            329
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Test-time ensembling techniques take inspiration from approaches used in the large language model (LLM) community [31], where multiple generations are produced to assess consistency. Given a prompt  $c$ , we generate two images  $\\{\\hat{x}_1, \\hat{x}_2\\}$  and measure the similarity between the two outputs. We denote this technique 2XDM. The underlying rationale is that if the model has high confidence, it will generate similar images. However, if uncertainty is high, it will generate substantially different images. We note that this is similar to the approach proposed in [5], however, that approach uses an ensemble of diffusion models, which is not practical in the setting where we want to quantify the uncertainty of a single pretrained T2I model.",
+        "bbox": [
+            88,
+            332,
+            483,
+            527
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "A key question remains regarding how to find the best measure similarity, and we propose using the same techniques applied in the time-step-based methods. A significant advantage of test-time ensembling approaches is that they can be applied generically to all T2I models. In contrast, timestep-based approaches require a multi-step generation process, limiting the broadness of their applicability. We note that all above approaches require multiple generations, incurring a notable computational overhead at inference time. For example, DDPM-OOD requires  $\\sim 50\\times$  the number of network inferences as a single generation (if the hyperparameters of the original paper are used) [23].",
+        "bbox": [
+            88,
+            530,
+            483,
+            712
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Prompt-based Uncertainty Estimation",
+        "text_level": 1,
+        "bbox": [
+            89,
+            720,
+            419,
+            736
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we propose a new method, Prompt-based Uncertainty estimation for text-to-image generation (PUNC), which leverages Large Vision-Language Models (LVLM) in conjunction with T2I models. We aim to assess the uncertainty in generated images by analyzing the alignment between initial text prompts and reconstructed textual descriptions of the generated images.",
+        "bbox": [
+            89,
+            744,
+            483,
+            849
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Background on Large Vision-Language Model (LVLM)",
+        "text_level": 1,
+        "bbox": [
+            89,
+            869,
+            482,
+            883
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "A LVLM integrates a large language model (LLM) with",
+        "bbox": [
+            89,
+            885,
+            482,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "an image encoder, allowing for robust, multimodal understanding. Given a text prompt  $c$  and an image  $x$ , a LVLM model, with parameters  $\\omega$ , includes:",
+        "bbox": [
+            511,
+            314,
+            903,
+            358
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Image Encoder:  $f_{\\omega}^{\\mathrm{img}}(\\cdot)$  processes the input image  $x$  and produces an embedding:  $z^{\\mathrm{img}} = f_{\\omega}^{\\mathrm{img}}(x)$",
+            "- Text Processor with LLM:  $f_{\\omega}^{\\mathrm{txt}}(\\cdot, \\cdot)$  takes the prompt  $c$  and the image embedding  $z^{\\mathrm{img}}$  as inputs, generating a descriptive or answer-based response:  $\\hat{c} = f_{\\omega}^{\\mathrm{txt}}(c, z^{\\mathrm{img}})$"
+        ],
+        "bbox": [
+            511,
+            359,
+            903,
+            438
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Thus, a LVLM system generates an interpretation  $\\hat{c}$  that reflects the content of the image given an initial prompt.",
+        "bbox": [
+            511,
+            440,
+            903,
+            470
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Prompt-Based Uncertainty Estimation for T2I Generation (PUNC) The previously introduced approaches to uncertainty estimation rely on measuring similarity between images. For a given text prompt  $c$ , a T2I model generates an image  $x \\sim p_{\\theta}(x|c)$ . We reason that, even under low uncertainty, the model can produce images that vary in aspects such as color, texture, or object positioning, while still faithfully representing the semantics of prompt  $c$ . As previously discussed in Sec. 3, this may result in image-based similarity metrics failing to capture uncertainty at a semantic level, since large differences in the image space may not necessarily correspond to meaningful semantic differences.",
+        "bbox": [
+            511,
+            488,
+            905,
+            669
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We propose that advanced LVLMs, such as Molmo [11], LLAMA 3 [15], and GPT-4 [1], are a strong candidate for extracting the semantics from an image, through describing an image via natural language. Building on this insight, our proposed method, PUNC, introduces a novel approach that bypasses direct image comparison and instead evaluates the alignment between generated images and the original prompts by calculating an uncertainty score in the more semantically meaningful text-space. The steps for implementing PUNC are as follows:",
+        "bbox": [
+            511,
+            669,
+            905,
+            819
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Step 1: Initial Prompt and Image Generation Given a test prompt  $c^*$ , we use the T2I model to sample an image  $x$ :",
+        "bbox": [
+            511,
+            820,
+            903,
+            849
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\boldsymbol {x} \\sim p _ {\\boldsymbol {\\theta}} (\\boldsymbol {x} | \\boldsymbol {c} ^ {*})\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            658,
+            859,
+            758,
+            876
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Step 2: Image Interpretation via LVLM With the",
+        "bbox": [
+            511,
+            885,
+            903,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "8066",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/bfd49245a54fb630faa3175b2d199995fca0356c85229b1aa37a4ac1798f6ca0.jpg",
+        "image_caption": [
+            "Figure 5. Disentangling Aleatoric and Epistemic Uncertainty."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            87,
+            446,
+            208
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "generated image  $\\pmb{x}$  and the initial prompt  $c^*$ , the LVLM produces a new descriptive caption  $\\hat{c}$ :",
+        "bbox": [
+            89,
+            258,
+            482,
+            291
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\boldsymbol {c}} = f _ {\\omega} ^ {\\mathrm {t x t}} \\left(\\boldsymbol {c} ^ {*}, f _ {\\omega} ^ {\\mathrm {i m g}} (\\boldsymbol {x})\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            212,
+            301,
+            359,
+            319
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Step 3: Uncertainty Score Calculation The core of PUNC lies in measuring the alignment between  $\\mathbf{c}$  and  $\\hat{\\mathbf{c}}$  to assess the diffusion model's confidence. Specifically, the similarity score  $S(\\mathbf{c},\\hat{\\mathbf{c}})$  is computed as:",
+        "bbox": [
+            89,
+            330,
+            483,
+            391
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS (\\boldsymbol {c} ^ {*}, \\hat {\\boldsymbol {c}}) = \\operatorname {s i m} (\\boldsymbol {c} ^ {*}, \\hat {\\boldsymbol {c}})\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            210,
+            402,
+            362,
+            420
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "High similarity suggests low uncertainty, indicating the generated image closely matches the initial prompt. Conversely, a lower score reveals higher uncertainty, signifying potential divergence from the prompt's intended meaning.",
+        "bbox": [
+            89,
+            431,
+            482,
+            492
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Thus, PUNC is able to predict semantic uncertainties by utilising the powerful image understanding ability of LVLMs and computing similarities in the text space.",
+        "bbox": [
+            89,
+            492,
+            483,
+            537
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Prompt Similarity: Aleatoric and Epistemic via Precision and Recall. In Sec. 3 we discussed the different types of semantic uncertainty for T2I generation. Here we will present a method for disentangling these uncertainties using the precision and recall of the caption given the prompt. Consider an abstract scenario where the prompt and the generated image each contain a discrete set of semantic concepts. We can then calculate the precision and recall of the image concepts with respect to the prompt:",
+        "bbox": [
+            89,
+            537,
+            482,
+            674
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Intuitively, a lack of knowledge about concepts in the prompt (epistemic uncertainty) will result in fewer concepts being preserved in the image, reducing recall. On the other hand, a lack of specificity in the prompt (aleatoric uncertainty) will result in additional concepts being generated, reducing precision. Thus we can use this abstract form of semantic-concept-based recall to measure epistemic uncertainty, and use precision to measure aleatoric uncertainty.",
+        "bbox": [
+            89,
+            674,
+            482,
+            795
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Although we cannot directly measure concepts, we can approximate them using natural language. Moreover, existing similarity scores for text already have the above notions of precision and recall integrated within [36, 43, 72], allowing PUNC to disentangle uncertainty by comparing the caption of the generated image with the original prompt. We note that the image-based approaches in",
+        "bbox": [
+            89,
+            795,
+            482,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Sec. 4.1 are unable to disentangle aleatoric from epistemic uncertainty. We focus on two different existing text-based similarity metrics, ROUGE [36], and BERTScore [72]. The former calculates similarity (precision or recall) by directly matching (sequences of) words between two text sequences. The latter on the other hand measures similarity in the deep feature space of a text encoder.",
+        "bbox": [
+            511,
+            90,
+            903,
+            196
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As illustrated in Fig. 5, we define Prompt Similarity as the semantic overlap between the input prompt and the generated descriptive caption. More formally, let  $P_{\\mathrm{GT}}$  denote the set of semantic concepts in the ground-truth prompt used to generate an image, and  $P_{\\mathrm{VQA}}$  the set of concepts extracted from the LVLM-generated caption. The resulting uncertainties can be characterized as follows:",
+        "bbox": [
+            511,
+            196,
+            906,
+            301
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Epistemic Uncertainty, which captures missing concepts in  $P_{\\mathrm{VQA}}$  that are present in  $P_{\\mathrm{GT}}$ , reflects an incomplete understanding of the intended semantics. It is defined as:",
+        "bbox": [
+            511,
+            301,
+            903,
+            348
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nU _ {\\text {e p i s t e m i c}} = 1 - \\frac {\\left| P _ {\\mathrm {G T}} \\cap P _ {\\mathrm {V Q A}} \\right|}{\\left| P _ {\\mathrm {G T}} \\right|} = 1 - \\text {R e c a l l} \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            560,
+            358,
+            903,
+            391
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Aleatoric Uncertainty, on the other hand, arises from spurious or extra concepts in  $P_{\\mathrm{VQA}}$  that do not appear in  $P_{\\mathrm{GT}}$ , indicating ambiguity in the model's response. It is defined as:",
+        "bbox": [
+            511,
+            402,
+            903,
+            460
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nU _ {\\text {a l e a t o r i c}} = 1 - \\frac {\\left| P _ {\\mathrm {G T}} \\cap P _ {\\mathrm {V Q A}} \\right|}{\\left| P _ {\\mathrm {V Q A}} \\right|} = 1 - \\text {P r e c i s i o n} \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            553,
+            470,
+            903,
+            503
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Further details on the computation of precision and recall can be found in Appendix A.",
+        "bbox": [
+            511,
+            513,
+            906,
+            545
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments & Discussions",
+        "text_level": 1,
+        "bbox": [
+            511,
+            558,
+            766,
+            574
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We test PUNC on Various task and various dataset and compared PUNC to DDPM-OOD [23], LMD [39] and 2XDM. We first introduce on Sec. 5 how we have constitute the dataset of prompts. Then on Sec. 5.1 we discuss of different results on these dataset. Then Appendix F explains how we can use the uncertainty of text-to-image model to detect if a model has been trained with some concept related to some fictional character of politicians. Appendix F.3 shows how uncertainty can help to detect bias of T2I models. To conduct our experiments and evaluate the various techniques for quantifying uncertainty, we utilized four T2I models: Stable Diffusion 1.5 (SDv1.5) [50], SDXS [68], SDXL [48], and PixArt-  $\\Sigma$  [7]. These diverse T2I models allow us to test a range of architectures, each with unique characteristics and potential weaknesses. SDv1.5, being one of the earliest models, likely has a smaller and less diverse training dataset compared to the others. In contrast, SDXS generates in a single-step, different to the multi-step processes in other models.",
+        "bbox": [
+            511,
+            583,
+            906,
+            869
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Dataset of prompts To conduct our experiments, we require both an in-distribution dataset of prompts and",
+        "bbox": [
+            511,
+            871,
+            903,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "8067",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "datasets with uncertainty. For the in-distribution prompts, we use the dataset proposed by [13], where high-quality descriptions of ImageNet [12] images are generated using GPT-4[13], referred to as Normal. To simulate epistemic uncertainty, we create out-of-distribution (OOD) datasets using images from underrepresented domains like remote sensing, texture, and microscopic datasets, captioned with LLAVA Next[35], labeled as Remote Sensing, Texture, and Microscopic. For aleatoric uncertainty, we generate two datasets: Vague, containing 2,000 minimally descriptive prompts (e.g., \"An image of ***\"), and Corrupted, derived from Normal by adding grammatical errors (Level 1) and removing  $50\\%$  of words (Level 2). A detailed summary of these datasets is presented in Tab. 1 and Appendix E. In Appendix C, we provide the details of the hyperparameters used in our experiments.",
+        "bbox": [
+            89,
+            90,
+            480,
+            333
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/171fe859994b014f2125487c611fc0c28b4bf4d771eef1699d8169708a968c43.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset Type</td><td>Dataset Name</td><td>Description</td></tr><tr><td>In-Distribution (ID)</td><td>Normal</td><td>20k High-quality detailed prompts from ImageNet [12] images using GPT-4[13].</td></tr><tr><td>Out-of-Distribution (OOD)</td><td>Remote Sensing Texture Microscopic</td><td>20k Prompts from remote sensing images [47]5k Prompts from texture datasets10k Prompts from microscopic images [44]</td></tr><tr><td>Aleatoric Uncertainty</td><td>VagueAdversarialCorrupted</td><td>2k prompts with minimal context, e.g., &quot;An image of ***,&quot; where &quot;****&quot; is an ImageNet class name, creating ambiguous input.1k altered prompts from Normal using Unlearn-DiffAtk [73]20k Altered prompts from Normal:• Level 1: Grammatical mistakes and spelling errors.• Level 2: Random removal of 50% of words in Level 1 prompts.</td></tr></table>",
+        "bbox": [
+            130,
+            345,
+            444,
+            474
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Summary of prompt datasets for unc. quantification.",
+        "bbox": [
+            104,
+            486,
+            467,
+            500
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1. Results for Uncertainty Estimation",
+        "text_level": 1,
+        "bbox": [
+            89,
+            527,
+            393,
+            544
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Tab. 3 presents the results of PUNC when using Molmo as the LVLM to assess epistemic uncertainty in the T2I models. Our approach, PUNC, outperforms most state-of-the-art techniques on average. However, we observe more mixed results on the Texture dataset, likely due to uncertainty about the presence of texture images in the training data of the T2I models.",
+        "bbox": [
+            89,
+            550,
+            482,
+            655
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In Tab. 4, we show the results of PUNC under aleatoric uncertainty, again using Molmo as the LVLM. Our approach outperforms most state-of-the-art methods in this setting as well. Notably, while the ROUGE score tends to perform better for epistemic uncertainty, the optimal metric is less clear for aleatoric uncertainty. Here, the choice between ROUGE and BERTScore may depend on the specific application: because corruption introduces spelling and grammatical errors, BERTScore—which captures semantic similarity—may be more suitable. For additional results using different LVLMs, please refer to Appendix D and Appendix B. In Appendix G, we present qualitative results.",
+        "bbox": [
+            89,
+            656,
+            482,
+            838
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2. Applications of uncertainty",
+        "text_level": 1,
+        "bbox": [
+            89,
+            848,
+            336,
+            864
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Our study explores multiple applications of uncertainty quantification in T2I models, with detailed results provided",
+        "bbox": [
+            89,
+            869,
+            482,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "in Appendix F. First, we investigate deepfake detection by examining the model's ability to generate recognizable images of prominent politicians (see Appendix F.1). We examined whether diffusion models could generate convincing images of G7 politicians. Using 100 prompts containing their names, we generated images with four diffusion models and queried a LVLM to determine if the intended politician was present in each image (yes/no), thus measuring concept recognition uncertainty. Results are presented in Table 2.",
+        "bbox": [
+            511,
+            90,
+            903,
+            241
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Next, we address copyright concerns by evaluating whether T2I models can accurately replicate iconic, copyrighted characters such as Mickey Mouse and Darth Vader (see Appendix F.2). Finally, we analyze bias in T2I models by testing their capacity to generate individuals across different genders and races for specific job roles (see Appendix F.3).",
+        "bbox": [
+            511,
+            242,
+            903,
+            347
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Using a LVLM, we evaluate the presence of specific individuals, copyrighted characters, or demographic concepts in the generated images, demonstrating how uncertainty estimation can aid in detecting potential misuses of generative models. These experiments illustrate the value of text-based uncertainty quantification in analyzing T2I model performance across ethically and legally sensitive applications, enabling assessments that are challenging to achieve with image-based generation alone.",
+        "bbox": [
+            511,
+            348,
+            906,
+            484
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            513,
+            497,
+            633,
+            513
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this work, we set out to be first to investigate uncertainty estimation T2I generation, where the uncertainty is with respect to the text prompt. We both adapt existing image-space uncertainty approaches, as well as present Prompt-based UNCertainty Estimation for T2I models (PUNC), the first approach designed specifically to quantify uncertainty in T2I generation. By leveraging the interpretive power of Large Vision-Language Models to extract semantics from a generated image, PUNC addresses the challenges associated with measuring multimodal uncertainty, enabling a deeper understanding of model behavior with respect to prompt-based conditioning. Our results validate that PUNC provides reliable, fine-grained assessments of aleatoric and epistemic uncertainties, outperforming existing methods that work for image generation.",
+        "bbox": [
+            511,
+            523,
+            903,
+            750
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "PUNC opens avenues for new applications, such as improving model robustness in biased or OOD scenarios and safeguarding against unauthorized content generation, including deepfakes and copyrighted characters. Our experiments show that diffusion models, while powerful, still exhibit significant variances in their ability to adhere to prompt details, particularly with ambiguous, corrupt prompts or OOD prompts. By identifying such uncertainties, PUNC not only supports the development of more trustworthy generative AI but also aids in aligning these",
+        "bbox": [
+            511,
+            750,
+            906,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "8068",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/b80375c0c9bb1ca15cccb13e9d02894a56d3ece7d1a15a401578a1fdb70a4e62.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Justin Trudeau</td><td>King Charles</td><td>Fumio Kishida</td><td>Olaf Scholz</td><td>Emmanuel Macron</td><td>Joe Biden</td><td>Giorgia Meloni</td><td>AVERAGE</td></tr><tr><td>SDXL</td><td>94</td><td>94</td><td>88</td><td>88</td><td>98</td><td>96</td><td>21</td><td>95.5</td></tr><tr><td>SDv1.5</td><td>78</td><td>60</td><td>62</td><td>88</td><td>94</td><td>80</td><td>18</td><td>78</td></tr><tr><td>SDXS</td><td>80</td><td>40</td><td>70</td><td>82</td><td>94</td><td>76</td><td>28</td><td>72.5</td></tr><tr><td>PixArt</td><td>36</td><td>76</td><td>62</td><td>78</td><td>94</td><td>90</td><td>15</td><td>74</td></tr><tr><td>AVERAGE</td><td>72</td><td>67.5</td><td>70.5</td><td>84</td><td>95</td><td>85.5</td><td>20.5</td><td>70.71</td></tr></table>",
+        "bbox": [
+            215,
+            88,
+            781,
+            151
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/b660b9d8dd8679298c7986b79d8b083a83775bebe3e10881d71011565a1153b1.jpg",
+        "table_caption": [
+            "Table 2. Accuracy (%) of diffusion models in generating recognizable images of G7 politicians. Higher accuracy indicates lower uncertainty in the model's ability to represent specific individuals."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\"></td><td>auroc ↑</td><td>SDXS aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>PixArt aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>SDv1.5 aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>SDXL aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>Average aupr ↑</td><td>fpr95↓</td></tr><tr><td rowspan=\"6\">microscopic</td><td>DDPM-OOD mse_and_lpips</td><td></td><td></td><td></td><td>39.61%</td><td>56.38%</td><td>100.00%</td><td>71.64%</td><td>71.52%</td><td>93.62%</td><td>87.25%</td><td>83.67%</td><td>51.43%</td><td>66.17%</td><td>70.53%</td><td>81.68%</td></tr><tr><td>LMD lpips alex</td><td></td><td></td><td></td><td>45.82%</td><td>45.24%</td><td>99.10%</td><td>16.81%</td><td>31.61%</td><td>99.88%</td><td>5.75%</td><td>28.12%</td><td>100.00%</td><td>22.79%</td><td>34.99%</td><td>99.66%</td></tr><tr><td>2XDM lpips alex</td><td>24.49%</td><td>32.60%</td><td>99.91%</td><td>24.32%</td><td>33.21%</td><td>99.99%</td><td>16.19%</td><td>32.42%</td><td>99.94%</td><td>30.58%</td><td>36.99%</td><td>98.25%</td><td>23.90%</td><td>33.81%</td><td>99.52%</td></tr><tr><td>PUNC ROUGE (1_recall)</td><td>84.00%</td><td>85.10%</td><td>57.07%</td><td>95.47%</td><td>97.06%</td><td>30.84%</td><td>87.24%</td><td>90.66%</td><td>48.90%</td><td>83.56%</td><td>84.31%</td><td>52.78%</td><td>87.57%</td><td>89.28%</td><td>47.40%</td></tr><tr><td>PUNC ROUGE (L_recall)</td><td>87.64%</td><td>89.56%</td><td>55.28%</td><td>96.96%</td><td>98.02%</td><td>16.85%</td><td>90.20%</td><td>93.40%</td><td>43.56%</td><td>88.19%</td><td>89.84%</td><td>47.44%</td><td>90.75%</td><td>92.71%</td><td>40.78%</td></tr><tr><td>PUNC BERT (recall)</td><td>88.45%</td><td>89.71%</td><td>48.68%</td><td>97.83%</td><td>98.41%</td><td>10.32%</td><td>87.75%</td><td>91.70%</td><td>51.03%</td><td>83.10%</td><td>85.46%</td><td>58.14%</td><td>89.28%</td><td>91.32%</td><td>42.04%</td></tr><tr><td rowspan=\"6\">remote sensing</td><td>DDPM-OOD mse_and_lpips</td><td></td><td></td><td></td><td>80.00%</td><td>85.37%</td><td>99.28%</td><td>74.36%</td><td>71.89%</td><td>71.00%</td><td>65.47%</td><td>58.25%</td><td>71.97%</td><td>73.28%</td><td>71.84%</td><td>80.75%</td></tr><tr><td>LMD lpips alex</td><td></td><td></td><td></td><td>90.65%</td><td>93.11%</td><td>74.49%</td><td>78.93%</td><td>79.31%</td><td>81.05%</td><td>86.60%</td><td>85.39%</td><td>49.19%</td><td>85.39%</td><td>85.94%</td><td>68.24%</td></tr><tr><td>2XDM lpips alex</td><td>53.88%</td><td>59.21%</td><td>98.23%</td><td>54.43%</td><td>57.99%</td><td>98.78%</td><td>30.66%</td><td>38.80%</td><td>98.76%</td><td>23.54%</td><td>35.35%</td><td>96.23%</td><td>40.63%</td><td>47.83%</td><td>98.00%</td></tr><tr><td>PUNC ROUGE (1_recall)</td><td>85.59%</td><td>86.59%</td><td>63.87%</td><td>94.14%</td><td>95.97%</td><td>49.29%</td><td>73.24%</td><td>72.93%</td><td>79.56%</td><td>73.07%</td><td>74.54%</td><td>83.48%</td><td>81.51%</td><td>82.51%</td><td>69.05%</td></tr><tr><td>PUNC ROUGE (L_recall)</td><td>89.85%</td><td>90.63%</td><td>51.96%</td><td>97.01%</td><td>97.94%</td><td>16.08%</td><td>83.63%</td><td>84.65%</td><td>65.35%</td><td>85.68%</td><td>87.09%</td><td>64.00%</td><td>89.04%</td><td>90.08%</td><td>49.34%</td></tr><tr><td>PUNC BERT (recall)</td><td>80.50%</td><td>82.56%</td><td>75.19%</td><td>91.93%</td><td>93.08%</td><td>43.30%</td><td>61.27%</td><td>64.74%</td><td>90.08%</td><td>60.71%</td><td>65.32%</td><td>90.94%</td><td>73.60%</td><td>76.42%</td><td>74.88%</td></tr><tr><td rowspan=\"6\">Texture</td><td>DDPM-OOD mse_and_lpips</td><td></td><td></td><td></td><td>73.22%</td><td>62.02%</td><td>89.24%</td><td>70.21%</td><td>55.97%</td><td>99.80%</td><td>74.24%</td><td>51.97%</td><td>84.71%</td><td>72.56%</td><td>56.65%</td><td>91.25%</td></tr><tr><td>LMD lpips alex</td><td></td><td></td><td></td><td>63.40%</td><td>50.47%</td><td>97.52%</td><td>26.74%</td><td>16.06%</td><td>99.89%</td><td>35.66%</td><td>23.22%</td><td>99.95%</td><td>41.93%</td><td>29.92%</td><td>99.12%</td></tr><tr><td>2XDM lpips alex</td><td>32.80%</td><td>20.84%</td><td>99.98%</td><td>37.21%</td><td>23.36%</td><td>99.37%</td><td>29.58%</td><td>24.27%</td><td>99.93%</td><td>28.87%</td><td>16.49%</td><td>98.88%</td><td>32.11%</td><td>21.24%</td><td>99.54%</td></tr><tr><td>PUNC ROUGE (1_recall)</td><td>40.77%</td><td>72.88%</td><td>95.54%</td><td>83.26%</td><td>94.80%</td><td>73.48%</td><td>46.61%</td><td>75.64%</td><td>90.87%</td><td>49.16%</td><td>78.04%</td><td>93.57%</td><td>54.95%</td><td>80.34%</td><td>88.36%</td></tr><tr><td>PUNC ROUGE (L_recall)</td><td>45.64%</td><td>76.90%</td><td>95.99%</td><td>84.10%</td><td>95.13%</td><td>74.21%</td><td>47.60%</td><td>77.70%</td><td>93.14%</td><td>50.71%</td><td>80.05%</td><td>94.35%</td><td>57.01%</td><td>82.44%</td><td>89.42%</td></tr><tr><td>PUNC BERT (recall)</td><td>49.11%</td><td>79.11%</td><td>94.65%</td><td>84.40%</td><td>95.14%</td><td>69.99%</td><td>47.60%</td><td>78.80%</td><td>93.85%</td><td>46.72%</td><td>79.20%</td><td>96.87%</td><td>56.96%</td><td>83.06%</td><td>88.84%</td></tr></table>",
+        "bbox": [
+            135,
+            205,
+            864,
+            393
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/4433cec5bf99037f5c9ec0bab56b699020f6329088988d04ac39c4039ab20c8f.jpg",
+        "table_caption": [
+            "Table 3. Performance comparison of PUNC applied to different Out of Distribution concepts and different text-to-image models"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\"></td><td>auroc ↑</td><td>SDXS aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>PixArt aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>SDv1.5 aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>SDXL aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>Average aupr ↑</td><td>fpr95↓</td></tr><tr><td rowspan=\"6\">Vague</td><td>DDPM-OOD mse.and.Lpips</td><td></td><td></td><td></td><td>31.41%</td><td>6.06%</td><td>96.45%</td><td>50.06%</td><td>9.30%</td><td>96.52%</td><td>40.59%</td><td>5.36%</td><td>97.37%</td><td>40.69%</td><td>6.91%</td><td>96.78%</td></tr><tr><td>LMD lipps alex</td><td></td><td></td><td></td><td>39.26%</td><td>6.91%</td><td>96.35%</td><td>47.62%</td><td>9.54%</td><td>97.70%</td><td>40.77%</td><td>5.70%</td><td>97.95%</td><td>42.55%</td><td>7.38%</td><td>97.33%</td></tr><tr><td>2XDM lipps alex</td><td>51.28%</td><td>10.10%</td><td>94.35%</td><td>43.82%</td><td>8.57%</td><td>96.89%</td><td>50.97%</td><td>10.09%</td><td>95.96%</td><td>57.83%</td><td>10.01%</td><td>94.37%</td><td>50.98%</td><td>9.69%</td><td>95.39%</td></tr><tr><td>PUNC ROUGE (1_precision)</td><td>99.99%</td><td>100.00%</td><td>0.00%</td><td>100.00%</td><td>100.00%</td><td>0.00%</td><td>99.99%</td><td>100.00%</td><td>0.00%</td><td>99.99%</td><td>100.00%</td><td>0.00%</td><td>99.99%</td><td>100.00%</td><td>0.00%</td></tr><tr><td>PUNC ROUGE (L_precision)</td><td>99.99%</td><td>100.00%</td><td>0.00%</td><td>100.00%</td><td>100.00%</td><td>0.00%</td><td>99.99%</td><td>100.00%</td><td>0.00%</td><td>100.00%</td><td>100.00%</td><td>0.00%</td><td>100.00%</td><td>100.00%</td><td>0.00%</td></tr><tr><td>PUNC BERT (precision)</td><td>60.63%</td><td>92.24%</td><td>82.85%</td><td>67.02%</td><td>94.54%</td><td>83.10%</td><td>37.35%</td><td>86.51%</td><td>92.70%</td><td>70.85%</td><td>95.13%</td><td>78.00%</td><td>58.96%</td><td>92.11%</td><td>84.16%</td></tr><tr><td rowspan=\"6\">Adversarial</td><td>DDPM-OOD mse.and.Lpips</td><td></td><td></td><td></td><td>39.54%</td><td>1.78%</td><td>95.72%</td><td>50.89%</td><td>2.54%</td><td>95.91%</td><td>38.97%</td><td>1.10%</td><td>98.38%</td><td>43.13%</td><td>1.81%</td><td>96.67%</td></tr><tr><td>LMD lipps alex</td><td></td><td></td><td></td><td>43.71%</td><td>1.95%</td><td>96.24%</td><td>53.29%</td><td>2.89%</td><td>94.43%</td><td>41.64%</td><td>1.17%</td><td>97.48%</td><td>46.21%</td><td>2.00%</td><td>96.05%</td></tr><tr><td>2XDM lipps alex</td><td>36.94%</td><td>1.94%</td><td>98.92%</td><td>47.95%</td><td>2.54%</td><td>96.03%</td><td>53.04%</td><td>2.83%</td><td>95.69%</td><td>55.19%</td><td>1.96%</td><td>96.20%</td><td>48.28%</td><td>2.32%</td><td>96.71%</td></tr><tr><td>PUNC ROUGE (1_precision)</td><td>64.14%</td><td>97.77%</td><td>48.80%</td><td>76.32%</td><td>98.90%</td><td>48.00%</td><td>63.59%</td><td>97.73%</td><td>49.50%</td><td>62.97%</td><td>97.64%</td><td>49.20%</td><td>66.76%</td><td>98.01%</td><td>48.87%</td></tr><tr><td>PUNC ROUGE (L_precision)</td><td>63.20%</td><td>97.75%</td><td>49.00%</td><td>76.26%</td><td>98.90%</td><td>48.40%</td><td>64.09%</td><td>97.90%</td><td>49.70%</td><td>63.76%</td><td>97.87%</td><td>49.60%</td><td>66.83%</td><td>98.11%</td><td>49.17%</td></tr><tr><td>PUNC BERT (precision)</td><td>73.98%</td><td>98.72%</td><td>78.31%</td><td>72.95%</td><td>98.93%</td><td>86.40%</td><td>45.09%</td><td>96.93%</td><td>94.18%</td><td>72.69%</td><td>98.80%</td><td>81.40%</td><td>66.18%</td><td>98.34%</td><td>85.07%</td></tr><tr><td rowspan=\"6\">Corrupt l11</td><td>DDPM-OOD mse.and.Lpips</td><td></td><td></td><td></td><td>48.87%</td><td>44.82%</td><td>87.43%</td><td>53.20%</td><td>54.43%</td><td>95.48%</td><td>41.02%</td><td>29.54%</td><td>96.32%</td><td>47.70%</td><td>42.93%</td><td>93.08%</td></tr><tr><td>LMD lipps alex</td><td></td><td></td><td></td><td>51.23%</td><td>47.48%</td><td>89.99%</td><td>57.53%</td><td>56.49%</td><td>90.92%</td><td>40.94%</td><td>30.30%</td><td>96.32%</td><td>49.90%</td><td>44.76%</td><td>92.41%</td></tr><tr><td>2XDM lipps alex</td><td>25.48%</td><td>36.16%</td><td>99.29%</td><td>50.43%</td><td>49.86%</td><td>90.10%</td><td>54.97%</td><td>54.29%</td><td>95.69%</td><td>54.73%</td><td>38.27%</td><td>93.29%</td><td>46.40%</td><td>44.65%</td><td>94.59%</td></tr><tr><td>PUNC ROUGE (1_precision)</td><td>28.75%</td><td>38.34%</td><td>98.77%</td><td>54.58%</td><td>55.29%</td><td>96.02%</td><td>27.69%</td><td>37.57%</td><td>98.78%</td><td>27.45%</td><td>37.27%</td><td>98.54%</td><td>34.62%</td><td>42.12%</td><td>98.03%</td></tr><tr><td>PUNC ROUGE (L_precision)</td><td>27.29%</td><td>38.11%</td><td>99.15%</td><td>54.32%</td><td>55.59%</td><td>96.16%</td><td>27.30%</td><td>38.07%</td><td>99.45%</td><td>27.35%</td><td>37.89%</td><td>99.37%</td><td>34.07%</td><td>42.42%</td><td>98.53%</td></tr><tr><td>PUNC BERT (precision)</td><td>85.85%</td><td>84.45%</td><td>73.94%</td><td>79.43%</td><td>78.83%</td><td>87.21%</td><td>54.97%</td><td>53.73%</td><td>94.51%</td><td>75.76%</td><td>72.91%</td><td>84.65%</td><td>74.00%</td><td>72.48%</td><td>85.08%</td></tr><tr><td rowspan=\"6\">Corrupt l12</td><td>DDPM-OOD lipps alex</td><td></td><td></td><td></td><td>43.23%</td><td>43.94%</td><td>91.23%</td><td>59.46%</td><td>60.14%</td><td>93.71%</td><td>48.79%</td><td>34.44%</td><td>95.74%</td><td>50.49%</td><td>46.17%</td><td>93.56%</td></tr><tr><td>LMD lipps alex</td><td></td><td></td><td></td><td>46.90%</td><td>45.65%</td><td>92.04%</td><td>59.94%</td><td>58.90%</td><td>90.54%</td><td>40.88%</td><td>30.42%</td><td>96.35%</td><td>49.24%</td><td>44.99%</td><td>92.98%</td></tr><tr><td>2XDM lipps alex</td><td>58.15%</td><td>58.50%</td><td>93.98%</td><td>52.08%</td><td>53.36%</td><td>90.19%</td><td>53.57%</td><td>54.17%</td><td>96.34%</td><td>60.24%</td><td>44.23%</td><td>91.78%</td><td>56.01%</td><td>52.57%</td><td>93.07%</td></tr><tr><td>PUNC ROUGE (1_precision)</td><td>28.76%</td><td>38.10%</td><td>98.59%</td><td>54.77%</td><td>55.49%</td><td>96.02%</td><td>27.66%</td><td>37.42%</td><td>98.75%</td><td>27.39%</td><td>37.23%</td><td>98.53%</td><td>34.64%</td><td>42.06%</td><td>97.97%</td></tr><tr><td>PUNC ROUGE (L_precision)</td><td>27.33%</td><td>37.87%</td><td>99.16%</td><td>54.51%</td><td>55.87%</td><td>96.11%</td><td>27.31%</td><td>37.92%</td><td>99.35%</td><td>27.33%</td><td>37.83%</td><td>99.36%</td><td>34.12%</td><td>42.37%</td><td>98.50%</td></tr><tr><td>PUNC BERT (precision)</td><td>85.96%</td><td>84.23%</td><td>73.54%</td><td>79.47%</td><td>78.80%</td><td>87.05%</td><td>54.78%</td><td>53.51%</td><td>94.79%</td><td>76.04%</td><td>73.44%</td><td>84.58%</td><td>74.06%</td><td>72.49%</td><td>84.99%</td></tr></table>",
+        "bbox": [
+            133,
+            433,
+            864,
+            667
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Performance comparison all applications (Vague, Adversarial, Corrupt Lvl1, Corrup Lvl2)",
+        "bbox": [
+            189,
+            679,
+            799,
+            693
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "models with ethical and regulatory standards.",
+        "bbox": [
+            89,
+            720,
+            390,
+            734
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our contributions also include a dataset specifically tailored to uncertainty quantification in text-to-image models, which we hope will serve as a benchmark for future research. Ultimately, this work highlights the potential of prompt-based uncertainty estimation as a critical tool for enhancing the reliability and safety of generative models in real-world applications.",
+        "bbox": [
+            89,
+            742,
+            482,
+            848
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitations. While PUNC demonstrates promising results in quantifying uncertainty for text-to-image generation, several limitations remain. First, our approach",
+        "bbox": [
+            89,
+            854,
+            483,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "relies heavily on the interpretive abilities of LVLMs, which, despite their effectiveness, may introduce biases or inaccuracies inherited from their training data. Second, our analysis focuses primarily on English-language prompts, limiting our method's applicability to multilingual or culturally diverse datasets.",
+        "bbox": [
+            511,
+            719,
+            906,
+            811
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments. This work was performed using HPC resources from GENCI-IDRIS (Grant 2023-[AD011011970R3]), and EuroHPC Development access to LEONARDO, hosted by CINECA.",
+        "bbox": [
+            511,
+            829,
+            906,
+            891
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "8069",
+        "bbox": [
+            482,
+            944,
+            514,
+            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 preprint arXiv:2303.08774, 2023. 5",
+            "[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2",
+            "[3] Nouar AlDahoul, Talal Rahwan, and Yasir Zaki. Aigenerated faces free from racial and gender stereotypes. arXiv preprint arXiv:2402.01002, 2024. 6",
+            "[4] Georges Le Bellier and Nicolas Audebert. Detecting Out-Of-Distribution Earth Observation Images with Diffusion Models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 481–491, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 2, 3",
+            "[5] Lucas Berry, Axel Brando, and David Meger. Shedding light on large generative networks: Estimating epistemic uncertainty in diffusion models. In The 40th Conference on Uncertainty in Artificial Intelligence, 2024. 2, 3, 5",
+            "[6] Matthew A Chan, Maria J Molina, and Christopher A Metzler. Hyper-diffusion: Estimating epistemic and aleatoric uncertainty with a single model. arXiv preprint arXiv:2402.03478, 2024. 3",
+            "[7] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart $\\backslash$ sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 3, 6, 2",
+            "[8] Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\\mathbb{S}\\backslash \\mathbb{alpha}\\mathbb{S}$ : Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2, 4",
+            "[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 2, 3",
+            "[10] Charles Corbière, Nicolas Thome, Avner Bar-Hen, Matthieu Cord, and Patrick Pérez. Addressing failure prediction by learning model confidence. Advances in Neural Information Processing Systems, 32, 2019. 2",
+            "[11] Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 5, 2",
+            "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7, 2"
+        ],
+        "bbox": [
+            93,
+            114,
+            483,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[13] Ning Ding, Yehui Tang, Zhongqian Fu, Chao Xu, Kai Han, and Yunhe Wang. Gpt4image: Can large pre-trained models help vision models on perception tasks? arXiv e-prints, pages arXiv-2306, 2023. 7, 2",
+            "[14] Zhekai Du and Jingjing Li. Diffusion-based probabilistic uncertainty estimation for active domain adaptation. Advances in Neural Information Processing Systems, 36: 17129-17155, 2023. 3",
+            "[15] 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 preprint arXiv:2407.21783, 2024. 5, 2",
+            "[16] Nikita Durasov, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Masksembles for uncertainty estimation. In CVPR, 2021. 2",
+            "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 4",
+            "[18] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In Proceedings of the 41st International Conference on Machine Learning, pages 12606-12633. PMLR, 2024. 2, 4",
+            "[19] Gianni Franchi, Andrei Bursuc, Emanuel Aldea, Séverine Dubuisson, and Isabelle Bloch. TRADI: Tracking deep neural network weight distributions. In ECCV, 2020. 2",
+            "[20] Gianni Franchi, Xuanlong Yu, Andrei Bursuc, Emanuel Aldea, Severine Dubuisson, and David Filliat. Latent discriminant deterministic uncertainty. In Computer Vision – ECCV 2022, pages 243–260, Cham, 2022. Springer Nature Switzerland. 2, 3",
+            "[21] Yarin Gal. Uncertainty in Deep Learning. PhD thesis, University of Cambridge, 2016. 2, 3",
+            "[22] Yonatan Geifman and Ran El-Yaniv. Selective classification for deep neural networks. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 4885-4894, Red Hook, NY, USA, 2017. Curran Associates Inc. 3, 4",
+            "[23] Mark S Graham, Walter HL Pinaya, Petru-Daniel Tudosiu, Parashkev Nachev, Sebastien Ourselin, and Jorge Cardoso. Denoising diffusion models for out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2948-2957, 2023. 2, 3, 4, 5, 6",
+            "[24] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2017. 3, 4",
+            "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4"
+        ],
+        "bbox": [
+            93,
+            92,
+            903,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "8070",
+        "bbox": [
+            482,
+            944,
+            516,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[26] Eyke Hüllermeier and Willem Waegeman. *Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods*. *Machine Learning*, 110(3), 2021. 2, 3",
+            "[27] Pavel Izmailov, Sharad Vikram, Matthew D Hoffman, and Andrew Gordon Gordon Wilson. What are bayesian neural network posteriors really like? In ICML, 2021. 2",
+            "[28] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583-589, 2021. 2",
+            "[29] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, pages 26565-26577. Curran Associates, Inc., 2022. 4",
+            "[30] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 2",
+            "[31] Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Representations, 2023. 3, 5",
+            "[32] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeurIPS, 2017. 2, 3",
+            "[33] Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416-1421, 2023. 2",
+            "[34] Olivier Laurent, Adrien Lafage, Enzo Tartaglione, Geoffrey Daniel, Jean-Marc Martinez, Andrei Bursuc, and Gianni Franchi. Packed-ensembles for efficient uncertainty estimation. In ICLR, 2023. 2, 3",
+            "[35] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 7, 2, 3",
+            "[36] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6, 1",
+            "[37] Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. In *Forty-first International Conference on Machine Learning*. 2",
+            "[38] Krzysztof Lis, Krishna Nakka, Pascal Fua, and Mathieu Salzmann. Detecting the unexpected via image resynthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2152-2161, 2019. 3",
+            "[39] Zhenzhen Liu, Jin Peng Zhou, Yufan Wang, and Kilian Q Weinberger. Unsupervised out-of-distribution detection with diffusion inpainting. In International Conference on Machine Learning, pages 22528-22538. PMLR, 2023. 2, 3, 4, 6",
+            "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan LI, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Advances in Neural Information Processing Systems, pages 5775-5787. Curran Associates, Inc., 2022. 4"
+        ],
+        "bbox": [
+            91,
+            90,
+            480,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[41] Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3",
+            "[42] Radford M Neal et al. Mcmc using hamiltonian dynamics. Handbook of markov chain monte carlo, 2011. 2",
+            "[43] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6",
+            "[44] Gangadhar Payyavula. Nuclei segmentation in microscope cell images. https://www.kaggle.com/datasets/gangadhar/nuclei-segmentation-in-microscope-cell-images, 2018. Accessed: 2024-09-01. 7, 2, 3",
+            "[45] Tim Pearce, Alexandra Brintrup, and Jun Zhu. Understanding softmax confidence and uncertainty, 2021. 3",
+            "[46] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2",
+            "[47] Jeremy Pinto. Bigearthnet. https://github.com/ jerpint/bigearthnet, 2022. Accessed: 2024-09-01. 7, 2, 3",
+            "[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. In The Twelfth International Conference on Learning Representations, 2023. 2, 4, 6",
+            "[49] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021.",
+            "[50] 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. 2, 4, 6",
+            "[51] 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. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2, 3",
+            "[52] Jaskirat Singh and Liang Zheng. Divide, evaluate, and refine: Evaluating and improving text-to-image alignment with iterative vqa feedback. Advances in Neural Information Processing Systems, 36:70799-70811, 2023. 2",
+            "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 4",
+            "[54] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 4"
+        ],
+        "bbox": [
+            514,
+            90,
+            903,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "8071",
+        "bbox": [
+            482,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[55] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2409.17146, 2024. 4",
+            "[56] Tomas Vojir, Tomáš Šipka, Rahaf Aljundi, Nikolay Chumerin, Daniel Olmeda Reino, and Jiri Matas. Road anomaly detection by partial image reconstruction with segmentation coupling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15651-15660, 2021. 3",
+            "[57] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In ICML, 2011. 2",
+            "[58] Yeming Wen, Dustin Tran, and Jimmy Ba. BatchEnsemble: an alternative approach to efficient ensemble and lifelong learning. In ICLR, 2019. 2",
+            "[59] Guoxuan Xia and Christos-Savvas Bouganis. Augmenting softmax information for selective classification with out-of-distribution data. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1995-2012, 2022. 3",
+            "[60] Guoxuan Xia and Christos-Savvas Bouganis. On the usefulness of deep ensemble diversity for out-of-distribution detection, 2022. 3",
+            "[61] Guoxuan Xia and Christos-Savvas Bouganis. Window-based early-exit cascades for uncertainty estimation: When deep ensembles are more efficient than single models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17368-17380, 2023. 2, 3",
+            "[62] Guoxuan Xia, Duolikun Danier, Ayan Das, Stathi Fotiadis, Farhang Nabiei, Ushnish Sengupta, and Alberto Bernacchia. Score Normalization for a Faster Diffusion Exponential Integrator Sampler. In NeurIPS 2023 Workshop on Diffusion Models, 2023. 4",
+            "[63] Guoxuan Xia, Olivier Laurent, Gianni Franchi, and Christos-Savvas Bouganis. Towards understanding why label smoothing degrades selective classification and how to fix it, 2024. 3",
+            "[64] Yingda Xia, Yi Zhang, Fengze Liu, Wei Shen, and Alan L Yuille. Synthesize then compare: Detecting failures and anomalies for semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 145-161. Springer, 2020. 3",
+            "[65] Yilun Xu, Mingyang Deng, Xiang Cheng, Yonglong Tian, Ziming Liu, and Tommi S. Jaakkola. Restart sampling for improving generative processes. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 4",
+            "[66] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, 2024. 3",
+            "[67] Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roeah Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. What you see is what you read? improving text-image alignment evaluation. Advances in Neural Information Processing Systems, 36:1601-1619, 2023. 2",
+            "[68] Xuanwu Yin Yuda Song, Zehao Sun. Sdxs: Real-time one-step latent diffusion models with image conditions. arxiv, 2024. 6, 2",
+            "[69] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. In The Eleventh"
+        ],
+        "bbox": [
+            91,
+            92,
+            480,
+            900
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "International Conference on Learning Representations, 2023.4",
+            "[70] 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. 4",
+            "[71] 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. 1",
+            "[72] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 6, 1",
+            "[73] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now, 2024. 7, 3"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            359
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "8072",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/c3ad3744-8ee2-4117-a4a1-7d23c58e3fe4_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7948aacacc50c00553efe146feebba3aa782fd6102d0378a179d6e9822c9079a
+size 4573219

2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/full.md

@@ -0,0 +1,310 @@
+# Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation
+
+Gianni Franchi, $^{1}$ * Nacim Belkhir, $^{2}$ * Dat Nguyen Trong, $^{1}$  Guoxuan Xia, $^{3}$  Andrea Pilzer $^{4}$   
+U2IS, ENSTA Paris, Institut Polytechnique de Paris,  $^{1}$ Mirai, $^{2}$  Imperial College London, $^{3}$  NVIDIA $^{4}$
+
+![](images/0df14e7c0fe8aff5150a47ddde96930d88820f306268a239773e70e5602c43fa.jpg)  
+Figure 1. Examples of Applications for Uncertainty Quantification in Text-to-Image Generation. Text-to-image generation models may exhibit uncertainty, and that need to be quantified since it can provide insights into the model's training dataset, aiding in deepfake prevention, detecting model biases, and protecting copyrighted content from unauthorized generation.
+
+![](images/e4cb42dfba865827dd7ac22e034ddfa0a402644ec557b371ebb37dce40429cc8.jpg)
+
+# Abstract
+
+Uncertainty quantification in text-to-image (T2I) generative models is crucial for understanding model behavior and improving output reliability. In this paper, we are the first to quantify and evaluate the uncertainty of T2I models with respect to the prompt. Alongside adapting existing approaches designed to measure uncertainty in the image space, we also introduce Prompt-based Uncertainty Estimation for T2I models (PUNC), a novel method leveraging Large Vision-Language Models (LVLMs) to better address uncertainties arising from the semantics of the prompt and generated images. PUNC utilizes a LVLM to caption a gen
+
+erated image, and then compares the caption with the original prompt in the more semantically meaningful text space. PUNC also enables the disentanglement of both aleatoric and epistemic uncertainties via precision and recall, which image-space approaches are unable to do. Extensive experiments demonstrate that PUNC outperforms state-of-the-art uncertainty estimation techniques across various settings. Uncertainty quantification in text-to-image generation models can be used on various applications including bias detection, copyright protection, and OOD detection. We also introduce a comprehensive dataset of text prompts and generation pairs to foster further research in uncertainty quantification for generative models. Our findings illustrate that PUNC not only achieves competitive performance but also enables novel applications in evaluating
+
+and improving the trustworthiness of text-to-image models. The code is available at https://github.com/ENSTA-U2IS-AI/Uncertainty_diffusion
+
+# 1. Introduction
+
+Generative models have revolutionized artificial intelligence, making significant strides in fields such as robotics [2], healthcare [28], and physics [33]. These models have elevated the performance of AI systems by enabling the generation of complex data representations. However, while generative models offer immense potential, they do not address all challenges, especially those related to robustness and uncertainty in Deep Neural Networks(DNNs)—issues that have become even more pronounced in generative settings.
+
+In this paper, we focus on the uncertainty quantification in deep generative models, in particular, we are the first to investigate modern text-to-image (T2I) models where the input is a natural language prompt [8, 18, 48-50]. These models have surged in popularity in recent years due to their capacity to generate high-quality images conveniently in a user-guided fashion, finding usage in a wide array of real-world applications. We are the first to investigate and propose a dedicated approach for quantifying uncertainty in text-to-image generation, aiming to fill a critical gap in generative AI research. The task of quantifying uncertainty in these models is not only complex but also crucial. Understanding the uncertainty in generative models can be invaluable for applications such as Out-of-Distribution (OOD) detection, trustworthy AI, and decision-making in high-stakes domains. To the best of our knowledge, no existing work has comprehensively addressed the challenge of uncertainty quantification in text-to-image generative models. However, given the semantic nature of the uncertainty we aim to measure, our work is closely related to recent studies on text-to-image alignment [37, 52, 67].
+
+Existing research explores uncertainty estimation and OOD detection in the image/output space [4, 5, 23, 39]. For example, given a test image, can we determine whether or not it is OOD? On the other hand, the uncertainty of a generation with respect to conditioning (such as text) is underexplored. We argue that this is an important area to be addressed, as it represents a widespread real-world application of deep generative models. In this case, it remains an open question as to how should we define uncertainty, and given a definition, how should we quantify it? As illustrated in Fig. 1, focusing on uncertainty quantification in text-to-image generation could enable novel applications, allowing us to extract and interpret the knowledge embedded within models using language.
+
+To answer these questions, we first argue that uncertainties should be rooted in the semantics of the image
+
+and the prompt. We then adapt and evaluate a number of existing image-space approaches, as well as proposing a novel method for uncertainty quantification in T2I generative models, which we refer to as "Prompt-based Uncertainty Estimation for T2I Generation" (PUNC). Our approach leverages the growing capabilities of Large Vision-Language Models (LVLMs) to extract the semantics from generated images. LVLMs, which have been trained on vast amounts of text and visual data, can serve as powerful tools for interpreting the underlying meaning of text/image inputs and assessing how variations in phrasing or semantics impact the uncertainty of the generated outputs. Our hypothesis is that as LVLMs become increasingly adept at understanding complex relationships between text and images, they will allow us to better analyze and quantify uncertainty. Through this paper, we will demonstrate Prompt-based Uncertainty Estimation for T2I (PUNC) effectiveness in quantifying uncertainties in generative models. Showcasing its ability to decompose and interpret different types of uncertainties arising from the multimodal nature of the input, offering a new perspective on uncertainty quantification in generative AI.
+
+To summarize, our contributions are as follows: (1) We introduce a new task aimed at quantifying uncertainty (with respect to the prompt) in text-to-image generation models, and we share a dataset of prompts to support future research on uncertainty in these models. (2) We both adapt existing image-space approaches to our tasks, as well as propose a novel straightforward method, PUNC, designed to better quantify semantic uncertainty in text-to-image generation. (3) Extensive experiments demonstrate that PUNC, despite its simplicity and computational efficiency, achieves competitive performance. (4) We highlight the practical utility of uncertainty quantification in text-to-image generation by showcasing additional applications where this approach could be beneficial.
+
+# 2. Related Work
+
+Deep learning models typically encounter two primary types of uncertainty: aleatoric and epistemic uncertainty [26]. Over the years, a variety of methods have been developed to quantify these uncertainties. Ensemble methods [16, 32, 34, 58, 61], for instance, have often achieved state-of-the-art performance by leveraging multiple models to assess prediction confidence. However, these methods tend to be computationally expensive due to the need to train and maintain several models. Another approach involves Bayesian methods [21], especially Bayesian Neural Networks (BNNs) [19, 27, 42, 57], which offer a theoretically grounded way to quantify uncertainty. Additionally, there are techniques aimed at directly predicting uncertainty within DNNs [10, 20, 30]. Despite their effectiveness, most of these methods are challenging to apply to text-to-image
+
+![](images/f22022d38cc577c40f1d07107d04862ce1c9139eb16d6a4dce45345b8c7c4a8d.jpg)  
+Generation/Image-space Uncertainty (existing work)  
+Figure 2. Diagram illustrating generation/image-space uncertainty which is considered in previous work and condition/prompt-space uncertainty which is investigated in our work.
+
+![](images/61e51754cec18673e6e57272b130161e06703729facb467446902ddcc7dc5863.jpg)  
+Figure 3. Three generations from PixArt- $\Sigma$  [7] illustrating uncertainty with regards to prompt semantics. For the Normal image, we used an ImageNet-inspired prompt; for the corrupted image, additional corruption was applied to the prompt (e.g. fish was perturbed to fis) increasing the aleatoric uncertainty; and for the out-of-distribution (OOD) case, the model was prompted to generate an image of the Prime Minister of Japan Kishida Fumio under similar conditions, where epistemic uncertainty was injected in the form of a semantic concept the model was not familiar with.
+
+generation, as they rely on ensembling predictions or accessing model uncertainty in ways that are difficult to implement for generative models without explicit access to intermediate predictions.
+
+Generative models, as a subset of deep learning, also exhibit uncertainty. In the past, traditional generative models were employed for uncertainty quantification and used to help estimate the uncertainty within standard DNNs [6, 38, 56, 64]. With the recent popularity of diffusion models, research has begun exploring whether image-based diffusion models can detect out-of-distribution (OOD) samples [4, 23, 39] or quantify their own uncertainty [5, 6, 14]. Whilst these approaches align with our goals, they are limited to image-space tasks, where the input is an image rather than a textual prompt. This input difference fundamentally alters the nature of uncertainty in the model. Our approach, in contrast, specifically addresses text-to-image generation, a setting where prompt ambiguity and multimodal complexity introduce new dimensions of uncertainty. We demonstrate that quantifying uncertainty in this context can lead to valuable insights and applications across various downstream tasks.
+
+# 3. Uncertainty in Image Generation Models
+
+Consider a deep generative model  $p_{\theta}(\boldsymbol{x})$  with parameters  $\theta$  that models the data distribution  $p_{\mathrm{data}}(\boldsymbol{x})$  of  $\boldsymbol{x} \in \mathbb{R}^D$ . We may be interested in measuring the uncertainty of our model  $p_{\theta}$  with respect to some test datum  $\boldsymbol{x}^*$ . For example, consider the case where we want to determine whether or not  $\boldsymbol{x}^*$  was drawn from  $p_{\mathrm{data}}(\boldsymbol{x})$ , i.e. out-of-distribution (OOD) detection. This can be achieved by quantifying the uncertainty of  $p_{\theta}$  that arises from a lack of knowledge of  $\boldsymbol{x}^*$ , i.e. epistemic uncertainty [26]. A range of existing work that leverages diffusion-based generative models for OOD image detection [23, 39] fall into the above paradigm. From hereon we will refer to this as generation/image-space uncertainty. This is illustrated in Fig. 2 (top).
+
+In this work, however, we are primarily interested in a different scenario: text-to-image generation. Consider instead a conditional model  $p_{\theta}(\pmb{x}|\pmb{c})$  that models the conditional data distribution  $p_{\mathrm{data}}(\pmb{x}|\pmb{c})$ . This describes real-world text-to-image generation, where image  $\pmb{x}$  is drawn from  $p_{\theta}(\pmb{x}|\pmb{c}^*)$  conditioned on some test text prompt  $\pmb{c}^*$ . In this case, we wish to quantify the uncertainty of  $p_{\theta}$  not with respect to some test  $\pmb{x}^*$  (e.g. image) but instead with respect to a test condition  $\pmb{c}^*$ . We will hereon refer to this as condition/prompt-space uncertainty. This is illustrated in Fig. 2 (bottom).
+
+We argue that in this case, uncertainties should be concerned with semantics, as intuitively, these are what inherently link the user's intentions, represented by the text prompt, and the generated output image. This is similar to recent discourse on uncertainty estimation for large language models [31]. If we consider the semantic concepts in the prompt and generated image, then we can define useful concepts of uncertainty (illustrated in Fig. 3):
+
+- Aleatoric uncertainty is irreducible uncertainty in the data distribution  $p_{\mathrm{data}}(\boldsymbol{x}|\boldsymbol{c})$ . From the perspective of semantics, high aleatoric uncertainty would be where a variation of generated concepts may arise from a single prompt. For example, a spelling mistake where fish is mistyped as fis in the prompt may result in either a "fish" or a "fist" appearing in the image.  
+- Epistemic uncertainty should correspond to a model's lack of knowledge of semantic concepts found in a prompt. For example, a model trained on ImageNet will not know what the prime minister of Japan Kishida Fumio looks like, and thus will have high epistemic uncertainty for this semantic concept.
+
+We note that broadly speaking uncertainty estimation for deep conditional models is well explored [20-22, 24, 26, 32, 34, 41, 59-61, 63, 66]. However, research in this area is generally concerned with problem settings where 1) the
+
+models explicitly evaluate  $p_{\theta}(\pmb{x}|c)$  and 2) downstream applications of uncertainty relate directly to the numerical values of  $\pmb{x}$  and  $p_{\theta}(\pmb{x}|c)$ . For example, a cross-entropy-trained image classifier explicitly predicts the probability mass  $P_{\theta}(\pmb{x}|c^{*})$  of label  $\pmb{x}$  given test image  $c^{*}$ .  $P_{\theta}(\pmb{x}|c^{*})$  directly relates to the likelihood of classification error – which is an event that we may want to detect using an uncertainty measure, i.e. downstream misclassification detection [22, 24]. Thus, we are able to easily extract useful measures of uncertainty from the model  $P_{\theta}$ , e.g. the maximum of the softmax.
+
+On the other hand, text-to-image generation models are not as clear-cut. They may not explicitly evaluate  $p_{\theta}(\pmb{x}|c^{*})$  (e.g. diffusion-based models [25, 54]), and even when they do (e.g. vector-quantised autoregressive models [17, 55]) it is hard to directly relate (latent) pixel values directly to our above definitions of uncertainty. For example, simple variations in contrast and brightness that may leave semantics completely unchanged could result in a very high variance in  $p_{\theta}(\pmb{x}|c^{*})$ . To give a simple example, inverting the black and white stripes on a zebra would lead to a large Euclidean distance in the image space whilst preserving the semantics of the image. In the next section, we present and discuss a range of potential methods to quantify the uncertainty of text-to-image models with respect to a given prompt.
+
+# 4. Techniques to Quantify Uncertainty
+
+We quantify text-to-image uncertainty by adapting existing methods and proposing a novel approach named Prompt-based UNCertainty estimation (PUNC)(Fig. 4).
+
+# 4.1. Adapting Existing Approaches
+
+There are a number of existing methods for quantifying the image-space uncertainty of diffusion models. As diffusion-based models represent a majority of current text-to-image foundation models [8, 18, 48, 50], we choose to adapt these in order to quantify prompt-space uncertainty (Fig. 4). We first present a brief primer on diffusion models to the reader.
+
+# 4.1.1. Diffusion Models
+
+Diffusion models [25] generate data by iteratively denoising pure noise. The forward process over time  $t \in [0,1]$  for  $\boldsymbol{x}_t \in \mathbb{R}^D$ , conditioned on a text prompt  $c$ , follows:
+
+$$
+p \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {0}, \boldsymbol {c}\right) = \mathcal {N} \left(\boldsymbol {x} _ {t}; a _ {t} \boldsymbol {x} _ {0}, \sigma_ {t} ^ {2} \boldsymbol {I}\right), \tag {1}
+$$
+
+where  $\pmb{x}_0$  is sampled from an unknown data distribution  $p(\pmb{x}_0)$ , and  $a_{t},\sigma_{t}$  define the noise schedule. This formulation allows diffusion models to progressively refine noisy inputs, enabling high-quality data generation through a learned denoising process. By leveraging stochastic differential equations (SDEs) or their deterministic counterparts,
+
+these models sample from complex distributions in a structured manner. Various solvers [29, 40, 53, 54, 62, 65, 69] can then be employed to efficiently generate realistic samples with a limited number of denoising steps.
+
+Latent diffusion models (LDMs) [39] perform this process within a compressed latent space rather than directly on pixel space. By first encoding images into a more compact representation, LDMs reduce computational demands, enabling high-quality high-resolution generations. We will not consider the latent representation to keep the notation simple, even though experiments will use LDMs.
+
+# 4.1.2. Time-step based approaches
+
+DDPM-OOD [23]. This approach, developed for image-space OOD detection, involves noising a test sample  $\boldsymbol{x}_0^*$  to different points in the forward process to get  $\{\boldsymbol{x}_{t_1}^*,\boldsymbol{x}_{t_2}^*,\dots \}$  and denoising them, generating multiple reconstructions  $\{\hat{x}_{0_1},\hat{x}_{0_2},\ldots \}$ . Similarity is then measured between each reconstruction and the original image. The idea is that test samples  $\boldsymbol{x}_0^*$  for which the model has higher epistemic uncertainty will lead to reconstructions that differ from the test sample. However, in our case, we do not have a test image  $\boldsymbol{x}_0^*$ . In order to adapt DDPM-OOD we propose measuring similarity between reconstructions at different  $t$  and a baseline generation,  $S(\hat{x}_{0_i},\hat{x}_0)$ , where  $\hat{x}_0$  represents an image directly generated using the original test prompt  $c^*$  alone. Both aleatoric uncertainty (e.g. ambiguous prompt) and epistemic uncertainty (e.g. unfamiliar concepts in prompt) could lead to more divergent reconstructions.
+
+LMD [39]. Another similar approach for image-space OOD detection, known as Learned Manifold Denoising (LMD), instead explores alternative transforms on  $\boldsymbol{x}_0^*$  to noising. The idea is to lift  $\boldsymbol{x}_0^*$  from its original data manifold and then map it back using the trained diffusion model. If the sample is in-distribution, the mapping process should return it close to its original position on the data manifold. Conversely, if it is OOD, the mapped position will likely diverge significantly from the original. The corruption process typically involves masking portions of the image, with various mask patterns applied. Similar to DDPM-OOD, similarity is measured between the original image and the reconstructed version after applying the diffusion process. Similar to before, since we lack a testimage, we measure the similarity between reconstructed images and  $\hat{\boldsymbol{x}}_{0_i}$ , the image generated directly from the prompt without perturbations.
+
+Similarity Metric. Various methods are used to evaluate similarity between images. Mean-squared error (MSE) is commonly applied between input and reconstruction, while some approaches favor the Learned Perceptual Image Patch Similarity (LPIPS) [70] metric. The choice of similarity metric is a crucial component of these techniques.
+
+![](images/028fd6b0c4c53b8460bc355b984d6d2584ea5339d9111611314b3422a07d66e0.jpg)  
+Figure 4. Illustration showing the different baselines and PUNC. PUNC leverages a LVLM to describe generated images and assess similarity with the original prompt, providing a refined uncertainty score. In contrast, baseline methods employ traditional techniques such as noise injection, ensembling, or masking to quantify uncertainty, followed by image-based similarity scoring.
+
+# 4.1.3. Test-Time Ensembling Approaches
+
+Test-time ensembling techniques take inspiration from approaches used in the large language model (LLM) community [31], where multiple generations are produced to assess consistency. Given a prompt  $c$ , we generate two images  $\{\hat{x}_1, \hat{x}_2\}$  and measure the similarity between the two outputs. We denote this technique 2XDM. The underlying rationale is that if the model has high confidence, it will generate similar images. However, if uncertainty is high, it will generate substantially different images. We note that this is similar to the approach proposed in [5], however, that approach uses an ensemble of diffusion models, which is not practical in the setting where we want to quantify the uncertainty of a single pretrained T2I model.
+
+A key question remains regarding how to find the best measure similarity, and we propose using the same techniques applied in the time-step-based methods. A significant advantage of test-time ensembling approaches is that they can be applied generically to all T2I models. In contrast, timestep-based approaches require a multi-step generation process, limiting the broadness of their applicability. We note that all above approaches require multiple generations, incurring a notable computational overhead at inference time. For example, DDPM-OOD requires  $\sim 50\times$  the number of network inferences as a single generation (if the hyperparameters of the original paper are used) [23].
+
+# 4.2. Prompt-based Uncertainty Estimation
+
+In this section, we propose a new method, Prompt-based Uncertainty estimation for text-to-image generation (PUNC), which leverages Large Vision-Language Models (LVLM) in conjunction with T2I models. We aim to assess the uncertainty in generated images by analyzing the alignment between initial text prompts and reconstructed textual descriptions of the generated images.
+
+# Background on Large Vision-Language Model (LVLM)
+
+A LVLM integrates a large language model (LLM) with
+
+an image encoder, allowing for robust, multimodal understanding. Given a text prompt  $c$  and an image  $x$ , a LVLM model, with parameters  $\omega$ , includes:
+
+- Image Encoder:  $f_{\omega}^{\mathrm{img}}(\cdot)$  processes the input image  $x$  and produces an embedding:  $z^{\mathrm{img}} = f_{\omega}^{\mathrm{img}}(x)$  
+- Text Processor with LLM:  $f_{\omega}^{\mathrm{txt}}(\cdot, \cdot)$  takes the prompt  $c$  and the image embedding  $z^{\mathrm{img}}$  as inputs, generating a descriptive or answer-based response:  $\hat{c} = f_{\omega}^{\mathrm{txt}}(c, z^{\mathrm{img}})$
+
+Thus, a LVLM system generates an interpretation  $\hat{c}$  that reflects the content of the image given an initial prompt.
+
+Prompt-Based Uncertainty Estimation for T2I Generation (PUNC) The previously introduced approaches to uncertainty estimation rely on measuring similarity between images. For a given text prompt  $c$ , a T2I model generates an image  $x \sim p_{\theta}(x|c)$ . We reason that, even under low uncertainty, the model can produce images that vary in aspects such as color, texture, or object positioning, while still faithfully representing the semantics of prompt  $c$ . As previously discussed in Sec. 3, this may result in image-based similarity metrics failing to capture uncertainty at a semantic level, since large differences in the image space may not necessarily correspond to meaningful semantic differences.
+
+We propose that advanced LVLMs, such as Molmo [11], LLAMA 3 [15], and GPT-4 [1], are a strong candidate for extracting the semantics from an image, through describing an image via natural language. Building on this insight, our proposed method, PUNC, introduces a novel approach that bypasses direct image comparison and instead evaluates the alignment between generated images and the original prompts by calculating an uncertainty score in the more semantically meaningful text-space. The steps for implementing PUNC are as follows:
+
+Step 1: Initial Prompt and Image Generation Given a test prompt  $c^*$ , we use the T2I model to sample an image  $x$ :
+
+$$
+\boldsymbol {x} \sim p _ {\boldsymbol {\theta}} (\boldsymbol {x} | \boldsymbol {c} ^ {*})
+$$
+
+Step 2: Image Interpretation via LVLM With the
+
+![](images/bfd49245a54fb630faa3175b2d199995fca0356c85229b1aa37a4ac1798f6ca0.jpg)  
+Figure 5. Disentangling Aleatoric and Epistemic Uncertainty.
+
+generated image  $\pmb{x}$  and the initial prompt  $c^*$ , the LVLM produces a new descriptive caption  $\hat{c}$ :
+
+$$
+\hat {\boldsymbol {c}} = f _ {\omega} ^ {\mathrm {t x t}} \left(\boldsymbol {c} ^ {*}, f _ {\omega} ^ {\mathrm {i m g}} (\boldsymbol {x})\right)
+$$
+
+Step 3: Uncertainty Score Calculation The core of PUNC lies in measuring the alignment between  $\mathbf{c}$  and  $\hat{\mathbf{c}}$  to assess the diffusion model's confidence. Specifically, the similarity score  $S(\mathbf{c},\hat{\mathbf{c}})$  is computed as:
+
+$$
+S (\boldsymbol {c} ^ {*}, \hat {\boldsymbol {c}}) = \operatorname {s i m} (\boldsymbol {c} ^ {*}, \hat {\boldsymbol {c}})
+$$
+
+High similarity suggests low uncertainty, indicating the generated image closely matches the initial prompt. Conversely, a lower score reveals higher uncertainty, signifying potential divergence from the prompt's intended meaning.
+
+Thus, PUNC is able to predict semantic uncertainties by utilising the powerful image understanding ability of LVLMs and computing similarities in the text space.
+
+Prompt Similarity: Aleatoric and Epistemic via Precision and Recall. In Sec. 3 we discussed the different types of semantic uncertainty for T2I generation. Here we will present a method for disentangling these uncertainties using the precision and recall of the caption given the prompt. Consider an abstract scenario where the prompt and the generated image each contain a discrete set of semantic concepts. We can then calculate the precision and recall of the image concepts with respect to the prompt:
+
+Intuitively, a lack of knowledge about concepts in the prompt (epistemic uncertainty) will result in fewer concepts being preserved in the image, reducing recall. On the other hand, a lack of specificity in the prompt (aleatoric uncertainty) will result in additional concepts being generated, reducing precision. Thus we can use this abstract form of semantic-concept-based recall to measure epistemic uncertainty, and use precision to measure aleatoric uncertainty.
+
+Although we cannot directly measure concepts, we can approximate them using natural language. Moreover, existing similarity scores for text already have the above notions of precision and recall integrated within [36, 43, 72], allowing PUNC to disentangle uncertainty by comparing the caption of the generated image with the original prompt. We note that the image-based approaches in
+
+Sec. 4.1 are unable to disentangle aleatoric from epistemic uncertainty. We focus on two different existing text-based similarity metrics, ROUGE [36], and BERTScore [72]. The former calculates similarity (precision or recall) by directly matching (sequences of) words between two text sequences. The latter on the other hand measures similarity in the deep feature space of a text encoder.
+
+As illustrated in Fig. 5, we define Prompt Similarity as the semantic overlap between the input prompt and the generated descriptive caption. More formally, let  $P_{\mathrm{GT}}$  denote the set of semantic concepts in the ground-truth prompt used to generate an image, and  $P_{\mathrm{VQA}}$  the set of concepts extracted from the LVLM-generated caption. The resulting uncertainties can be characterized as follows:
+
+Epistemic Uncertainty, which captures missing concepts in  $P_{\mathrm{VQA}}$  that are present in  $P_{\mathrm{GT}}$ , reflects an incomplete understanding of the intended semantics. It is defined as:
+
+$$
+U _ {\text {e p i s t e m i c}} = 1 - \frac {\left| P _ {\mathrm {G T}} \cap P _ {\mathrm {V Q A}} \right|}{\left| P _ {\mathrm {G T}} \right|} = 1 - \text {R e c a l l} \tag {2}
+$$
+
+Aleatoric Uncertainty, on the other hand, arises from spurious or extra concepts in  $P_{\mathrm{VQA}}$  that do not appear in  $P_{\mathrm{GT}}$ , indicating ambiguity in the model's response. It is defined as:
+
+$$
+U _ {\text {a l e a t o r i c}} = 1 - \frac {\left| P _ {\mathrm {G T}} \cap P _ {\mathrm {V Q A}} \right|}{\left| P _ {\mathrm {V Q A}} \right|} = 1 - \text {P r e c i s i o n} \tag {3}
+$$
+
+Further details on the computation of precision and recall can be found in Appendix A.
+
+# 5. Experiments & Discussions
+
+We test PUNC on Various task and various dataset and compared PUNC to DDPM-OOD [23], LMD [39] and 2XDM. We first introduce on Sec. 5 how we have constitute the dataset of prompts. Then on Sec. 5.1 we discuss of different results on these dataset. Then Appendix F explains how we can use the uncertainty of text-to-image model to detect if a model has been trained with some concept related to some fictional character of politicians. Appendix F.3 shows how uncertainty can help to detect bias of T2I models. To conduct our experiments and evaluate the various techniques for quantifying uncertainty, we utilized four T2I models: Stable Diffusion 1.5 (SDv1.5) [50], SDXS [68], SDXL [48], and PixArt-  $\Sigma$  [7]. These diverse T2I models allow us to test a range of architectures, each with unique characteristics and potential weaknesses. SDv1.5, being one of the earliest models, likely has a smaller and less diverse training dataset compared to the others. In contrast, SDXS generates in a single-step, different to the multi-step processes in other models.
+
+Dataset of prompts To conduct our experiments, we require both an in-distribution dataset of prompts and
+
+datasets with uncertainty. For the in-distribution prompts, we use the dataset proposed by [13], where high-quality descriptions of ImageNet [12] images are generated using GPT-4[13], referred to as Normal. To simulate epistemic uncertainty, we create out-of-distribution (OOD) datasets using images from underrepresented domains like remote sensing, texture, and microscopic datasets, captioned with LLAVA Next[35], labeled as Remote Sensing, Texture, and Microscopic. For aleatoric uncertainty, we generate two datasets: Vague, containing 2,000 minimally descriptive prompts (e.g., "An image of ***"), and Corrupted, derived from Normal by adding grammatical errors (Level 1) and removing  $50\%$  of words (Level 2). A detailed summary of these datasets is presented in Tab. 1 and Appendix E. In Appendix C, we provide the details of the hyperparameters used in our experiments.
+
+<table><tr><td>Dataset Type</td><td>Dataset Name</td><td>Description</td></tr><tr><td>In-Distribution (ID)</td><td>Normal</td><td>20k High-quality detailed prompts from ImageNet [12] images using GPT-4[13].</td></tr><tr><td>Out-of-Distribution (OOD)</td><td>Remote Sensing Texture Microscopic</td><td>20k Prompts from remote sensing images [47]5k Prompts from texture datasets10k Prompts from microscopic images [44]</td></tr><tr><td>Aleatoric Uncertainty</td><td>VagueAdversarialCorrupted</td><td>2k prompts with minimal context, e.g., &quot;An image of ***,&quot; where &quot;****&quot; is an ImageNet class name, creating ambiguous input.1k altered prompts from Normal using Unlearn-DiffAtk [73]20k Altered prompts from Normal:• Level 1: Grammatical mistakes and spelling errors.• Level 2: Random removal of 50% of words in Level 1 prompts.</td></tr></table>
+
+Table 1. Summary of prompt datasets for unc. quantification.
+
+# 5.1. Results for Uncertainty Estimation
+
+Tab. 3 presents the results of PUNC when using Molmo as the LVLM to assess epistemic uncertainty in the T2I models. Our approach, PUNC, outperforms most state-of-the-art techniques on average. However, we observe more mixed results on the Texture dataset, likely due to uncertainty about the presence of texture images in the training data of the T2I models.
+
+In Tab. 4, we show the results of PUNC under aleatoric uncertainty, again using Molmo as the LVLM. Our approach outperforms most state-of-the-art methods in this setting as well. Notably, while the ROUGE score tends to perform better for epistemic uncertainty, the optimal metric is less clear for aleatoric uncertainty. Here, the choice between ROUGE and BERTScore may depend on the specific application: because corruption introduces spelling and grammatical errors, BERTScore—which captures semantic similarity—may be more suitable. For additional results using different LVLMs, please refer to Appendix D and Appendix B. In Appendix G, we present qualitative results.
+
+# 5.2. Applications of uncertainty
+
+Our study explores multiple applications of uncertainty quantification in T2I models, with detailed results provided
+
+in Appendix F. First, we investigate deepfake detection by examining the model's ability to generate recognizable images of prominent politicians (see Appendix F.1). We examined whether diffusion models could generate convincing images of G7 politicians. Using 100 prompts containing their names, we generated images with four diffusion models and queried a LVLM to determine if the intended politician was present in each image (yes/no), thus measuring concept recognition uncertainty. Results are presented in Table 2.
+
+Next, we address copyright concerns by evaluating whether T2I models can accurately replicate iconic, copyrighted characters such as Mickey Mouse and Darth Vader (see Appendix F.2). Finally, we analyze bias in T2I models by testing their capacity to generate individuals across different genders and races for specific job roles (see Appendix F.3).
+
+Using a LVLM, we evaluate the presence of specific individuals, copyrighted characters, or demographic concepts in the generated images, demonstrating how uncertainty estimation can aid in detecting potential misuses of generative models. These experiments illustrate the value of text-based uncertainty quantification in analyzing T2I model performance across ethically and legally sensitive applications, enabling assessments that are challenging to achieve with image-based generation alone.
+
+# 6. Conclusion
+
+In this work, we set out to be first to investigate uncertainty estimation T2I generation, where the uncertainty is with respect to the text prompt. We both adapt existing image-space uncertainty approaches, as well as present Prompt-based UNCertainty Estimation for T2I models (PUNC), the first approach designed specifically to quantify uncertainty in T2I generation. By leveraging the interpretive power of Large Vision-Language Models to extract semantics from a generated image, PUNC addresses the challenges associated with measuring multimodal uncertainty, enabling a deeper understanding of model behavior with respect to prompt-based conditioning. Our results validate that PUNC provides reliable, fine-grained assessments of aleatoric and epistemic uncertainties, outperforming existing methods that work for image generation.
+
+PUNC opens avenues for new applications, such as improving model robustness in biased or OOD scenarios and safeguarding against unauthorized content generation, including deepfakes and copyrighted characters. Our experiments show that diffusion models, while powerful, still exhibit significant variances in their ability to adhere to prompt details, particularly with ambiguous, corrupt prompts or OOD prompts. By identifying such uncertainties, PUNC not only supports the development of more trustworthy generative AI but also aids in aligning these
+
+<table><tr><td></td><td>Justin Trudeau</td><td>King Charles</td><td>Fumio Kishida</td><td>Olaf Scholz</td><td>Emmanuel Macron</td><td>Joe Biden</td><td>Giorgia Meloni</td><td>AVERAGE</td></tr><tr><td>SDXL</td><td>94</td><td>94</td><td>88</td><td>88</td><td>98</td><td>96</td><td>21</td><td>95.5</td></tr><tr><td>SDv1.5</td><td>78</td><td>60</td><td>62</td><td>88</td><td>94</td><td>80</td><td>18</td><td>78</td></tr><tr><td>SDXS</td><td>80</td><td>40</td><td>70</td><td>82</td><td>94</td><td>76</td><td>28</td><td>72.5</td></tr><tr><td>PixArt</td><td>36</td><td>76</td><td>62</td><td>78</td><td>94</td><td>90</td><td>15</td><td>74</td></tr><tr><td>AVERAGE</td><td>72</td><td>67.5</td><td>70.5</td><td>84</td><td>95</td><td>85.5</td><td>20.5</td><td>70.71</td></tr></table>
+
+Table 2. Accuracy (%) of diffusion models in generating recognizable images of G7 politicians. Higher accuracy indicates lower uncertainty in the model's ability to represent specific individuals.  
+
+<table><tr><td colspan="2"></td><td>auroc ↑</td><td>SDXS aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>PixArt aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>SDv1.5 aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>SDXL aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>Average aupr ↑</td><td>fpr95↓</td></tr><tr><td rowspan="6">microscopic</td><td>DDPM-OOD mse_and_lpips</td><td></td><td></td><td></td><td>39.61%</td><td>56.38%</td><td>100.00%</td><td>71.64%</td><td>71.52%</td><td>93.62%</td><td>87.25%</td><td>83.67%</td><td>51.43%</td><td>66.17%</td><td>70.53%</td><td>81.68%</td></tr><tr><td>LMD lpips alex</td><td></td><td></td><td></td><td>45.82%</td><td>45.24%</td><td>99.10%</td><td>16.81%</td><td>31.61%</td><td>99.88%</td><td>5.75%</td><td>28.12%</td><td>100.00%</td><td>22.79%</td><td>34.99%</td><td>99.66%</td></tr><tr><td>2XDM lpips alex</td><td>24.49%</td><td>32.60%</td><td>99.91%</td><td>24.32%</td><td>33.21%</td><td>99.99%</td><td>16.19%</td><td>32.42%</td><td>99.94%</td><td>30.58%</td><td>36.99%</td><td>98.25%</td><td>23.90%</td><td>33.81%</td><td>99.52%</td></tr><tr><td>PUNC ROUGE (1_recall)</td><td>84.00%</td><td>85.10%</td><td>57.07%</td><td>95.47%</td><td>97.06%</td><td>30.84%</td><td>87.24%</td><td>90.66%</td><td>48.90%</td><td>83.56%</td><td>84.31%</td><td>52.78%</td><td>87.57%</td><td>89.28%</td><td>47.40%</td></tr><tr><td>PUNC ROUGE (L_recall)</td><td>87.64%</td><td>89.56%</td><td>55.28%</td><td>96.96%</td><td>98.02%</td><td>16.85%</td><td>90.20%</td><td>93.40%</td><td>43.56%</td><td>88.19%</td><td>89.84%</td><td>47.44%</td><td>90.75%</td><td>92.71%</td><td>40.78%</td></tr><tr><td>PUNC BERT (recall)</td><td>88.45%</td><td>89.71%</td><td>48.68%</td><td>97.83%</td><td>98.41%</td><td>10.32%</td><td>87.75%</td><td>91.70%</td><td>51.03%</td><td>83.10%</td><td>85.46%</td><td>58.14%</td><td>89.28%</td><td>91.32%</td><td>42.04%</td></tr><tr><td rowspan="6">remote sensing</td><td>DDPM-OOD mse_and_lpips</td><td></td><td></td><td></td><td>80.00%</td><td>85.37%</td><td>99.28%</td><td>74.36%</td><td>71.89%</td><td>71.00%</td><td>65.47%</td><td>58.25%</td><td>71.97%</td><td>73.28%</td><td>71.84%</td><td>80.75%</td></tr><tr><td>LMD lpips alex</td><td></td><td></td><td></td><td>90.65%</td><td>93.11%</td><td>74.49%</td><td>78.93%</td><td>79.31%</td><td>81.05%</td><td>86.60%</td><td>85.39%</td><td>49.19%</td><td>85.39%</td><td>85.94%</td><td>68.24%</td></tr><tr><td>2XDM lpips alex</td><td>53.88%</td><td>59.21%</td><td>98.23%</td><td>54.43%</td><td>57.99%</td><td>98.78%</td><td>30.66%</td><td>38.80%</td><td>98.76%</td><td>23.54%</td><td>35.35%</td><td>96.23%</td><td>40.63%</td><td>47.83%</td><td>98.00%</td></tr><tr><td>PUNC ROUGE (1_recall)</td><td>85.59%</td><td>86.59%</td><td>63.87%</td><td>94.14%</td><td>95.97%</td><td>49.29%</td><td>73.24%</td><td>72.93%</td><td>79.56%</td><td>73.07%</td><td>74.54%</td><td>83.48%</td><td>81.51%</td><td>82.51%</td><td>69.05%</td></tr><tr><td>PUNC ROUGE (L_recall)</td><td>89.85%</td><td>90.63%</td><td>51.96%</td><td>97.01%</td><td>97.94%</td><td>16.08%</td><td>83.63%</td><td>84.65%</td><td>65.35%</td><td>85.68%</td><td>87.09%</td><td>64.00%</td><td>89.04%</td><td>90.08%</td><td>49.34%</td></tr><tr><td>PUNC BERT (recall)</td><td>80.50%</td><td>82.56%</td><td>75.19%</td><td>91.93%</td><td>93.08%</td><td>43.30%</td><td>61.27%</td><td>64.74%</td><td>90.08%</td><td>60.71%</td><td>65.32%</td><td>90.94%</td><td>73.60%</td><td>76.42%</td><td>74.88%</td></tr><tr><td rowspan="6">Texture</td><td>DDPM-OOD mse_and_lpips</td><td></td><td></td><td></td><td>73.22%</td><td>62.02%</td><td>89.24%</td><td>70.21%</td><td>55.97%</td><td>99.80%</td><td>74.24%</td><td>51.97%</td><td>84.71%</td><td>72.56%</td><td>56.65%</td><td>91.25%</td></tr><tr><td>LMD lpips alex</td><td></td><td></td><td></td><td>63.40%</td><td>50.47%</td><td>97.52%</td><td>26.74%</td><td>16.06%</td><td>99.89%</td><td>35.66%</td><td>23.22%</td><td>99.95%</td><td>41.93%</td><td>29.92%</td><td>99.12%</td></tr><tr><td>2XDM lpips alex</td><td>32.80%</td><td>20.84%</td><td>99.98%</td><td>37.21%</td><td>23.36%</td><td>99.37%</td><td>29.58%</td><td>24.27%</td><td>99.93%</td><td>28.87%</td><td>16.49%</td><td>98.88%</td><td>32.11%</td><td>21.24%</td><td>99.54%</td></tr><tr><td>PUNC ROUGE (1_recall)</td><td>40.77%</td><td>72.88%</td><td>95.54%</td><td>83.26%</td><td>94.80%</td><td>73.48%</td><td>46.61%</td><td>75.64%</td><td>90.87%</td><td>49.16%</td><td>78.04%</td><td>93.57%</td><td>54.95%</td><td>80.34%</td><td>88.36%</td></tr><tr><td>PUNC ROUGE (L_recall)</td><td>45.64%</td><td>76.90%</td><td>95.99%</td><td>84.10%</td><td>95.13%</td><td>74.21%</td><td>47.60%</td><td>77.70%</td><td>93.14%</td><td>50.71%</td><td>80.05%</td><td>94.35%</td><td>57.01%</td><td>82.44%</td><td>89.42%</td></tr><tr><td>PUNC BERT (recall)</td><td>49.11%</td><td>79.11%</td><td>94.65%</td><td>84.40%</td><td>95.14%</td><td>69.99%</td><td>47.60%</td><td>78.80%</td><td>93.85%</td><td>46.72%</td><td>79.20%</td><td>96.87%</td><td>56.96%</td><td>83.06%</td><td>88.84%</td></tr></table>
+
+Table 3. Performance comparison of PUNC applied to different Out of Distribution concepts and different text-to-image models  
+
+<table><tr><td colspan="2"></td><td>auroc ↑</td><td>SDXS aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>PixArt aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>SDv1.5 aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>SDXL aupr ↑</td><td>fpr95↓</td><td>auroc ↑</td><td>Average aupr ↑</td><td>fpr95↓</td></tr><tr><td rowspan="6">Vague</td><td>DDPM-OOD mse.and.Lpips</td><td></td><td></td><td></td><td>31.41%</td><td>6.06%</td><td>96.45%</td><td>50.06%</td><td>9.30%</td><td>96.52%</td><td>40.59%</td><td>5.36%</td><td>97.37%</td><td>40.69%</td><td>6.91%</td><td>96.78%</td></tr><tr><td>LMD lipps alex</td><td></td><td></td><td></td><td>39.26%</td><td>6.91%</td><td>96.35%</td><td>47.62%</td><td>9.54%</td><td>97.70%</td><td>40.77%</td><td>5.70%</td><td>97.95%</td><td>42.55%</td><td>7.38%</td><td>97.33%</td></tr><tr><td>2XDM lipps alex</td><td>51.28%</td><td>10.10%</td><td>94.35%</td><td>43.82%</td><td>8.57%</td><td>96.89%</td><td>50.97%</td><td>10.09%</td><td>95.96%</td><td>57.83%</td><td>10.01%</td><td>94.37%</td><td>50.98%</td><td>9.69%</td><td>95.39%</td></tr><tr><td>PUNC ROUGE (1_precision)</td><td>99.99%</td><td>100.00%</td><td>0.00%</td><td>100.00%</td><td>100.00%</td><td>0.00%</td><td>99.99%</td><td>100.00%</td><td>0.00%</td><td>99.99%</td><td>100.00%</td><td>0.00%</td><td>99.99%</td><td>100.00%</td><td>0.00%</td></tr><tr><td>PUNC ROUGE (L_precision)</td><td>99.99%</td><td>100.00%</td><td>0.00%</td><td>100.00%</td><td>100.00%</td><td>0.00%</td><td>99.99%</td><td>100.00%</td><td>0.00%</td><td>100.00%</td><td>100.00%</td><td>0.00%</td><td>100.00%</td><td>100.00%</td><td>0.00%</td></tr><tr><td>PUNC BERT (precision)</td><td>60.63%</td><td>92.24%</td><td>82.85%</td><td>67.02%</td><td>94.54%</td><td>83.10%</td><td>37.35%</td><td>86.51%</td><td>92.70%</td><td>70.85%</td><td>95.13%</td><td>78.00%</td><td>58.96%</td><td>92.11%</td><td>84.16%</td></tr><tr><td rowspan="6">Adversarial</td><td>DDPM-OOD mse.and.Lpips</td><td></td><td></td><td></td><td>39.54%</td><td>1.78%</td><td>95.72%</td><td>50.89%</td><td>2.54%</td><td>95.91%</td><td>38.97%</td><td>1.10%</td><td>98.38%</td><td>43.13%</td><td>1.81%</td><td>96.67%</td></tr><tr><td>LMD lipps alex</td><td></td><td></td><td></td><td>43.71%</td><td>1.95%</td><td>96.24%</td><td>53.29%</td><td>2.89%</td><td>94.43%</td><td>41.64%</td><td>1.17%</td><td>97.48%</td><td>46.21%</td><td>2.00%</td><td>96.05%</td></tr><tr><td>2XDM lipps alex</td><td>36.94%</td><td>1.94%</td><td>98.92%</td><td>47.95%</td><td>2.54%</td><td>96.03%</td><td>53.04%</td><td>2.83%</td><td>95.69%</td><td>55.19%</td><td>1.96%</td><td>96.20%</td><td>48.28%</td><td>2.32%</td><td>96.71%</td></tr><tr><td>PUNC ROUGE (1_precision)</td><td>64.14%</td><td>97.77%</td><td>48.80%</td><td>76.32%</td><td>98.90%</td><td>48.00%</td><td>63.59%</td><td>97.73%</td><td>49.50%</td><td>62.97%</td><td>97.64%</td><td>49.20%</td><td>66.76%</td><td>98.01%</td><td>48.87%</td></tr><tr><td>PUNC ROUGE (L_precision)</td><td>63.20%</td><td>97.75%</td><td>49.00%</td><td>76.26%</td><td>98.90%</td><td>48.40%</td><td>64.09%</td><td>97.90%</td><td>49.70%</td><td>63.76%</td><td>97.87%</td><td>49.60%</td><td>66.83%</td><td>98.11%</td><td>49.17%</td></tr><tr><td>PUNC BERT (precision)</td><td>73.98%</td><td>98.72%</td><td>78.31%</td><td>72.95%</td><td>98.93%</td><td>86.40%</td><td>45.09%</td><td>96.93%</td><td>94.18%</td><td>72.69%</td><td>98.80%</td><td>81.40%</td><td>66.18%</td><td>98.34%</td><td>85.07%</td></tr><tr><td rowspan="6">Corrupt l11</td><td>DDPM-OOD mse.and.Lpips</td><td></td><td></td><td></td><td>48.87%</td><td>44.82%</td><td>87.43%</td><td>53.20%</td><td>54.43%</td><td>95.48%</td><td>41.02%</td><td>29.54%</td><td>96.32%</td><td>47.70%</td><td>42.93%</td><td>93.08%</td></tr><tr><td>LMD lipps alex</td><td></td><td></td><td></td><td>51.23%</td><td>47.48%</td><td>89.99%</td><td>57.53%</td><td>56.49%</td><td>90.92%</td><td>40.94%</td><td>30.30%</td><td>96.32%</td><td>49.90%</td><td>44.76%</td><td>92.41%</td></tr><tr><td>2XDM lipps alex</td><td>25.48%</td><td>36.16%</td><td>99.29%</td><td>50.43%</td><td>49.86%</td><td>90.10%</td><td>54.97%</td><td>54.29%</td><td>95.69%</td><td>54.73%</td><td>38.27%</td><td>93.29%</td><td>46.40%</td><td>44.65%</td><td>94.59%</td></tr><tr><td>PUNC ROUGE (1_precision)</td><td>28.75%</td><td>38.34%</td><td>98.77%</td><td>54.58%</td><td>55.29%</td><td>96.02%</td><td>27.69%</td><td>37.57%</td><td>98.78%</td><td>27.45%</td><td>37.27%</td><td>98.54%</td><td>34.62%</td><td>42.12%</td><td>98.03%</td></tr><tr><td>PUNC ROUGE (L_precision)</td><td>27.29%</td><td>38.11%</td><td>99.15%</td><td>54.32%</td><td>55.59%</td><td>96.16%</td><td>27.30%</td><td>38.07%</td><td>99.45%</td><td>27.35%</td><td>37.89%</td><td>99.37%</td><td>34.07%</td><td>42.42%</td><td>98.53%</td></tr><tr><td>PUNC BERT (precision)</td><td>85.85%</td><td>84.45%</td><td>73.94%</td><td>79.43%</td><td>78.83%</td><td>87.21%</td><td>54.97%</td><td>53.73%</td><td>94.51%</td><td>75.76%</td><td>72.91%</td><td>84.65%</td><td>74.00%</td><td>72.48%</td><td>85.08%</td></tr><tr><td rowspan="6">Corrupt l12</td><td>DDPM-OOD lipps alex</td><td></td><td></td><td></td><td>43.23%</td><td>43.94%</td><td>91.23%</td><td>59.46%</td><td>60.14%</td><td>93.71%</td><td>48.79%</td><td>34.44%</td><td>95.74%</td><td>50.49%</td><td>46.17%</td><td>93.56%</td></tr><tr><td>LMD lipps alex</td><td></td><td></td><td></td><td>46.90%</td><td>45.65%</td><td>92.04%</td><td>59.94%</td><td>58.90%</td><td>90.54%</td><td>40.88%</td><td>30.42%</td><td>96.35%</td><td>49.24%</td><td>44.99%</td><td>92.98%</td></tr><tr><td>2XDM lipps alex</td><td>58.15%</td><td>58.50%</td><td>93.98%</td><td>52.08%</td><td>53.36%</td><td>90.19%</td><td>53.57%</td><td>54.17%</td><td>96.34%</td><td>60.24%</td><td>44.23%</td><td>91.78%</td><td>56.01%</td><td>52.57%</td><td>93.07%</td></tr><tr><td>PUNC ROUGE (1_precision)</td><td>28.76%</td><td>38.10%</td><td>98.59%</td><td>54.77%</td><td>55.49%</td><td>96.02%</td><td>27.66%</td><td>37.42%</td><td>98.75%</td><td>27.39%</td><td>37.23%</td><td>98.53%</td><td>34.64%</td><td>42.06%</td><td>97.97%</td></tr><tr><td>PUNC ROUGE (L_precision)</td><td>27.33%</td><td>37.87%</td><td>99.16%</td><td>54.51%</td><td>55.87%</td><td>96.11%</td><td>27.31%</td><td>37.92%</td><td>99.35%</td><td>27.33%</td><td>37.83%</td><td>99.36%</td><td>34.12%</td><td>42.37%</td><td>98.50%</td></tr><tr><td>PUNC BERT (precision)</td><td>85.96%</td><td>84.23%</td><td>73.54%</td><td>79.47%</td><td>78.80%</td><td>87.05%</td><td>54.78%</td><td>53.51%</td><td>94.79%</td><td>76.04%</td><td>73.44%</td><td>84.58%</td><td>74.06%</td><td>72.49%</td><td>84.99%</td></tr></table>
+
+Table 4. Performance comparison all applications (Vague, Adversarial, Corrupt Lvl1, Corrup Lvl2)
+
+models with ethical and regulatory standards.
+
+Our contributions also include a dataset specifically tailored to uncertainty quantification in text-to-image models, which we hope will serve as a benchmark for future research. Ultimately, this work highlights the potential of prompt-based uncertainty estimation as a critical tool for enhancing the reliability and safety of generative models in real-world applications.
+
+Limitations. While PUNC demonstrates promising results in quantifying uncertainty for text-to-image generation, several limitations remain. First, our approach
+
+relies heavily on the interpretive abilities of LVLMs, which, despite their effectiveness, may introduce biases or inaccuracies inherited from their training data. Second, our analysis focuses primarily on English-language prompts, limiting our method's applicability to multilingual or culturally diverse datasets.
+
+Acknowledgments. This work was performed using HPC resources from GENCI-IDRIS (Grant 2023-[AD011011970R3]), and EuroHPC Development access to LEONARDO, hosted by CINECA.
+
+# References
+
+[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 preprint arXiv:2303.08774, 2023. 5  
+[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2  
+[3] Nouar AlDahoul, Talal Rahwan, and Yasir Zaki. Aigenerated faces free from racial and gender stereotypes. arXiv preprint arXiv:2402.01002, 2024. 6  
+[4] Georges Le Bellier and Nicolas Audebert. Detecting Out-Of-Distribution Earth Observation Images with Diffusion Models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 481–491, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 2, 3  
+[5] Lucas Berry, Axel Brando, and David Meger. Shedding light on large generative networks: Estimating epistemic uncertainty in diffusion models. In The 40th Conference on Uncertainty in Artificial Intelligence, 2024. 2, 3, 5  
+[6] Matthew A Chan, Maria J Molina, and Christopher A Metzler. Hyper-diffusion: Estimating epistemic and aleatoric uncertainty with a single model. arXiv preprint arXiv:2402.03478, 2024. 3  
+[7] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart $\backslash$ sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 3, 6, 2  
+[8] Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\mathbb{S}\backslash \mathbb{alpha}\mathbb{S}$ : Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2, 4  
+[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 2, 3  
+[10] Charles Corbière, Nicolas Thome, Avner Bar-Hen, Matthieu Cord, and Patrick Pérez. Addressing failure prediction by learning model confidence. Advances in Neural Information Processing Systems, 32, 2019. 2  
+[11] Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 5, 2  
+[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7, 2
+
+[13] Ning Ding, Yehui Tang, Zhongqian Fu, Chao Xu, Kai Han, and Yunhe Wang. Gpt4image: Can large pre-trained models help vision models on perception tasks? arXiv e-prints, pages arXiv-2306, 2023. 7, 2  
+[14] Zhekai Du and Jingjing Li. Diffusion-based probabilistic uncertainty estimation for active domain adaptation. Advances in Neural Information Processing Systems, 36: 17129-17155, 2023. 3  
+[15] 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 preprint arXiv:2407.21783, 2024. 5, 2  
+[16] Nikita Durasov, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Masksembles for uncertainty estimation. In CVPR, 2021. 2  
+[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 4  
+[18] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In Proceedings of the 41st International Conference on Machine Learning, pages 12606-12633. PMLR, 2024. 2, 4  
+[19] Gianni Franchi, Andrei Bursuc, Emanuel Aldea, Séverine Dubuisson, and Isabelle Bloch. TRADI: Tracking deep neural network weight distributions. In ECCV, 2020. 2  
+[20] Gianni Franchi, Xuanlong Yu, Andrei Bursuc, Emanuel Aldea, Severine Dubuisson, and David Filliat. Latent discriminant deterministic uncertainty. In Computer Vision – ECCV 2022, pages 243–260, Cham, 2022. Springer Nature Switzerland. 2, 3  
+[21] Yarin Gal. Uncertainty in Deep Learning. PhD thesis, University of Cambridge, 2016. 2, 3  
+[22] Yonatan Geifman and Ran El-Yaniv. Selective classification for deep neural networks. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 4885-4894, Red Hook, NY, USA, 2017. Curran Associates Inc. 3, 4  
+[23] Mark S Graham, Walter HL Pinaya, Petru-Daniel Tudosiu, Parashkev Nachev, Sebastien Ourselin, and Jorge Cardoso. Denoising diffusion models for out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2948-2957, 2023. 2, 3, 4, 5, 6  
+[24] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2017. 3, 4  
+[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4
+
+[26] Eyke Hüllermeier and Willem Waegeman. *Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods*. *Machine Learning*, 110(3), 2021. 2, 3  
+[27] Pavel Izmailov, Sharad Vikram, Matthew D Hoffman, and Andrew Gordon Gordon Wilson. What are bayesian neural network posteriors really like? In ICML, 2021. 2  
+[28] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583-589, 2021. 2  
+[29] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, pages 26565-26577. Curran Associates, Inc., 2022. 4  
+[30] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 2  
+[31] Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Representations, 2023. 3, 5  
+[32] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeurIPS, 2017. 2, 3  
+[33] Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416-1421, 2023. 2  
+[34] Olivier Laurent, Adrien Lafage, Enzo Tartaglione, Geoffrey Daniel, Jean-Marc Martinez, Andrei Bursuc, and Gianni Franchi. Packed-ensembles for efficient uncertainty estimation. In ICLR, 2023. 2, 3  
+[35] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 7, 2, 3  
+[36] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6, 1  
+[37] Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. In *Forty-first International Conference on Machine Learning*. 2  
+[38] Krzysztof Lis, Krishna Nakka, Pascal Fua, and Mathieu Salzmann. Detecting the unexpected via image resynthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2152-2161, 2019. 3  
+[39] Zhenzhen Liu, Jin Peng Zhou, Yufan Wang, and Kilian Q Weinberger. Unsupervised out-of-distribution detection with diffusion inpainting. In International Conference on Machine Learning, pages 22528-22538. PMLR, 2023. 2, 3, 4, 6  
+[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan LI, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Advances in Neural Information Processing Systems, pages 5775-5787. Curran Associates, Inc., 2022. 4
+
+[41] Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3  
+[42] Radford M Neal et al. Mcmc using hamiltonian dynamics. Handbook of markov chain monte carlo, 2011. 2  
+[43] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6  
+[44] Gangadhar Payyavula. Nuclei segmentation in microscope cell images. https://www.kaggle.com/datasets/gangadhar/nuclei-segmentation-in-microscope-cell-images, 2018. Accessed: 2024-09-01. 7, 2, 3  
+[45] Tim Pearce, Alexandra Brintrup, and Jun Zhu. Understanding softmax confidence and uncertainty, 2021. 3  
+[46] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2  
+[47] Jeremy Pinto. Bigearthnet. https://github.com/ jerpint/bigearthnet, 2022. Accessed: 2024-09-01. 7, 2, 3  
+[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. In The Twelfth International Conference on Learning Representations, 2023. 2, 4, 6  
+[49] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021.  
+[50] 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. 2, 4, 6  
+[51] 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. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2, 3  
+[52] Jaskirat Singh and Liang Zheng. Divide, evaluate, and refine: Evaluating and improving text-to-image alignment with iterative vqa feedback. Advances in Neural Information Processing Systems, 36:70799-70811, 2023. 2  
+[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 4  
+[54] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 4
+
+[55] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2409.17146, 2024. 4  
+[56] Tomas Vojir, Tomáš Šipka, Rahaf Aljundi, Nikolay Chumerin, Daniel Olmeda Reino, and Jiri Matas. Road anomaly detection by partial image reconstruction with segmentation coupling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15651-15660, 2021. 3  
+[57] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In ICML, 2011. 2  
+[58] Yeming Wen, Dustin Tran, and Jimmy Ba. BatchEnsemble: an alternative approach to efficient ensemble and lifelong learning. In ICLR, 2019. 2  
+[59] Guoxuan Xia and Christos-Savvas Bouganis. Augmenting softmax information for selective classification with out-of-distribution data. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1995-2012, 2022. 3  
+[60] Guoxuan Xia and Christos-Savvas Bouganis. On the usefulness of deep ensemble diversity for out-of-distribution detection, 2022. 3  
+[61] Guoxuan Xia and Christos-Savvas Bouganis. Window-based early-exit cascades for uncertainty estimation: When deep ensembles are more efficient than single models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17368-17380, 2023. 2, 3  
+[62] Guoxuan Xia, Duolikun Danier, Ayan Das, Stathi Fotiadis, Farhang Nabiei, Ushnish Sengupta, and Alberto Bernacchia. Score Normalization for a Faster Diffusion Exponential Integrator Sampler. In NeurIPS 2023 Workshop on Diffusion Models, 2023. 4  
+[63] Guoxuan Xia, Olivier Laurent, Gianni Franchi, and Christos-Savvas Bouganis. Towards understanding why label smoothing degrades selective classification and how to fix it, 2024. 3  
+[64] Yingda Xia, Yi Zhang, Fengze Liu, Wei Shen, and Alan L Yuille. Synthesize then compare: Detecting failures and anomalies for semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 145-161. Springer, 2020. 3  
+[65] Yilun Xu, Mingyang Deng, Xiang Cheng, Yonglong Tian, Ziming Liu, and Tommi S. Jaakkola. Restart sampling for improving generative processes. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 4  
+[66] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, 2024. 3  
+[67] Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roeah Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. What you see is what you read? improving text-image alignment evaluation. Advances in Neural Information Processing Systems, 36:1601-1619, 2023. 2  
+[68] Xuanwu Yin Yuda Song, Zehao Sun. Sdxs: Real-time one-step latent diffusion models with image conditions. arxiv, 2024. 6, 2  
+[69] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. In The Eleventh
+
+International Conference on Learning Representations, 2023.4  
+[70] 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. 4  
+[71] 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. 1  
+[72] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 6, 1  
+[73] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now, 2024. 7, 3

2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:737a2eaa794beb17489c38a478c2f7944d0e2a984f8c7ffcead45e528a0b16d6
+size 679094

2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/c23bce64-f04c-4e24-b700-3d6473e34043_content_list.json

@@ -0,0 +1,1359 @@
+[
+    {
+        "type": "text",
+        "text": "Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network",
+        "text_level": 1,
+        "bbox": [
+            107,
+            128,
+            890,
+            174
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Haifeng Zhang, Qinghui He, Xiuli Bi, Weisheng Li, Bo Liu*, Bin Xiao Chongqing University of Posts and Telecommunications",
+        "bbox": [
+            209,
+            203,
+            785,
+            239
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "s220331143@stu.cqupt.edu.cn, s220232014@stu.cqupt.edu.cn",
+        "bbox": [
+            250,
+            241,
+            746,
+            256
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{bix1, liws, boliu, xiaobin}@cqupt.edu.cn",
+        "bbox": [
+            313,
+            258,
+            676,
+            273
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            308,
+            326,
+            325
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The rapid advancement of generative models has significantly improved the quality of generated images. Meanwhile, it challenges information authenticity and credibility. Current generated image detection methods based on large-scale pre-trained multimodal models have achieved impressive results. Although these models provide abundant features, the authentication task-related features are often submerged. Consequently, those authentication task-irrelated features cause models to learn superficial biases, thereby harming their generalization performance across different model genera (e.g., GANs and Diffusion Models). To this end, we proposed VIB-Net, which uses Variational Information Bottlenecks to enforce authentication task-related feature learning. We tested and analyzed the proposed method and existing methods on samples generated by 17 different generative models. Compared to SOTA methods, VIB-Net achieved a  $5.55\\%$  improvement in mAP and a  $9.33\\%$  increase in accuracy. Notably, in generalization tests on unseen generative models from different series, VIB-Net improved mAP by  $12.48\\%$  and accuracy by  $23.59\\%$  over SOTA methods. The code is available at https://github.com/oceanzhf/VIBAIGCDetect.",
+        "bbox": [
+            89,
+            340,
+            485,
+            674
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            91,
+            700,
+            220,
+            717
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The continuous advancements in deep learning have led to the emergence of numerous well-known image generation models, such as GANs[9][15] and Diffusion[12][31]. Images generated by these models have reached a level where observers can hardly distinguish them from real images, and the flexibility of tools based on these models has also greatly improved [6]. With a few simple steps, people can easily obtain high-quality images as needed. While this presents a tremendous opportunity for fields like visual arts, it also serves as a paradise for individuals looking to fab",
+        "bbox": [
+            89,
+            727,
+            483,
+            878
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/432848555de54ec22b2ec2160073a539e39993a3a52b1f214925a0ea15c4b742.jpg",
+        "image_caption": [
+            "Figure 1. Each method's overall performance('P') on test data and generalization performance ('G') on unseen data are linked by dashed lines. Our detector demonstrates a clear advantage over all SOTA methods in both metrics."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            516,
+            308,
+            905,
+            486
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ricate misinformation. Given the ease of creating manipulated images today and the potential risk of using them for malicious purposes, there is an urgent need to develop detectors that help distinguish between real and generated images.",
+        "bbox": [
+            511,
+            580,
+            906,
+            656
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Existing detection methods [4][39][25][43] focus on identifying specific traces left by generators in either the spatial or frequency domain of generated images, thus establishing a decision boundary between real and generated images. These detection methods are highly effective in detecting images generated by GANs and earlier diffusion models. However, their performance drops significantly when facing images generated by recent, popular models that visually resemble real images. Although some error reconstruction-based methods [40][23] have been developed specifically for detecting images generated by diffusion models, they similarly experience performance declines when applied to GAN-generated images. Consequently, current research increasingly focuses on improving the cross-model genera detection capabilities of detectors.",
+        "bbox": [
+            511,
+            657,
+            908,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Generalization across different domains is a primary re",
+        "bbox": [
+            531,
+            885,
+            903,
+            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,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Corresponding Author",
+        "bbox": [
+            107,
+            887,
+            241,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "23828",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "search challenge in deep learning. In this context, the emergence of large-scale pre-trained multimodal models has introduced new approaches and promising results. For instance, by using the image encoder from the CLIP [30] model to extract image features and training a classifier with linear probing, there is a notable improvement in generalization performance over models like ResNet50[28][6]. However, compared to traditional classification tasks, real and generated images share subtle, harder-to-capture distinguishing features. This implies that the features extracted by large-scale pre-trained models for the task of distinguishing real from fake images may contain a considerable amount of task-irrelevant features. Task-irrelevant features can either have no impact on the realizness of the image or, though effective in isolation, become redundant when combined with other features. These features not only interfere with the classifier's ability to accurately learn decision boundaries but also hinder the model's ability to capture and distinguish task-relevant characteristics. More critically, task-irrelevant features may lead the model to learn superficial biases. For example, if a fake image contains a tree while the real image does not, the model might erroneously associate the presence of the tree with being a fake image. Consequently, when detecting other images, the model may mistakenly classify a real image as fake if it contains a tree. Therefore, when dealing with high-quality generated images with subtle distinguishing features, it is challenging to fine-tune the model to separate essential from unnecessary information. This can lead to an overfitting of the incidental statistical correlations between unnecessary information and target images, ultimately undermining the model's generalization and discrimination capabilities.",
+        "bbox": [
+            93,
+            90,
+            480,
+            571
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "How can we harness the powerful feature extraction capability of large models while minimizing the interference of unnecessary information? By approaching this from a novel perspective, we propose addressing the issue using Information Bottleneck Theory. This theory aims to create latent encodings that not only effectively represent the target but also maximally compress the original input, thereby avoiding the impact of task-irrelevant features on the model. Specifically, we introduce a method called VIB-Net, which couples a large-scale pre-trained multimodal model with Variational Information Bottleneck (VIB) [1]. Considering that feature representations extracted by the model's image encoder are more task-specific, we decouple and filter these features using latent variables. This method maximizes the mutual information between the latent feature and the classification output, while minimizing the mutual information between the latent feature and the input, resulting in a concise feature that should only contain the effective information needed for the task. This concise latent feature helps the model avoid overfitting to irrelevant features and the incidental correlations with the task, preventing the model",
+        "bbox": [
+            93,
+            584,
+            480,
+            898
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "from learning superficial biases, which is beneficial to the model's generalization performance.",
+        "bbox": [
+            516,
+            92,
+            903,
+            119
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In summary, our contributions are as follows: (1) To the best of our knowledge, we are the first to apply Information Bottleneck theory and techniques to the detection of generated images, demonstrating the potential of this theory and approach in this field. (2) We analyze the limitations of methods based on large-scale pre-trained models in detecting images generated by unseen generative models, and present our theory of what could be wrong with the existing paradigm. (3) We propose VIB-Net for the detection of generated images, which uses VIB to decouple and filter the general features extracted by large-scale pre-trained multimodal models, removing task-irrelevant information to enhance detection performance. (4) A large number of experiments validate the advantages of our method (see Fig. 1), especially in the improvement of generalization performance  $(+23.59\\%$  accuracy).",
+        "bbox": [
+            516,
+            122,
+            903,
+            361
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related work",
+        "text_level": 1,
+        "bbox": [
+            516,
+            377,
+            648,
+            392
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Generalization Performance. The continuous advancement of generated image models presents ongoing challenges for generated image detection, making the generalization performance of detection models on unseen generators an increasing focus for researchers[14][21]. [39] explored the effects of data preprocessing and augmentation methods on the generalization performance of a ResNet50[11]-based detector. [44] argues that the differences between real and fake images are subtle and localized, framing forgery detection as a fine-grained task that could positively impact generalization performance. [3] takes a frequency-domain approach, mapping real images to a dense subspace and synthetic images outside this subspace to enhance generalization. While these methods have achieved certain successes on different models within the same generation series, they show limitations across different series. This is due to the fact that, with the development of generative algorithms, subtle traces in generated images are becoming increasingly difficult to capture. Additionally, artifacts in the frequency domain differ between model series and are gradually decreasing in visibility.",
+        "bbox": [
+            516,
+            402,
+            903,
+            718
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The impressive results of large-scale pre-trained models in terms of transferability and generalization have sparked significant interest among researchers in the field of generated image detection. [28] argues that existing detectors perform asymmetric learning, where they use CLIP's image encoder to extract image features, exploring the potential of large models in detecting different model families. [6] suggests that detectors based on large models rely on higher-level semantic features, which enhances the generalization ability of the detectors. [19] uses both image and text features extracted by the CLIP model, and then employs a multilayer perceptron to detect images generated by",
+        "bbox": [
+            516,
+            720,
+            903,
+            898
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "23829",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "diffusion models, improving generalization performance. Information Bottleneck. The Information Bottleneck (IB) [38] is an innovative information-theoretic method used for pattern analysis and representation learning. It provides an elegant balance between data compression and information retention[13]. VIB [1] is a variational approximation of IB, aimed at parameterizing IB using neural networks. [2] demonstrated VIB's capability in improving classification calibration and out-of-distribution data. [24] explored VIB as a regularization technique to address overfitting issues in NLP models. [41] proposed a dual-stream information bottleneck method to utilize in-distribution data to enhance out-of-distribution target detection tasks.",
+        "bbox": [
+            89,
+            90,
+            485,
+            287
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Based on the above analysis, current methods for detecting generated images using large-scale pre-trained models simply utilize the general representations extracted from images, without thoroughly exploring the relationship between these general features and downstream tasks. Additionally, Information Bottleneck is an effective technique for improving performance in cross-domain and cross-distribution tasks. Therefore, unlike existing methods, our approach leverages Information Bottleneck techniques to explore the potential of large-scale pre-trained models in generated image detection from a completely new perspective, achieving outstanding performance.",
+        "bbox": [
+            89,
+            289,
+            483,
+            470
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            89,
+            484,
+            230,
+            500
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Problem background",
+        "text_level": 1,
+        "bbox": [
+            89,
+            511,
+            290,
+            526
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "A common paradigm for using large-scale pretrained models is to leverage a large-scale pretrained multimodal model (e.g., CLIP) to learn the general representation of real and generated images, and then apply this feature representation to downstream tasks[28][6]. Building on their findings, we further consider the following points: (1) It is well-known that the features extracted by large-scale pre-trained visual encoders are general-purpose, meaning they contain many general features, such as semantic information. (2) Another widely accepted notion is that some critical features for distinguishing real from generated images are subtle and local.",
+        "bbox": [
+            89,
+            532,
+            483,
+            699
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Further combining these two points, we propose the following hypotheses: (i) The general features extracted by large-scale pre-trained multimodal models are more beneficial for real-fake image classification tasks than those extracted by pre-trained visual models. (ii) These general features contain task-irrelevant features for real-fake classification. (iii) These task-irrelevant features affect the classification of real and fake images. Next, we will delve into these three hypotheses.",
+        "bbox": [
+            89,
+            700,
+            483,
+            835
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Analysis of general feature representation",
+        "text_level": 1,
+        "bbox": [
+            89,
+            847,
+            447,
+            864
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For the first hypothesis, We use image pairs of real and generated images to aid our analysis[6], generating textual de",
+        "bbox": [
+            89,
+            869,
+            483,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/e1410341985ff647e745d20a946c655965b19f0cbbf303dfb911f10c27cde35e.jpg",
+        "image_caption": [
+            "Figure 2. UMAP visualization of features extracted from Dalle2 and Dalle3 generated images using large-scale pretrained multimodal(CLIP) and unimodal models(ResNet)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            89,
+            905,
+            308
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "scriptions for 1,000 real images and using these descriptions to produce 1,000 corresponding images with DALL-E 2 and DALL-E 3. We then compare features extracted by large-scale pre-trained multimodal models(CLIP[30]) and pre-trained visual models(ResNet[11]) through UMAP [26] visualization of 3,000 images. The results show that features extracted by multimodal models (e.g., CLIP) exhibit greater distribution differences for real images and DALL-E 2-generated images, suggesting multimodal models may be better at distinguishing real from generated images. However, for DALL-E 3-generated images, both model types show similar feature distributions, possibly due to advancements in image generation tools, which complicate the classification task and require further fine-tuning.",
+        "bbox": [
+            511,
+            397,
+            906,
+            609
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To further explore hypotheses (ii) and (iii), we visualized the extracted features using both category and authenticity as labels to observe their distribution differences. Specifically, we prepared images of two categories (car and cat), each containing 500 real images and 500 generated images. We then used CLIP to extract their features and performed UMAP dimensionality reduction for visualization, with the results shown in Fig. 3a. We can see that the features extracted by CLIP clearly form two clusters: the left cluster mainly consists of light blue and light red points, while the right cluster is mainly composed of light pink and light purple points. These two clusters correspond to category 1 and category 2, respectively. This indicates that CLIP can significantly distinguish category information. However, this clear separation of category information may affect the decision-making process for classifying real vs. generated images. When examining the distinction between real and generated images, we notice that the right cluster shows a noticeable separation of real and generated features, with",
+        "bbox": [
+            511,
+            613,
+            908,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "23830",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/fe8f885df46bbd6b886ae2693b1dc8ce8b2ffa83bcc61143cf31da8d0c17917d.jpg",
+        "image_caption": [
+            "(a) Visualization of features extracted by CLIP.",
+            "Figure 3. UMAP visualization of features extracted from real and generated images of two category using CLIP and VIB-Net."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            89,
+            284,
+            212
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/1ab822a33b6db285e7ab7f352b8aedae2588e0ed8278793c8c97bcbb3a4451d0.jpg",
+        "image_caption": [
+            "(b) Visualization of features extracted by VIB-Net."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            285,
+            90,
+            480,
+            212
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "two different colors observed within the cluster. This suggests that CLIP is capable of distinguishing real and generated features within category 2. However, in the left cluster, CLIP does not show significant ability to differentiate between real and generated features. We observe that in the left cluster (category 1), the two colors are mixed together, showing poor distinguishability. From these observations, it is easy to infer that the internal separation of real or generated images caused by semantic category information makes it more difficult for the classifier to learn the decision boundary. More importantly, the strong sensitivity of large-scale pre-trained multimodal models like CLIP to semantic information leads the model to tend to learn superficial biases.",
+        "bbox": [
+            88,
+            313,
+            482,
+            523
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Based on this viewpoint, we consider whether it is possible to minimize the irrelevant features within these multimodal features while retaining the ones that are relevant to the classification task. This would help the model avoid learning superficial biases, allow the classifier to more easily learn decision boundaries, and thereby enhance the model's discrimination ability and generalization capability. This is the motivation of this paper. As shown in Fig. 3b, the method(VIB-Net) proposed in this paper achieves significant results.",
+        "bbox": [
+            88,
+            527,
+            482,
+            679
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4. Approach",
+        "text_level": 1,
+        "bbox": [
+            89,
+            705,
+            199,
+            722
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Based on the issues analyzed in Sec. 3, an effective approach is to filter the multimodal features in order to obtain the representations most beneficial for the real-fake classification task, which serves as the starting point of this paper. Overall, our method is illustrated in Fig. 4. VIB attempts to filter out irrelevant features and to find the most concise representation, making it easier for the classifier to learn the features needed for distinguishing between real and fake tasks, thereby significantly improving the accuracy of the classification task. Below, we will provide a detailed description of our method.",
+        "bbox": [
+            88,
+            734,
+            482,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1. Information bottleneck constraint",
+        "text_level": 1,
+        "bbox": [
+            511,
+            90,
+            808,
+            104
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To specifically optimize the filtering of irrelevant information in the input features while retaining task-relevant information, we apply the information bottleneck (IB) principle as a constraint on the features to be optimized. Let  $X$ ,  $Z$ , and  $Y$  represent the input, the latent compressed representation, and the output target, respectively. The goal of IB is to find the maximum compressed representation  $Z$  of the input  $X$ , while preserving as much information about the output  $Y$  as possible. The maximization objective function is defined as follows:",
+        "bbox": [
+            511,
+            112,
+            906,
+            263
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {I B} = I (Z, Y) - \\beta I (Z, X) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            609,
+            272,
+            906,
+            289
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Where  $I(.,.)$  denotes mutual information and  $\\beta$  is a trade-off parameter. The first term constrains the latent compressed representation to retain sufficient information for downstream tasks, while the second term ensures that irrelevant and redundant information is filtered out.",
+        "bbox": [
+            511,
+            296,
+            906,
+            372
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2. Variational information bottleneck",
+        "text_level": 1,
+        "bbox": [
+            511,
+            380,
+            815,
+            393
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To parameterize the information bottleneck (IB) using deep neural networks, we perform variational inference on Equation (1). First, based on the definition of mutual information, Equation (1) can be rewritten as:",
+        "bbox": [
+            511,
+            402,
+            905,
+            463
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} L _ {I B} = \\int d y d z p (y, z) \\log {\\frac {p (y | z)}{p (y)}} \\\\ - \\beta \\int d x d z p (x, z) \\log \\frac {p (z \\mid x)}{p (z)} \\tag {2} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            589,
+            482,
+            906,
+            550
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Where  $p(y|z)$  is defined by the encoder and the Markov chain as follows:",
+        "bbox": [
+            511,
+            556,
+            905,
+            585
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np (y \\mid z) = \\int d x \\frac {p (y \\mid x) p (z \\mid x) p (x)}{p (z)} \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            578,
+            592,
+            905,
+            625
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Since it is not computable here, we consider setting  $q(y|z)$  as a parameterized approximation of  $p(y|z)$ , serving as a \"decoder,\" which can be approximated by a linear layer. We introduce the Kullback-Leibler (KL) divergence to obtain an estimizable representation. As the KL divergence satisfies  $\\mathrm{KL}(p(y|z),q(y|z))\\geq 0$ , and by applying the Markov assumption,  $p(y,z)$  can be rewritten as  $p(y,z) = \\int dx p(x)p(y|x)p(z|x)$ , thereby providing a lower bound for the first term of the objective function:",
+        "bbox": [
+            511,
+            632,
+            906,
+            768
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nI (Z, Y) \\geq \\int d x d y d z p (x) p (y | x) p (z | x) \\log q (y | z) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            524,
+            775,
+            903,
+            818
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "At the same time, let  $r(z)$  be an estimate of the prior probability  $p(z)$ , used to approximate the marginal distribution, thereby obtaining an upper bound for the second term:",
+        "bbox": [
+            511,
+            819,
+            903,
+            864
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nI (Z, X) \\leq \\int d x d z p (x) p (z | x) \\log \\frac {p (z | x)}{r (z)} \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            555,
+            871,
+            903,
+            905
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "23831",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/f2654d25fbfe695a4e4ecf0ae41380f6fa32b64f9234168d03dc86580330d688.jpg",
+        "image_caption": [
+            "Figure 4. Overview of VIB-Net. We utilize Variational Information Bottleneck (VIB) techniques to map the image features extracted by CLIP's visual encoder into latent features  $z$ , where the information in  $z$  is filtered based on the Information Bottleneck (IB). Subsequently, the latent features  $z$  are used as input to an MLP to learn the decision boundary."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            89,
+            88,
+            903,
+            301
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The encoder we use is defined as  $p(z|x) = p_{\\theta}(z|x) = \\mathcal{N}\\left(z|f^{\\mu}(x), f^{\\Sigma}(x)\\right)$ , where  $f(.)$  represents a shallow MLP,  $\\mu$  denotes the K-dimensional mean vector of the output  $z$ , and  $\\Sigma$  represents the  $K \\times K$  covariance matrix. We select a parameterized Gaussian distribution for both  $p_{\\theta}(z|x)$  and  $r(z)$  to allow for the analytical computation of KL divergence. Additionally, we employ the reparameterization trick to estimate gradients [18]. Combining all the components above yields the following objective function:",
+        "bbox": [
+            88,
+            378,
+            483,
+            517
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} L _ {V I B} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} z \\sim p _ {\\theta} (z | x) [ - \\log q (y _ {\\mathrm {n}} | z _ {n}) ] \\\\ + \\beta \\mathrm {K L} [ p _ {\\theta} (z | x _ {n}), r (z) ] \\tag {6} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            138,
+            537,
+            482,
+            599
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3. Workflow of the overall method",
+        "text_level": 1,
+        "bbox": [
+            89,
+            607,
+            372,
+            622
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As shown in Fig. 4, the parts in VIB-Net work in the following way: First, the real and generated images  $\\{x_{1}, x_{2}, x_{3}, \\dots, x_{n}\\}$  are input into the CLIP image encoder, yielding the corresponding image features  $\\{t_{1}, t_{2}, t_{3}, \\dots, t_{n}\\}$ . Next, these feature vectors are passed into the Variational Information Bottleneck (VIB). Specifically, the image features are processed through a shallow MLP, then a linear layer  $f(.)$  calculates  $\\mu$  and  $\\Sigma$ , which are transformed using a softmax activation. This layer has  $K$  dimensions, representing the bottleneck size. At this point, we obtain the latent compressed feature  $z$ , which is then fed into a linear layer  $q(y|z)$  to output the target task label  $y$ .",
+        "bbox": [
+            89,
+            628,
+            483,
+            811
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            823,
+            223,
+            840
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1. Settings",
+        "text_level": 1,
+        "bbox": [
+            89,
+            848,
+            189,
+            864
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To better evaluate the model's performance, we utilize two mainstream public datasets: ForenSynths[39] and",
+        "bbox": [
+            89,
+            869,
+            483,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Genimage[46]. The ForenSynths dataset primarily consists of images generated by GANs, while the Genimage dataset mainly includes images generated by diffusion models. Consequently, we set up two training and testing schemes: (1) Training on ForenSynths and then testing on both ForenSynths and Genimage, and (2) similarly, training on Genimage and testing on both Genimage and ForenSynths. This setup allows us to more effectively validate the generalization performance of the detector on cross model genera scenarios. The details are as follows:",
+        "bbox": [
+            511,
+            378,
+            906,
+            529
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training set:",
+        "text_level": 1,
+        "bbox": [
+            513,
+            532,
+            607,
+            546
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The ForenSynths training set consists of 20 different categories, each containing 18,000 real images from the LSUN dataset [42] and an equal number of images generated using ProGAN[15]. Following the setup in Genimage, we use images generated by Stable Diffusion v1.4 [32] in the Genimage dataset, along with their corresponding real images, as the training set. In total, these include 32,000 real and generated images.",
+        "bbox": [
+            511,
+            547,
+            905,
+            667
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Testing set:",
+        "text_level": 1,
+        "bbox": [
+            513,
+            670,
+            596,
+            684
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "11 models from ForenSynths The test set includes images generated by 11 different models, primarily from the GAN family[15][45][29][5][16]. The real images are sourced from six datasets, including ImageNet[34], LSUN[42], FaceForensics++[33], CelebA-HQ[15], CelebA[22], and COCO[20].",
+        "bbox": [
+            511,
+            686,
+            905,
+            777
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "7 models from Genimage To validate the method across different generation models, we use images generated by seven advanced generators along with their corresponding real images as part of the test set. These seven generation models include Stable Diffusion v1.4/1.5[32], VQDM[10], ADM[7], GLIDE[27], Wukong, and Midjourney.",
+        "bbox": [
+            511,
+            777,
+            906,
+            869
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Real-vs-Fake classification baselines: We compare VIB-Net with existing state-of-the-art detection methods, in",
+        "bbox": [
+            511,
+            869,
+            905,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "23832",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/ec2b24faa11d5f7a5f1a5dcca0ab1010d0fb1a26e14f0229abea264ef53b80fa.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Detection method</td><td colspan=\"7\">Diffusion Models</td><td colspan=\"5\">Generative Adversarial Networks</td><td colspan=\"2\">Others</td><td>AP</td><td></td></tr><tr><td>SDV1.4</td><td>SDV1.5</td><td>ADM</td><td>GLIDE</td><td>Midjourney</td><td>Wukong</td><td>VQDM</td><td>Pro-GAN</td><td>Cycle-Big-GAN</td><td>Style-GAN</td><td>Star-GAN</td><td>Gau-GAN</td><td colspan=\"2\">Deep-fake</td><td>Average</td><td></td></tr><tr><td>CNNSpot[39]</td><td>99.98</td><td>99.83</td><td>51.10</td><td>58.80</td><td>67.93</td><td>99.80</td><td>49.92</td><td>53.15</td><td>50.23</td><td>49.79</td><td>55.98</td><td>47.07</td><td>56.08</td><td>54.86</td><td>54.03</td><td>63.24</td></tr><tr><td>Fusing[14]</td><td>99.90</td><td>97.98</td><td>69.30</td><td>94.20</td><td>81.20</td><td>99.90</td><td>84.60</td><td>67.63</td><td>87.79</td><td>69.37</td><td>67.90</td><td>91.20</td><td>43.08</td><td>72.56</td><td>89.42</td><td>81.07</td></tr><tr><td>Lgrad[36]</td><td>99.94</td><td>99.92</td><td>58.52</td><td>84.00</td><td>91.06</td><td>99.72</td><td>56.34</td><td>83.59</td><td>90.24</td><td>47.51</td><td>82.74</td><td>99.19</td><td>49.25</td><td>66.49</td><td>65.09</td><td>78.24</td></tr><tr><td>Univfd[28]</td><td>96.04</td><td>96.26</td><td>66.34</td><td>93.73</td><td>92.08</td><td>90.98</td><td>74.53</td><td>51.77</td><td>63.42</td><td>75.81</td><td>54.12</td><td>54.93</td><td>65.99</td><td>70.24</td><td>83.34</td><td>75.31</td></tr><tr><td>NPR[37]</td><td>100.00</td><td>99.97</td><td>94.70</td><td>95.80</td><td>95.50</td><td>100.00</td><td>86.30</td><td>83.30</td><td>94.90</td><td>72.00</td><td>82.70</td><td>97.30</td><td>66.00</td><td>85.30</td><td>95.90</td><td>89.98</td></tr><tr><td>CLIPping[17]</td><td>93.97</td><td>93.10</td><td>68.00</td><td>87.44</td><td>77.34</td><td>86.52</td><td>77.17</td><td>88.54</td><td>88.44</td><td>85.33</td><td>77.23</td><td>89.82</td><td>81.56</td><td>61.19</td><td>57.37</td><td>80.87</td></tr><tr><td>Ours</td><td>100.00</td><td>99.97</td><td>95.49</td><td>97.13</td><td>97.81</td><td>99.93</td><td>97.00</td><td>96.59</td><td>98.44</td><td>97.17</td><td>84.31</td><td>97.60</td><td>96.94</td><td>81.32</td><td>93.27</td><td>95.53</td></tr></table>",
+        "bbox": [
+            94,
+            88,
+            906,
+            258
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/2c33959f89ed5c07d0089de82cedaf079f6dd783e06c4f9afa7a02d5978a4abf.jpg",
+        "table_caption": [
+            "Table 1. The  $AP$  values of different methods based on diffusion sources. Models outside the Diffusion Models (DMs) column are considered as the generalization domain across generative model series."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Detection method</td><td colspan=\"7\">Diffusion Models</td><td colspan=\"5\">Generative Adversarial Networks</td><td>Others</td><td>ACC</td><td></td></tr><tr><td>SDV1.4</td><td>SDV1.5</td><td>ADM</td><td>GLIDE</td><td>Midjourney</td><td>Wukong</td><td>VQDM</td><td>Pro-GAN</td><td>Cycle-Big-GAN</td><td>Style-GAN</td><td>Star-GAN</td><td>Gau-GAN</td><td>Deep-SAN</td><td>Average</td><td></td></tr><tr><td>CNNSpot[39]</td><td>99.48</td><td>99.35</td><td>50.10</td><td>50.90</td><td>56.42</td><td>97.90</td><td>50.04</td><td>50.27</td><td>49.81</td><td>50.10</td><td>50.98</td><td>49.77</td><td>50.38</td><td>51.98</td><td>50.22</td></tr><tr><td>Fusing[14]</td><td>99.90</td><td>99.91</td><td>51.30</td><td>57.50</td><td>52.30</td><td>99.90</td><td>64.20</td><td>51.20</td><td>52.40</td><td>53.50</td><td>50.20</td><td>58.20</td><td>49.32</td><td>51.02</td><td>64.84</td></tr><tr><td>Lgrad[36]</td><td>99.12</td><td>99.05</td><td>53.00</td><td>64.24</td><td>76.34</td><td>97.53</td><td>50.93</td><td>61.61</td><td>60.74</td><td>48.82</td><td>61.43</td><td>50.17</td><td>49.70</td><td>50.17</td><td>56.49</td></tr><tr><td>Univfd[28]</td><td>83.55</td><td>84.80</td><td>53.35</td><td>75.30</td><td>71.60</td><td>73.55</td><td>55.10</td><td>58.65</td><td>59.30</td><td>61.45</td><td>56.80</td><td>61.45</td><td>55.30</td><td>58.40</td><td>72.00</td></tr><tr><td>NPR[37]</td><td>100.00</td><td>99.90</td><td>73.00</td><td>89.70</td><td>82.30</td><td>100.00</td><td>68.30</td><td>60.30</td><td>67.20</td><td>59.20</td><td>58.00</td><td>73.20</td><td>52.00</td><td>74.80</td><td>89.60</td></tr><tr><td>CLIPping[17]</td><td>96.07</td><td>95.48</td><td>70.14</td><td>85.00</td><td>77.66</td><td>88.86</td><td>79.35</td><td>88.78</td><td>88.48</td><td>89.57</td><td>80.69</td><td>90.82</td><td>85.94</td><td>66.91</td><td>61.64</td></tr><tr><td>Ours</td><td>99.55</td><td>99.20</td><td>73.85</td><td>74.25</td><td>88.05</td><td>98.25</td><td>89.35</td><td>89.70</td><td>88.60</td><td>91.20</td><td>74.10</td><td>80.70</td><td>87.15</td><td>72.00</td><td>81.50</td></tr></table>",
+        "bbox": [
+            94,
+            309,
+            906,
+            481
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Similar to Tab. 1, we compare these methods using classification accuracy, which is influenced by manually set thresholds. For a fair comparison, the classification threshold for all models is set to 0.5.",
+        "bbox": [
+            89,
+            492,
+            906,
+            521
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "cluding CNNDetection (CVPR 2020) [39], FreDect (ICML 2020) [8], Fusing (ICIP 2022) [14], LGrad (CVPR 2023) [36], DIRE (ICCV 2023) [40], Ojha (CVPR 2023) [28], NPR (CVPR 2024) [37], CLIPping(ICMR 2024) [17]. To ensure a fair comparison, we re-train and predict using the official code on ForenSynths and Genimage.",
+        "bbox": [
+            88,
+            544,
+            482,
+            651
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation metrics We follow existing research and consider two metrics: Average Precision (AP) and Accuracy (ACC). Average Precision (AP) is threshold-independent and evaluates the overall performance of the model. Accuracy is defined as the number of correct predictions divided by the number of test images, and this metric is influenced by the decision threshold. Given that there is no prior information in real-world scenarios, we standardize the decision threshold to 0.5.",
+        "bbox": [
+            88,
+            654,
+            482,
+            789
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details Our method is implemented using the PyTorch framework on the NVIDIA Tesla V100. The model is trained with the Adam optimizer, a learning rate of  $2 \\times 10^{-4}$ , and a batch size of 256. Following the same data augmentation strategy as in [39][28], which includes Gaussian blur and JPEG compression, experiments have shown that this is critical for the model's generalization ability.",
+        "bbox": [
+            88,
+            794,
+            482,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2. Performance evaluation based on Diffusion sources",
+        "text_level": 1,
+        "bbox": [
+            511,
+            544,
+            906,
+            574
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "It is well-known that images generated by the current state-of-the-art diffusion models are more realistic than those generated by other types of generative models. To validate the effectiveness of the proposed method, we designed a challenging experimental scenario. Specifically, we trained various detector using Stable Diffusion v1.4 and then tested them on images generated by 15 different models, which include the latest diffusion-based generative models as well as classic GANs. The results are presented in Tabs. 1 and 2. Bold indicates the highest value in the column, and underline indicates the second highest value in the column. This applies to the other tables as well.",
+        "bbox": [
+            509,
+            582,
+            906,
+            763
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We first evaluated the performance of each model on the latest generative models. As shown in Tabs. 1 and 2, advanced diffusion models significantly challenged most detection models, causing a sharp drop in performance. For example, Univfd and Lgrad experienced a  $40\\%$  decrease in accuracy and precision when tested on ADM, Midjourney, and VQDM. This decline is likely due to the high-quality images produced by these models, which make it difficult for detection methods to extract distinguishing fea",
+        "bbox": [
+            509,
+            763,
+            908,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "23833",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/96b08620198dd1fda7f1a026e64f21038d153e055719354ff20bd9af4f1aad29.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Detection method</td><td colspan=\"6\">Generative Adversarial Networks</td><td colspan=\"2\">Perceptual loss</td><td colspan=\"2\">Others</td><td colspan=\"7\">Diffusion Models</td><td>AP</td></tr><tr><td>Pro-GAN</td><td>Cycle-GAN</td><td>Big-GAN</td><td>Style-GAN</td><td>Star-GAN</td><td>Gau-GAN</td><td>CRN</td><td>IMLE</td><td>Deep-fake</td><td>SAN</td><td>SDV1.4</td><td>SDV1.5</td><td>ADM</td><td>GLIDE</td><td>Midjourney</td><td>Wukong</td><td>VQDM</td><td>Average</td></tr><tr><td>CNNSpot[39]</td><td>99.99</td><td>96.40</td><td>87.50</td><td>96.94</td><td>94.24</td><td>98.28</td><td>96.61</td><td>98.76</td><td>64.42</td><td>55.89</td><td>52.86</td><td>53.25</td><td>65.14</td><td>68.10</td><td>56.60</td><td>51.15</td><td>69.49</td><td>76.80</td></tr><tr><td>FreDect[8]</td><td>99.99</td><td>84.77</td><td>93.62</td><td>88.97</td><td>99.48</td><td>82.85</td><td>82.27</td><td>69.77</td><td>70.77</td><td>49.50</td><td>38.50</td><td>38.41</td><td>63.72</td><td>54.73</td><td>47.25</td><td>40.44</td><td>86.01</td><td>70.06</td></tr><tr><td>Fusing[14]</td><td>100.00</td><td>95.50</td><td>90.76</td><td>99.48</td><td>99.82</td><td>88.32</td><td>90.32</td><td>96.02</td><td>71.12</td><td>77.33</td><td>65.30</td><td>65.62</td><td>74.85</td><td>77.44</td><td>69.91</td><td>64.53</td><td>75.42</td><td>82.46</td></tr><tr><td>Lgrad[36]</td><td>99.90</td><td>94.01</td><td>90.75</td><td>99.80</td><td>99.98</td><td>79.29</td><td>61.64</td><td>67.05</td><td>71.71</td><td>45.09</td><td>70.90</td><td>71.72</td><td>71.83</td><td>75.96</td><td>71.42</td><td>66.51</td><td>70.23</td><td>76.93</td></tr><tr><td>Univfd[28]</td><td>100.00</td><td>99.21</td><td>98.31</td><td>97.98</td><td>99.35</td><td>99.80</td><td>96.72</td><td>99.00</td><td>82.04</td><td>82.18</td><td>85.48</td><td>82.30</td><td>84.34</td><td>84.04</td><td>69.10</td><td>90.13</td><td>94.96</td><td>90.88</td></tr><tr><td>NPR[37]</td><td>100.00</td><td>98.50</td><td>87.80</td><td>99.80</td><td>99.90</td><td>85.50</td><td>50.00</td><td>50.00</td><td>82.40</td><td>71.60</td><td>84.00</td><td>84.60</td><td>74.60</td><td>85.70</td><td>85.40</td><td>80.50</td><td>81.20</td><td>82.44</td></tr><tr><td>CLIPping[17]</td><td>99.85</td><td>94.03</td><td>91.27</td><td>95.24</td><td>99.05</td><td>89.52</td><td>79.87</td><td>80.33</td><td>73.89</td><td>56.62</td><td>58.28</td><td>57.85</td><td>77.25</td><td>77.93</td><td>52.43</td><td>60.65</td><td>80.00</td><td>77.89</td></tr><tr><td>Ours</td><td>100.00</td><td>99.80</td><td>99.29</td><td>98.79</td><td>99.72</td><td>99.99</td><td>90.89</td><td>96.83</td><td>92.64</td><td>91.62</td><td>87.24</td><td>86.98</td><td>87.88</td><td>88.53</td><td>75.68</td><td>90.92</td><td>96.51</td><td>93.14</td></tr></table>",
+        "bbox": [
+            94,
+            88,
+            906,
+            252
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/2b78035a5b0191d0a345b4c6ace643fa0f58d9a5fd5c3da9ffc538edebe59679.jpg",
+        "table_caption": [
+            "Table 3. The  $AP$  values of different methods trained on GAN source images. Generative models other than those in the GANs column can be considered as the generalization domain."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Detection method</td><td colspan=\"6\">Generative Adversarial Networks</td><td colspan=\"2\">Perceptual loss</td><td colspan=\"2\">Others</td><td colspan=\"7\">Diffusion Models</td><td>ACC</td></tr><tr><td>Pro-GAN</td><td>Cycle-GAN</td><td>Big-GAN</td><td>Style-GAN</td><td>Star-GAN</td><td>Gau-GAN</td><td>CRN</td><td>IMLE</td><td>Deep-fake</td><td>SAN</td><td>SDV1.4</td><td>SDV1.5</td><td>ADM</td><td>GLIDE</td><td>Midjourney</td><td>Wukong</td><td>VQDM</td><td>Average</td></tr><tr><td>CNNSpot[39]</td><td>99.99</td><td>87.59</td><td>71.18</td><td>89.95</td><td>94.60</td><td>81.44</td><td>85.87</td><td>85.87</td><td>51.69</td><td>50.00</td><td>50.82</td><td>50.88</td><td>60.20</td><td>57.85</td><td>50.77</td><td>51.13</td><td>56.20</td><td>69.18</td></tr><tr><td>FreDect[8]</td><td>99.36</td><td>78.76</td><td>81.97</td><td>78.01</td><td>94.62</td><td>80.57</td><td>60.04</td><td>60.13</td><td>63.29</td><td>50.00</td><td>40.02</td><td>40.38</td><td>64.66</td><td>55.43</td><td>46.89</td><td>41.54</td><td>78.95</td><td>65.57</td></tr><tr><td>Fusing[14]</td><td>99.90</td><td>87.01</td><td>77.32</td><td>85.21</td><td>97.04</td><td>76.95</td><td>85.17</td><td>88.54</td><td>53.76</td><td>54.56</td><td>51.04</td><td>51.34</td><td>56.50</td><td>57.15</td><td>52.12</td><td>51.67</td><td>55.09</td><td>69.43</td></tr><tr><td>DIRE[40]</td><td>100.00</td><td>76.65</td><td>72.12</td><td>85.03</td><td>100.00</td><td>67.79</td><td>-</td><td>-</td><td>-</td><td>-</td><td>49.74</td><td>49.83</td><td>75.78</td><td>71.75</td><td>58.01</td><td>54.46</td><td>53.68</td><td>-</td></tr><tr><td>Lgrad[36]</td><td>99.80</td><td>86.94</td><td>85.63</td><td>91.08</td><td>99.27</td><td>72.49</td><td>50.49</td><td>50.50</td><td>56.42</td><td>44.47</td><td>63.03</td><td>63.67</td><td>67.10</td><td>66.10</td><td>56.20</td><td>63.60</td><td>67.02</td><td>69.64</td></tr><tr><td>Univfd[28]</td><td>99.90</td><td>98.50</td><td>94.50</td><td>84.40</td><td>95.85</td><td>99.50</td><td>59.50</td><td>72.00</td><td>67.40</td><td>56.50</td><td>63.10</td><td>63.57</td><td>66.90</td><td>61.70</td><td>57.85</td><td>71.06</td><td>85.00</td><td>76.31</td></tr><tr><td>NPR[37]</td><td>99.80</td><td>96.10</td><td>84.40</td><td>97.70</td><td>99.30</td><td>82.50</td><td>50.00</td><td>50.00</td><td>80.20</td><td>69.20</td><td>76.60</td><td>77.90</td><td>69.70</td><td>77.30</td><td>77.80</td><td>76.10</td><td>78.10</td><td>78.98</td></tr><tr><td>CLIPping[17]</td><td>99.88</td><td>96.74</td><td>94.77</td><td>94.87</td><td>99.47</td><td>94.15</td><td>85.23</td><td>87.73</td><td>76.48</td><td>60.27</td><td>61.26</td><td>60.78</td><td>80.77</td><td>80.69</td><td>53.95</td><td>63.94</td><td>84.76</td><td>80.93</td></tr><tr><td>Ours</td><td>99.99</td><td>99.00</td><td>95.75</td><td>91.25</td><td>98.95</td><td>99.70</td><td>71.25</td><td>86.75</td><td>83.20</td><td>70.50</td><td>71.55</td><td>70.00</td><td>71.45</td><td>69.40</td><td>61.25</td><td>75.90</td><td>86.65</td><td>82.50</td></tr></table>",
+        "bbox": [
+            94,
+            304,
+            906,
+            479
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Accuracy of different methods trained on GAN source images. The data from [40] is obtained from [37], and ”-” indicates that the method did not show data for the corresponding model.",
+        "bbox": [
+            89,
+            491,
+            906,
+            520
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "tures. We also observed that these models performed better on generative models with distributions similar to the training data (e.g., SDV1.5, Wukong). This suggests that they may have learned superficial biases from irrelevant features, which are superficially correlated with labels and do not generalize well to images from different distributions. In contrast, NPR, which detects authenticity by analyzing pixel relationships, partially mitigates these biases, resulting in less performance decline. When testing on GANs, we observed significant performance drops, particularly on GauGAN, where the accuracy and precision were around  $50 - 60\\%$ . This further highlights how even subtle relevant features and surface biases can severely impact the generalization performance of these models.",
+        "bbox": [
+            88,
+            539,
+            482,
+            750
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In contrast, Our method outperforms the baseline model, Univfd, with improvements of  $20.22\\%$  in average precision and  $20.46\\%$  in accuracy. This gap is even more pronounced across generative model families (GANs), where our method surpasses Univfd by  $34.17\\%$  in precision and  $26.41\\%$  in accuracy. Compared to state-of-the-art methods, our approach improves average precision by  $5.55\\%$  and accuracy by  $9.33\\%$ . For other generative models (GANs), VIB-Net achieves gains of  $12.48\\%$  in average precision and",
+        "bbox": [
+            89,
+            763,
+            482,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "23.59% in accuracy. This is due to VIB-Net's ability to retain compact deep features while eliminating irrelevant surface features, enhancing generalization. Overall, these results highlight the advantages of our method.",
+        "bbox": [
+            511,
+            539,
+            906,
+            599
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3. Performance evaluation based on GAN sources",
+        "text_level": 1,
+        "bbox": [
+            511,
+            606,
+            903,
+            622
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "So far, we have explored the generalization performance of using images generated by specific diffusion models as training data. Next, we trained using images generated by ProGAN and then tested on 18 different generative models to explore the performance of the detector on unseen generative models. The results are shown in Tabs. 3 and 4. From the tables, we can see that almost all models perform well when detecting images generated by other GAN variants. This is because, compared to diffusion-based generation methods, images generated by these GAN variants are visually more distinguishable, making it easier for the model to learn the distinguishing features of real and fake. However, when detecting images generated by models other than GAN variants, the performance of most models sharply declines. This suggests that even when the training data contains rich features related to real/fake classification, it does not necessarily ensure better generalization performance. This is because the model may overfit the incidental correla",
+        "bbox": [
+            509,
+            628,
+            906,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "23834",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/03c6eff0437e82330a33e46b3b5553ef28669794ee5a81cf45ccb08c1d454389.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            88,
+            903,
+            188
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/446f4f542fe26f5297875262a6a6efa70423d50c08a5ff47202b7b01c2126714.jpg",
+        "image_caption": [
+            "Figure 5. Ablation study on the network architecture. The  $AP$  and  $ACC$  values of different VIB-Net variants on various models. CLIP:ViT-L/14 shows better performance."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            195,
+            903,
+            296
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "tions between irrelevant features and the task, which harms generalization performance. Univfd performs better across model families, but since it extracts general features, many of these features are unrelated to the real/fake classification task, which naturally affects the decision boundary.",
+        "bbox": [
+            89,
+            354,
+            482,
+            430
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In contrast, our method shows better performance in detecting real and fake images. Compared to the baseline Univfd, our method achieved improvements of  $2.26\\%$  and  $6.19\\%$  in average precision and accuracy. In OOD scenarios (excluding GANs), our method outperforms the baseline Univfd by  $3.22\\%$  and  $8.48\\%$ . Compared with advanced methods, our method achieved improvements of  $10.7\\%$  to  $23.08\\%$  in average precision. These results further validate the effectiveness of our proposed method.",
+        "bbox": [
+            89,
+            430,
+            483,
+            566
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4. Effect of VIB-Net across different backbone",
+        "text_level": 1,
+        "bbox": [
+            89,
+            575,
+            464,
+            590
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "So far, we have seen the impressive performance of VIB-Net with the CLIP:ViT-L/14 backbone. In this section, we will explore the effect of transferring the VIB component to other multimodal models. We consider the following backbone variants to form VIB-Net: (1) FLAVA-FULL[35] (2) CLIP:ResNet50 (3) CLIP:ViT-B/16 (4) CLIP:ViT-L/14[30]. For each VIB-Net variant, we train using images generated by Stable Diffusion v1.4 and corresponding real images. Fig. 5 show the AP and ACC results of these variants on the same generative model. We observe that among these multimodal models, ViT-L/14 achieves the best performance, which can be attributed to its larger architecture and attention-based global receptive field. More importantly, all these variants exhibit strong performance, which can be attributed to the unique capabilities of our method.",
+        "bbox": [
+            89,
+            597,
+            482,
+            824
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.5. Analysis of latent features",
+        "text_level": 1,
+        "bbox": [
+            89,
+            833,
+            326,
+            849
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Finally, we apply our method to address the issues analyzed in Sec. 3.2. In Sec. 3.2, we identified that some irrelevant features in CLIP's general features could affect the classi-",
+        "bbox": [
+            89,
+            854,
+            483,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "fier's decision boundary, causing the model to learn superficial biases, which in turn impacts its generalization performance. As a solution, we proposed VIB-Net. Following the same experimental setup as in Sec. 3.2, we visualized the latent features extracted by VIB-Net using UMAP, and obtained encouraging results, as shown in Fig. 3b. We found that the latent features no longer exhibit obvious irrelevant features related to the real/fake image classification task, as the two separated clusters now belong to the same category (real or generated), unlike in Fig. 3a, where features of the same category were significantly separated. This demonstrates the advantage of our method.",
+        "bbox": [
+            511,
+            354,
+            906,
+            536
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            511,
+            546,
+            633,
+            561
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We propose a universal real-fake image detection method, VIB-Net. First, we found that the general features extracted by current methods based on large-scale pre-trained models contain irrelevant features that are unrelated to the task of distinguishing real from fake images. Experiments have shown that these features not only affect classification decisions, but may also cause the model to learn superficial biases, thereby damaging its generalization performance. As an alternative, we use VIB to remove task-irrelevant features, aiming to obtain the most concise image representations. These simplified features are maximally correlated with the real-fake classification task.",
+        "bbox": [
+            511,
+            571,
+            908,
+            752
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            511,
+            763,
+            681,
+            781
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 62376046 and Grant 62172067, and in part by the Natural Science Foundation of Chongqing under Grant CSTB2023NSCQ-MSX0341, the Science and Technology Research Program of Chongqing Municipal Education Commission (Grant No. KJQN202200635).",
+        "bbox": [
+            511,
+            787,
+            908,
+            887
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "23835",
+        "bbox": [
+            478,
+            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] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 2, 3",
+            "[2] Alexander A Alemi, Ian Fischer, and Joshua V Dillon. Uncertainty in the variational information bottleneck. arXiv preprint arXiv:1807.00906, 2018. 3",
+            "[3] Xiuli Bi, Bo Liu, Fan Yang, Bin Xiao, Weisheng Li, Gao Huang, and Pamela C Cosman. Detecting generated images by real images only. arXiv preprint arXiv:2311.00962, 2023. 2",
+            "[4] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 103-120. Springer, 2020.",
+            "[5] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 5",
+            "[6] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of ai-generated image detection with clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4356-4366, 2024. 1, 2, 3",
+            "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5",
+            "[8] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 6, 7",
+            "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. 1",
+            "[10] 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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 5",
+            "[11] 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. 2, 3",
+            "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1",
+            "[13] Shizhe Hu, Zhengzheng Lou, Xiaoqiang Yan, and Yangdong Ye. A survey on information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3"
+        ],
+        "bbox": [
+            93,
+            114,
+            483,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[14] Yan Ju, Shan Jia, Lipeng Ke, Hongfei Xue, Koki Nagano, and Siwei Lyu. Fusing global and local features for generalized ai-synthesized image detection. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3465-3469. IEEE, 2022. 2, 6, 7",
+            "[15] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 1, 5",
+            "[16] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5",
+            "[17] Sohail Ahmed Khan and Duc-Tien Dang-Nguyen. Clipping the deception: Adapting vision-language models for universal deepfake detection. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 1006-1015, 2024. 6, 7",
+            "[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5",
+            "[19] Li Lin, Irene Amerini, Xin Wang, Shu Hu, et al. Robust clip-based detector for exposing diffusion model-generated images. arXiv preprint arXiv:2404.12908, 2024. 2",
+            "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, 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",
+            "[21] Bo Liu, Fan Yang, Xiuli Bi, Bin Xiao, Weisheng Li, and Xinbo Gao. Detecting generated images by real images. In European Conference on Computer Vision, pages 95-110. Springer, 2022. 2",
+            "[22] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5",
+            "[23] Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. Lare^ 2: Latent reconstruction error based method for diffusion-generated image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17006-17015, 2024. 1",
+            "[24] Rabeeh Karimi Mahabadi, Yonatan Belinkov, and James Henderson. Variational information bottleneck for effective low-resource fine-tuning. arXiv preprint arXiv:2106.05469, 2021.3",
+            "[25] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 1",
+            "[26] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 3",
+            "[27] 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 preprint arXiv:2112.10741, 2021. 5"
+        ],
+        "bbox": [
+            516,
+            92,
+            906,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "23836",
+        "bbox": [
+            478,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[28] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 2, 3, 6, 7",
+            "[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 5",
+            "[30] 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, 8",
+            "[31] 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",
+            "[32] 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",
+            "[33] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 5",
+            "[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5",
+            "[35] 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, pages 15638-15650, 2022. 8",
+            "[36] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 6, 7",
+            "[37] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 6, 7",
+            "[38] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3"
+        ],
+        "bbox": [
+            91,
+            90,
+            480,
+            901
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[39] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 1, 2, 5, 6, 7",
+            "[40] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 1, 6, 7",
+            "[41] Among Wu and Cheng Deng. Tib: Detecting unknown objects via two-stream information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3",
+            "[42] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5",
+            "[43] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 1",
+            "[44] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021. 2",
+            "[45] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 5",
+            "[46] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. Advances in Neural Information Processing Systems, 36, 2024. 5"
+        ],
+        "bbox": [
+            516,
+            90,
+            906,
+            599
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "23837",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/c23bce64-f04c-4e24-b700-3d6473e34043_model.json

@@ -0,0 +1,1859 @@
+[
+    [
+        {
+            "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": [
+                0.108,
+                0.13,
+                0.892,
+                0.175
+            ],
+            "angle": 0,
+            "content": "Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.21,
+                0.204,
+                0.787,
+                0.24
+            ],
+            "angle": 0,
+            "content": "Haifeng Zhang, Qinghui He, Xiuli Bi, Weisheng Li, Bo Liu*, Bin Xiao Chongqing University of Posts and Telecommunications"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.25,
+                0.242,
+                0.748,
+                0.257
+            ],
+            "angle": 0,
+            "content": "s220331143@stu.cqupt.edu.cn, s220232014@stu.cqupt.edu.cn"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.315,
+                0.259,
+                0.678,
+                0.275
+            ],
+            "angle": 0,
+            "content": "{bix1, liws, boliu, xiaobin}@cqupt.edu.cn"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.248,
+                0.309,
+                0.327,
+                0.327
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.341,
+                0.486,
+                0.675
+            ],
+            "angle": 0,
+            "content": "The rapid advancement of generative models has significantly improved the quality of generated images. Meanwhile, it challenges information authenticity and credibility. Current generated image detection methods based on large-scale pre-trained multimodal models have achieved impressive results. Although these models provide abundant features, the authentication task-related features are often submerged. Consequently, those authentication task-irrelated features cause models to learn superficial biases, thereby harming their generalization performance across different model genera (e.g., GANs and Diffusion Models). To this end, we proposed VIB-Net, which uses Variational Information Bottlenecks to enforce authentication task-related feature learning. We tested and analyzed the proposed method and existing methods on samples generated by 17 different generative models. Compared to SOTA methods, VIB-Net achieved a \\(5.55\\%\\) improvement in mAP and a \\(9.33\\%\\) increase in accuracy. Notably, in generalization tests on unseen generative models from different series, VIB-Net improved mAP by \\(12.48\\%\\) and accuracy by \\(23.59\\%\\) over SOTA methods. The code is available at https://github.com/oceanzhf/VIBAIGCDetect."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.092,
+                0.702,
+                0.222,
+                0.718
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.728,
+                0.485,
+                0.879
+            ],
+            "angle": 0,
+            "content": "The continuous advancements in deep learning have led to the emergence of numerous well-known image generation models, such as GANs[9][15] and Diffusion[12][31]. Images generated by these models have reached a level where observers can hardly distinguish them from real images, and the flexibility of tools based on these models has also greatly improved [6]. With a few simple steps, people can easily obtain high-quality images as needed. While this presents a tremendous opportunity for fields like visual arts, it also serves as a paradise for individuals looking to fab"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.517,
+                0.309,
+                0.906,
+                0.487
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.512,
+                0.498,
+                0.909,
+                0.554
+            ],
+            "angle": 0,
+            "content": "Figure 1. Each method's overall performance('P') on test data and generalization performance ('G') on unseen data are linked by dashed lines. Our detector demonstrates a clear advantage over all SOTA methods in both metrics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.582,
+                0.907,
+                0.657
+            ],
+            "angle": 0,
+            "content": "ricate misinformation. Given the ease of creating manipulated images today and the potential risk of using them for malicious purposes, there is an urgent need to develop detectors that help distinguish between real and generated images."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.658,
+                0.909,
+                0.885
+            ],
+            "angle": 0,
+            "content": "Existing detection methods [4][39][25][43] focus on identifying specific traces left by generators in either the spatial or frequency domain of generated images, thus establishing a decision boundary between real and generated images. These detection methods are highly effective in detecting images generated by GANs and earlier diffusion models. However, their performance drops significantly when facing images generated by recent, popular models that visually resemble real images. Although some error reconstruction-based methods [40][23] have been developed specifically for detecting images generated by diffusion models, they similarly experience performance declines when applied to GAN-generated images. Consequently, current research increasingly focuses on improving the cross-model genera detection capabilities of detectors."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.886,
+                0.905,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Generalization across different domains is a primary re"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.109,
+                0.888,
+                0.242,
+                0.901
+            ],
+            "angle": 0,
+            "content": "* Corresponding Author"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "23828"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.094,
+                0.092,
+                0.482,
+                0.573
+            ],
+            "angle": 0,
+            "content": "search challenge in deep learning. In this context, the emergence of large-scale pre-trained multimodal models has introduced new approaches and promising results. For instance, by using the image encoder from the CLIP [30] model to extract image features and training a classifier with linear probing, there is a notable improvement in generalization performance over models like ResNet50[28][6]. However, compared to traditional classification tasks, real and generated images share subtle, harder-to-capture distinguishing features. This implies that the features extracted by large-scale pre-trained models for the task of distinguishing real from fake images may contain a considerable amount of task-irrelevant features. Task-irrelevant features can either have no impact on the realizness of the image or, though effective in isolation, become redundant when combined with other features. These features not only interfere with the classifier's ability to accurately learn decision boundaries but also hinder the model's ability to capture and distinguish task-relevant characteristics. More critically, task-irrelevant features may lead the model to learn superficial biases. For example, if a fake image contains a tree while the real image does not, the model might erroneously associate the presence of the tree with being a fake image. Consequently, when detecting other images, the model may mistakenly classify a real image as fake if it contains a tree. Therefore, when dealing with high-quality generated images with subtle distinguishing features, it is challenging to fine-tune the model to separate essential from unnecessary information. This can lead to an overfitting of the incidental statistical correlations between unnecessary information and target images, ultimately undermining the model's generalization and discrimination capabilities."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.094,
+                0.585,
+                0.482,
+                0.9
+            ],
+            "angle": 0,
+            "content": "How can we harness the powerful feature extraction capability of large models while minimizing the interference of unnecessary information? By approaching this from a novel perspective, we propose addressing the issue using Information Bottleneck Theory. This theory aims to create latent encodings that not only effectively represent the target but also maximally compress the original input, thereby avoiding the impact of task-irrelevant features on the model. Specifically, we introduce a method called VIB-Net, which couples a large-scale pre-trained multimodal model with Variational Information Bottleneck (VIB) [1]. Considering that feature representations extracted by the model's image encoder are more task-specific, we decouple and filter these features using latent variables. This method maximizes the mutual information between the latent feature and the classification output, while minimizing the mutual information between the latent feature and the input, resulting in a concise feature that should only contain the effective information needed for the task. This concise latent feature helps the model avoid overfitting to irrelevant features and the incidental correlations with the task, preventing the model"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.517,
+                0.093,
+                0.904,
+                0.121
+            ],
+            "angle": 0,
+            "content": "from learning superficial biases, which is beneficial to the model's generalization performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.517,
+                0.123,
+                0.904,
+                0.362
+            ],
+            "angle": 0,
+            "content": "In summary, our contributions are as follows: (1) To the best of our knowledge, we are the first to apply Information Bottleneck theory and techniques to the detection of generated images, demonstrating the potential of this theory and approach in this field. (2) We analyze the limitations of methods based on large-scale pre-trained models in detecting images generated by unseen generative models, and present our theory of what could be wrong with the existing paradigm. (3) We propose VIB-Net for the detection of generated images, which uses VIB to decouple and filter the general features extracted by large-scale pre-trained multimodal models, removing task-irrelevant information to enhance detection performance. (4) A large number of experiments validate the advantages of our method (see Fig. 1), especially in the improvement of generalization performance \\((+23.59\\%\\) accuracy)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.517,
+                0.378,
+                0.649,
+                0.393
+            ],
+            "angle": 0,
+            "content": "2. Related work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.517,
+                0.403,
+                0.904,
+                0.719
+            ],
+            "angle": 0,
+            "content": "Generalization Performance. The continuous advancement of generated image models presents ongoing challenges for generated image detection, making the generalization performance of detection models on unseen generators an increasing focus for researchers[14][21]. [39] explored the effects of data preprocessing and augmentation methods on the generalization performance of a ResNet50[11]-based detector. [44] argues that the differences between real and fake images are subtle and localized, framing forgery detection as a fine-grained task that could positively impact generalization performance. [3] takes a frequency-domain approach, mapping real images to a dense subspace and synthetic images outside this subspace to enhance generalization. While these methods have achieved certain successes on different models within the same generation series, they show limitations across different series. This is due to the fact that, with the development of generative algorithms, subtle traces in generated images are becoming increasingly difficult to capture. Additionally, artifacts in the frequency domain differ between model series and are gradually decreasing in visibility."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.517,
+                0.721,
+                0.904,
+                0.9
+            ],
+            "angle": 0,
+            "content": "The impressive results of large-scale pre-trained models in terms of transferability and generalization have sparked significant interest among researchers in the field of generated image detection. [28] argues that existing detectors perform asymmetric learning, where they use CLIP's image encoder to extract image features, exploring the potential of large models in detecting different model families. [6] suggests that detectors based on large models rely on higher-level semantic features, which enhances the generalization ability of the detectors. [19] uses both image and text features extracted by the CLIP model, and then employs a multilayer perceptron to detect images generated by"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.518,
+                0.957
+            ],
+            "angle": 0,
+            "content": "23829"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.092,
+                0.486,
+                0.288
+            ],
+            "angle": 0,
+            "content": "diffusion models, improving generalization performance. Information Bottleneck. The Information Bottleneck (IB) [38] is an innovative information-theoretic method used for pattern analysis and representation learning. It provides an elegant balance between data compression and information retention[13]. VIB [1] is a variational approximation of IB, aimed at parameterizing IB using neural networks. [2] demonstrated VIB's capability in improving classification calibration and out-of-distribution data. [24] explored VIB as a regularization technique to address overfitting issues in NLP models. [41] proposed a dual-stream information bottleneck method to utilize in-distribution data to enhance out-of-distribution target detection tasks."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.29,
+                0.485,
+                0.471
+            ],
+            "angle": 0,
+            "content": "Based on the above analysis, current methods for detecting generated images using large-scale pre-trained models simply utilize the general representations extracted from images, without thoroughly exploring the relationship between these general features and downstream tasks. Additionally, Information Bottleneck is an effective technique for improving performance in cross-domain and cross-distribution tasks. Therefore, unlike existing methods, our approach leverages Information Bottleneck techniques to explore the potential of large-scale pre-trained models in generated image detection from a completely new perspective, achieving outstanding performance."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.486,
+                0.231,
+                0.501
+            ],
+            "angle": 0,
+            "content": "3. Preliminaries"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.512,
+                0.291,
+                0.527
+            ],
+            "angle": 0,
+            "content": "3.1. Problem background"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.534,
+                0.484,
+                0.7
+            ],
+            "angle": 0,
+            "content": "A common paradigm for using large-scale pretrained models is to leverage a large-scale pretrained multimodal model (e.g., CLIP) to learn the general representation of real and generated images, and then apply this feature representation to downstream tasks[28][6]. Building on their findings, we further consider the following points: (1) It is well-known that the features extracted by large-scale pre-trained visual encoders are general-purpose, meaning they contain many general features, such as semantic information. (2) Another widely accepted notion is that some critical features for distinguishing real from generated images are subtle and local."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.701,
+                0.484,
+                0.837
+            ],
+            "angle": 0,
+            "content": "Further combining these two points, we propose the following hypotheses: (i) The general features extracted by large-scale pre-trained multimodal models are more beneficial for real-fake image classification tasks than those extracted by pre-trained visual models. (ii) These general features contain task-irrelevant features for real-fake classification. (iii) These task-irrelevant features affect the classification of real and fake images. Next, we will delve into these three hypotheses."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.848,
+                0.449,
+                0.865
+            ],
+            "angle": 0,
+            "content": "3.2. Analysis of general feature representation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.871,
+                0.484,
+                0.902
+            ],
+            "angle": 0,
+            "content": "For the first hypothesis, We use image pairs of real and generated images to aid our analysis[6], generating textual de"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.522,
+                0.09,
+                0.906,
+                0.309
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.513,
+                0.32,
+                0.907,
+                0.362
+            ],
+            "angle": 0,
+            "content": "Figure 2. UMAP visualization of features extracted from Dalle2 and Dalle3 generated images using large-scale pretrained multimodal(CLIP) and unimodal models(ResNet)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.398,
+                0.907,
+                0.61
+            ],
+            "angle": 0,
+            "content": "scriptions for 1,000 real images and using these descriptions to produce 1,000 corresponding images with DALL-E 2 and DALL-E 3. We then compare features extracted by large-scale pre-trained multimodal models(CLIP[30]) and pre-trained visual models(ResNet[11]) through UMAP [26] visualization of 3,000 images. The results show that features extracted by multimodal models (e.g., CLIP) exhibit greater distribution differences for real images and DALL-E 2-generated images, suggesting multimodal models may be better at distinguishing real from generated images. However, for DALL-E 3-generated images, both model types show similar feature distributions, possibly due to advancements in image generation tools, which complicate the classification task and require further fine-tuning."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.614,
+                0.909,
+                0.901
+            ],
+            "angle": 0,
+            "content": "To further explore hypotheses (ii) and (iii), we visualized the extracted features using both category and authenticity as labels to observe their distribution differences. Specifically, we prepared images of two categories (car and cat), each containing 500 real images and 500 generated images. We then used CLIP to extract their features and performed UMAP dimensionality reduction for visualization, with the results shown in Fig. 3a. We can see that the features extracted by CLIP clearly form two clusters: the left cluster mainly consists of light blue and light red points, while the right cluster is mainly composed of light pink and light purple points. These two clusters correspond to category 1 and category 2, respectively. This indicates that CLIP can significantly distinguish category information. However, this clear separation of category information may affect the decision-making process for classifying real vs. generated images. When examining the distinction between real and generated images, we notice that the right cluster shows a noticeable separation of real and generated features, with"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.945,
+                0.52,
+                0.957
+            ],
+            "angle": 0,
+            "content": "23830"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.093,
+                0.09,
+                0.285,
+                0.213
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.093,
+                0.216,
+                0.285,
+                0.24
+            ],
+            "angle": 0,
+            "content": "(a) Visualization of features extracted by CLIP."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.287,
+                0.091,
+                0.482,
+                0.213
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.289,
+                0.216,
+                0.483,
+                0.24
+            ],
+            "angle": 0,
+            "content": "(b) Visualization of features extracted by VIB-Net."
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.09,
+                0.252,
+                0.483,
+                0.281
+            ],
+            "angle": 0,
+            "content": "Figure 3. UMAP visualization of features extracted from real and generated images of two category using CLIP and VIB-Net."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.314,
+                0.483,
+                0.525
+            ],
+            "angle": 0,
+            "content": "two different colors observed within the cluster. This suggests that CLIP is capable of distinguishing real and generated features within category 2. However, in the left cluster, CLIP does not show significant ability to differentiate between real and generated features. We observe that in the left cluster (category 1), the two colors are mixed together, showing poor distinguishability. From these observations, it is easy to infer that the internal separation of real or generated images caused by semantic category information makes it more difficult for the classifier to learn the decision boundary. More importantly, the strong sensitivity of large-scale pre-trained multimodal models like CLIP to semantic information leads the model to tend to learn superficial biases."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.529,
+                0.483,
+                0.68
+            ],
+            "angle": 0,
+            "content": "Based on this viewpoint, we consider whether it is possible to minimize the irrelevant features within these multimodal features while retaining the ones that are relevant to the classification task. This would help the model avoid learning superficial biases, allow the classifier to more easily learn decision boundaries, and thereby enhance the model's discrimination ability and generalization capability. This is the motivation of this paper. As shown in Fig. 3b, the method(VIB-Net) proposed in this paper achieves significant results."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.09,
+                0.706,
+                0.2,
+                0.723
+            ],
+            "angle": 0,
+            "content": "4. Approach"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.735,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Based on the issues analyzed in Sec. 3, an effective approach is to filter the multimodal features in order to obtain the representations most beneficial for the real-fake classification task, which serves as the starting point of this paper. Overall, our method is illustrated in Fig. 4. VIB attempts to filter out irrelevant features and to find the most concise representation, making it easier for the classifier to learn the features needed for distinguishing between real and fake tasks, thereby significantly improving the accuracy of the classification task. Below, we will provide a detailed description of our method."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.091,
+                0.81,
+                0.106
+            ],
+            "angle": 0,
+            "content": "4.1. Information bottleneck constraint"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.113,
+                0.907,
+                0.264
+            ],
+            "angle": 0,
+            "content": "To specifically optimize the filtering of irrelevant information in the input features while retaining task-relevant information, we apply the information bottleneck (IB) principle as a constraint on the features to be optimized. Let \\( X \\), \\( Z \\), and \\( Y \\) represent the input, the latent compressed representation, and the output target, respectively. The goal of IB is to find the maximum compressed representation \\( Z \\) of the input \\( X \\), while preserving as much information about the output \\( Y \\) as possible. The maximization objective function is defined as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.611,
+                0.273,
+                0.907,
+                0.29
+            ],
+            "angle": 0,
+            "content": "\\[\nL _ {I B} = I (Z, Y) - \\beta I (Z, X) \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.297,
+                0.907,
+                0.373
+            ],
+            "angle": 0,
+            "content": "Where \\( I(.,.) \\) denotes mutual information and \\( \\beta \\) is a trade-off parameter. The first term constrains the latent compressed representation to retain sufficient information for downstream tasks, while the second term ensures that irrelevant and redundant information is filtered out."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.381,
+                0.816,
+                0.395
+            ],
+            "angle": 0,
+            "content": "4.2. Variational information bottleneck"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.403,
+                0.906,
+                0.464
+            ],
+            "angle": 0,
+            "content": "To parameterize the information bottleneck (IB) using deep neural networks, we perform variational inference on Equation (1). First, based on the definition of mutual information, Equation (1) can be rewritten as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.591,
+                0.483,
+                0.907,
+                0.551
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} L _ {I B} = \\int d y d z p (y, z) \\log {\\frac {p (y | z)}{p (y)}} \\\\ - \\beta \\int d x d z p (x, z) \\log \\frac {p (z \\mid x)}{p (z)} \\tag {2} \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.558,
+                0.906,
+                0.587
+            ],
+            "angle": 0,
+            "content": "Where \\( p(y|z) \\) is defined by the encoder and the Markov chain as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.58,
+                0.593,
+                0.906,
+                0.626
+            ],
+            "angle": 0,
+            "content": "\\[\np (y \\mid z) = \\int d x \\frac {p (y \\mid x) p (z \\mid x) p (x)}{p (z)} \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.633,
+                0.907,
+                0.769
+            ],
+            "angle": 0,
+            "content": "Since it is not computable here, we consider setting \\( q(y|z) \\) as a parameterized approximation of \\( p(y|z) \\), serving as a \"decoder,\" which can be approximated by a linear layer. We introduce the Kullback-Leibler (KL) divergence to obtain an estimizable representation. As the KL divergence satisfies \\( \\mathrm{KL}(p(y|z),q(y|z))\\geq 0 \\), and by applying the Markov assumption, \\( p(y,z) \\) can be rewritten as \\( p(y,z) = \\int dx p(x)p(y|x)p(z|x) \\), thereby providing a lower bound for the first term of the objective function:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.526,
+                0.776,
+                0.905,
+                0.819
+            ],
+            "angle": 0,
+            "content": "\\[\nI (Z, Y) \\geq \\int d x d y d z p (x) p (y | x) p (z | x) \\log q (y | z) \\tag {4}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.82,
+                0.905,
+                0.865
+            ],
+            "angle": 0,
+            "content": "At the same time, let \\( r(z) \\) be an estimate of the prior probability \\( p(z) \\), used to approximate the marginal distribution, thereby obtaining an upper bound for the second term:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.557,
+                0.872,
+                0.905,
+                0.906
+            ],
+            "angle": 0,
+            "content": "\\[\nI (Z, X) \\leq \\int d x d z p (x) p (z | x) \\log \\frac {p (z | x)}{r (z)} \\tag {5}\n\\]"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.518,
+                0.957
+            ],
+            "angle": 0,
+            "content": "23831"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.091,
+                0.089,
+                0.905,
+                0.302
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.089,
+                0.311,
+                0.908,
+                0.355
+            ],
+            "angle": 0,
+            "content": "Figure 4. Overview of VIB-Net. We utilize Variational Information Bottleneck (VIB) techniques to map the image features extracted by CLIP's visual encoder into latent features \\( z \\), where the information in \\( z \\) is filtered based on the Information Bottleneck (IB). Subsequently, the latent features \\( z \\) are used as input to an MLP to learn the decision boundary."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.38,
+                0.484,
+                0.518
+            ],
+            "angle": 0,
+            "content": "The encoder we use is defined as \\( p(z|x) = p_{\\theta}(z|x) = \\mathcal{N}\\left(z|f^{\\mu}(x), f^{\\Sigma}(x)\\right) \\), where \\( f(.) \\) represents a shallow MLP, \\( \\mu \\) denotes the K-dimensional mean vector of the output \\( z \\), and \\( \\Sigma \\) represents the \\( K \\times K \\) covariance matrix. We select a parameterized Gaussian distribution for both \\( p_{\\theta}(z|x) \\) and \\( r(z) \\) to allow for the analytical computation of KL divergence. Additionally, we employ the reparameterization trick to estimate gradients [18]. Combining all the components above yields the following objective function:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.14,
+                0.538,
+                0.483,
+                0.6
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} L _ {V I B} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} z \\sim p _ {\\theta} (z | x) [ - \\log q (y _ {\\mathrm {n}} | z _ {n}) ] \\\\ + \\beta \\mathrm {K L} [ p _ {\\theta} (z | x _ {n}), r (z) ] \\tag {6} \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.09,
+                0.608,
+                0.373,
+                0.623
+            ],
+            "angle": 0,
+            "content": "4.3. Workflow of the overall method"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.63,
+                0.484,
+                0.812
+            ],
+            "angle": 0,
+            "content": "As shown in Fig. 4, the parts in VIB-Net work in the following way: First, the real and generated images \\(\\{x_{1}, x_{2}, x_{3}, \\dots, x_{n}\\}\\) are input into the CLIP image encoder, yielding the corresponding image features \\(\\{t_{1}, t_{2}, t_{3}, \\dots, t_{n}\\}\\). Next, these feature vectors are passed into the Variational Information Bottleneck (VIB). Specifically, the image features are processed through a shallow MLP, then a linear layer \\(f(.)\\) calculates \\(\\mu\\) and \\(\\Sigma\\), which are transformed using a softmax activation. This layer has \\(K\\) dimensions, representing the bottleneck size. At this point, we obtain the latent compressed feature \\(z\\), which is then fed into a linear layer \\(q(y|z)\\) to output the target task label \\(y\\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.09,
+                0.824,
+                0.224,
+                0.841
+            ],
+            "angle": 0,
+            "content": "5. Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.849,
+                0.191,
+                0.866
+            ],
+            "angle": 0,
+            "content": "5.1. Settings"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.871,
+                0.484,
+                0.902
+            ],
+            "angle": 0,
+            "content": "To better evaluate the model's performance, we utilize two mainstream public datasets: ForenSynths[39] and"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.38,
+                0.907,
+                0.53
+            ],
+            "angle": 0,
+            "content": "Genimage[46]. The ForenSynths dataset primarily consists of images generated by GANs, while the Genimage dataset mainly includes images generated by diffusion models. Consequently, we set up two training and testing schemes: (1) Training on ForenSynths and then testing on both ForenSynths and Genimage, and (2) similarly, training on Genimage and testing on both Genimage and ForenSynths. This setup allows us to more effectively validate the generalization performance of the detector on cross model genera scenarios. The details are as follows:"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.533,
+                0.609,
+                0.547
+            ],
+            "angle": 0,
+            "content": "Training set:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.548,
+                0.906,
+                0.669
+            ],
+            "angle": 0,
+            "content": "The ForenSynths training set consists of 20 different categories, each containing 18,000 real images from the LSUN dataset [42] and an equal number of images generated using ProGAN[15]. Following the setup in Genimage, we use images generated by Stable Diffusion v1.4 [32] in the Genimage dataset, along with their corresponding real images, as the training set. In total, these include 32,000 real and generated images."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.671,
+                0.598,
+                0.685
+            ],
+            "angle": 0,
+            "content": "Testing set:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.687,
+                0.906,
+                0.778
+            ],
+            "angle": 0,
+            "content": "11 models from ForenSynths The test set includes images generated by 11 different models, primarily from the GAN family[15][45][29][5][16]. The real images are sourced from six datasets, including ImageNet[34], LSUN[42], FaceForensics++[33], CelebA-HQ[15], CelebA[22], and COCO[20]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.779,
+                0.907,
+                0.87
+            ],
+            "angle": 0,
+            "content": "7 models from Genimage To validate the method across different generation models, we use images generated by seven advanced generators along with their corresponding real images as part of the test set. These seven generation models include Stable Diffusion v1.4/1.5[32], VQDM[10], ADM[7], GLIDE[27], Wukong, and Midjourney."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.871,
+                0.906,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Real-vs-Fake classification baselines: We compare VIB-Net with existing state-of-the-art detection methods, in"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "23832"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.095,
+                0.089,
+                0.907,
+                0.259
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Detection method</td><td colspan=\"7\">Diffusion Models</td><td colspan=\"5\">Generative Adversarial Networks</td><td colspan=\"2\">Others</td><td>AP</td><td></td></tr><tr><td>SDV1.4</td><td>SDV1.5</td><td>ADM</td><td>GLIDE</td><td>Midjourney</td><td>Wukong</td><td>VQDM</td><td>Pro-GAN</td><td>Cycle-Big-GAN</td><td>Style-GAN</td><td>Star-GAN</td><td>Gau-GAN</td><td colspan=\"2\">Deep-fake</td><td>Average</td><td></td></tr><tr><td>CNNSpot[39]</td><td>99.98</td><td>99.83</td><td>51.10</td><td>58.80</td><td>67.93</td><td>99.80</td><td>49.92</td><td>53.15</td><td>50.23</td><td>49.79</td><td>55.98</td><td>47.07</td><td>56.08</td><td>54.86</td><td>54.03</td><td>63.24</td></tr><tr><td>Fusing[14]</td><td>99.90</td><td>97.98</td><td>69.30</td><td>94.20</td><td>81.20</td><td>99.90</td><td>84.60</td><td>67.63</td><td>87.79</td><td>69.37</td><td>67.90</td><td>91.20</td><td>43.08</td><td>72.56</td><td>89.42</td><td>81.07</td></tr><tr><td>Lgrad[36]</td><td>99.94</td><td>99.92</td><td>58.52</td><td>84.00</td><td>91.06</td><td>99.72</td><td>56.34</td><td>83.59</td><td>90.24</td><td>47.51</td><td>82.74</td><td>99.19</td><td>49.25</td><td>66.49</td><td>65.09</td><td>78.24</td></tr><tr><td>Univfd[28]</td><td>96.04</td><td>96.26</td><td>66.34</td><td>93.73</td><td>92.08</td><td>90.98</td><td>74.53</td><td>51.77</td><td>63.42</td><td>75.81</td><td>54.12</td><td>54.93</td><td>65.99</td><td>70.24</td><td>83.34</td><td>75.31</td></tr><tr><td>NPR[37]</td><td>100.00</td><td>99.97</td><td>94.70</td><td>95.80</td><td>95.50</td><td>100.00</td><td>86.30</td><td>83.30</td><td>94.90</td><td>72.00</td><td>82.70</td><td>97.30</td><td>66.00</td><td>85.30</td><td>95.90</td><td>89.98</td></tr><tr><td>CLIPping[17]</td><td>93.97</td><td>93.10</td><td>68.00</td><td>87.44</td><td>77.34</td><td>86.52</td><td>77.17</td><td>88.54</td><td>88.44</td><td>85.33</td><td>77.23</td><td>89.82</td><td>81.56</td><td>61.19</td><td>57.37</td><td>80.87</td></tr><tr><td>Ours</td><td>100.00</td><td>99.97</td><td>95.49</td><td>97.13</td><td>97.81</td><td>99.93</td><td>97.00</td><td>96.59</td><td>98.44</td><td>97.17</td><td>84.31</td><td>97.60</td><td>96.94</td><td>81.32</td><td>93.27</td><td>95.53</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.09,
+                0.268,
+                0.907,
+                0.298
+            ],
+            "angle": 0,
+            "content": "Table 1. The \\(AP\\) values of different methods based on diffusion sources. Models outside the Diffusion Models (DMs) column are considered as the generalization domain across generative model series."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.095,
+                0.31,
+                0.907,
+                0.482
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Detection method</td><td colspan=\"7\">Diffusion Models</td><td colspan=\"5\">Generative Adversarial Networks</td><td>Others</td><td>ACC</td><td></td></tr><tr><td>SDV1.4</td><td>SDV1.5</td><td>ADM</td><td>GLIDE</td><td>Midjourney</td><td>Wukong</td><td>VQDM</td><td>Pro-GAN</td><td>Cycle-Big-GAN</td><td>Style-GAN</td><td>Star-GAN</td><td>Gau-GAN</td><td>Deep-SAN</td><td>Average</td><td></td></tr><tr><td>CNNSpot[39]</td><td>99.48</td><td>99.35</td><td>50.10</td><td>50.90</td><td>56.42</td><td>97.90</td><td>50.04</td><td>50.27</td><td>49.81</td><td>50.10</td><td>50.98</td><td>49.77</td><td>50.38</td><td>51.98</td><td>50.22</td></tr><tr><td>Fusing[14]</td><td>99.90</td><td>99.91</td><td>51.30</td><td>57.50</td><td>52.30</td><td>99.90</td><td>64.20</td><td>51.20</td><td>52.40</td><td>53.50</td><td>50.20</td><td>58.20</td><td>49.32</td><td>51.02</td><td>64.84</td></tr><tr><td>Lgrad[36]</td><td>99.12</td><td>99.05</td><td>53.00</td><td>64.24</td><td>76.34</td><td>97.53</td><td>50.93</td><td>61.61</td><td>60.74</td><td>48.82</td><td>61.43</td><td>50.17</td><td>49.70</td><td>50.17</td><td>56.49</td></tr><tr><td>Univfd[28]</td><td>83.55</td><td>84.80</td><td>53.35</td><td>75.30</td><td>71.60</td><td>73.55</td><td>55.10</td><td>58.65</td><td>59.30</td><td>61.45</td><td>56.80</td><td>61.45</td><td>55.30</td><td>58.40</td><td>72.00</td></tr><tr><td>NPR[37]</td><td>100.00</td><td>99.90</td><td>73.00</td><td>89.70</td><td>82.30</td><td>100.00</td><td>68.30</td><td>60.30</td><td>67.20</td><td>59.20</td><td>58.00</td><td>73.20</td><td>52.00</td><td>74.80</td><td>89.60</td></tr><tr><td>CLIPping[17]</td><td>96.07</td><td>95.48</td><td>70.14</td><td>85.00</td><td>77.66</td><td>88.86</td><td>79.35</td><td>88.78</td><td>88.48</td><td>89.57</td><td>80.69</td><td>90.82</td><td>85.94</td><td>66.91</td><td>61.64</td></tr><tr><td>Ours</td><td>99.55</td><td>99.20</td><td>73.85</td><td>74.25</td><td>88.05</td><td>98.25</td><td>89.35</td><td>89.70</td><td>88.60</td><td>91.20</td><td>74.10</td><td>80.70</td><td>87.15</td><td>72.00</td><td>81.50</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.09,
+                0.493,
+                0.907,
+                0.522
+            ],
+            "angle": 0,
+            "content": "Table 2. Similar to Tab. 1, we compare these methods using classification accuracy, which is influenced by manually set thresholds. For a fair comparison, the classification threshold for all models is set to 0.5."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.545,
+                0.483,
+                0.652
+            ],
+            "angle": 0,
+            "content": "cluding CNNDetection (CVPR 2020) [39], FreDect (ICML 2020) [8], Fusing (ICIP 2022) [14], LGrad (CVPR 2023) [36], DIRE (ICCV 2023) [40], Ojha (CVPR 2023) [28], NPR (CVPR 2024) [37], CLIPping(ICMR 2024) [17]. To ensure a fair comparison, we re-train and predict using the official code on ForenSynths and Genimage."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.655,
+                0.483,
+                0.79
+            ],
+            "angle": 0,
+            "content": "Evaluation metrics We follow existing research and consider two metrics: Average Precision (AP) and Accuracy (ACC). Average Precision (AP) is threshold-independent and evaluates the overall performance of the model. Accuracy is defined as the number of correct predictions divided by the number of test images, and this metric is influenced by the decision threshold. Given that there is no prior information in real-world scenarios, we standardize the decision threshold to 0.5."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.795,
+                0.483,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Implementation Details Our method is implemented using the PyTorch framework on the NVIDIA Tesla V100. The model is trained with the Adam optimizer, a learning rate of \\(2 \\times 10^{-4}\\), and a batch size of 256. Following the same data augmentation strategy as in [39][28], which includes Gaussian blur and JPEG compression, experiments have shown that this is critical for the model's generalization ability."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.545,
+                0.907,
+                0.575
+            ],
+            "angle": 0,
+            "content": "5.2. Performance evaluation based on Diffusion sources"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.511,
+                0.583,
+                0.907,
+                0.764
+            ],
+            "angle": 0,
+            "content": "It is well-known that images generated by the current state-of-the-art diffusion models are more realistic than those generated by other types of generative models. To validate the effectiveness of the proposed method, we designed a challenging experimental scenario. Specifically, we trained various detector using Stable Diffusion v1.4 and then tested them on images generated by 15 different models, which include the latest diffusion-based generative models as well as classic GANs. The results are presented in Tabs. 1 and 2. Bold indicates the highest value in the column, and underline indicates the second highest value in the column. This applies to the other tables as well."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.511,
+                0.765,
+                0.909,
+                0.903
+            ],
+            "angle": 0,
+            "content": "We first evaluated the performance of each model on the latest generative models. As shown in Tabs. 1 and 2, advanced diffusion models significantly challenged most detection models, causing a sharp drop in performance. For example, Univfd and Lgrad experienced a \\(40\\%\\) decrease in accuracy and precision when tested on ADM, Midjourney, and VQDM. This decline is likely due to the high-quality images produced by these models, which make it difficult for detection methods to extract distinguishing fea"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "23833"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.095,
+                0.089,
+                0.908,
+                0.253
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Detection method</td><td colspan=\"6\">Generative Adversarial Networks</td><td colspan=\"2\">Perceptual loss</td><td colspan=\"2\">Others</td><td colspan=\"7\">Diffusion Models</td><td>AP</td></tr><tr><td>Pro-GAN</td><td>Cycle-GAN</td><td>Big-GAN</td><td>Style-GAN</td><td>Star-GAN</td><td>Gau-GAN</td><td>CRN</td><td>IMLE</td><td>Deep-fake</td><td>SAN</td><td>SDV1.4</td><td>SDV1.5</td><td>ADM</td><td>GLIDE</td><td>Midjourney</td><td>Wukong</td><td>VQDM</td><td>Average</td></tr><tr><td>CNNSpot[39]</td><td>99.99</td><td>96.40</td><td>87.50</td><td>96.94</td><td>94.24</td><td>98.28</td><td>96.61</td><td>98.76</td><td>64.42</td><td>55.89</td><td>52.86</td><td>53.25</td><td>65.14</td><td>68.10</td><td>56.60</td><td>51.15</td><td>69.49</td><td>76.80</td></tr><tr><td>FreDect[8]</td><td>99.99</td><td>84.77</td><td>93.62</td><td>88.97</td><td>99.48</td><td>82.85</td><td>82.27</td><td>69.77</td><td>70.77</td><td>49.50</td><td>38.50</td><td>38.41</td><td>63.72</td><td>54.73</td><td>47.25</td><td>40.44</td><td>86.01</td><td>70.06</td></tr><tr><td>Fusing[14]</td><td>100.00</td><td>95.50</td><td>90.76</td><td>99.48</td><td>99.82</td><td>88.32</td><td>90.32</td><td>96.02</td><td>71.12</td><td>77.33</td><td>65.30</td><td>65.62</td><td>74.85</td><td>77.44</td><td>69.91</td><td>64.53</td><td>75.42</td><td>82.46</td></tr><tr><td>Lgrad[36]</td><td>99.90</td><td>94.01</td><td>90.75</td><td>99.80</td><td>99.98</td><td>79.29</td><td>61.64</td><td>67.05</td><td>71.71</td><td>45.09</td><td>70.90</td><td>71.72</td><td>71.83</td><td>75.96</td><td>71.42</td><td>66.51</td><td>70.23</td><td>76.93</td></tr><tr><td>Univfd[28]</td><td>100.00</td><td>99.21</td><td>98.31</td><td>97.98</td><td>99.35</td><td>99.80</td><td>96.72</td><td>99.00</td><td>82.04</td><td>82.18</td><td>85.48</td><td>82.30</td><td>84.34</td><td>84.04</td><td>69.10</td><td>90.13</td><td>94.96</td><td>90.88</td></tr><tr><td>NPR[37]</td><td>100.00</td><td>98.50</td><td>87.80</td><td>99.80</td><td>99.90</td><td>85.50</td><td>50.00</td><td>50.00</td><td>82.40</td><td>71.60</td><td>84.00</td><td>84.60</td><td>74.60</td><td>85.70</td><td>85.40</td><td>80.50</td><td>81.20</td><td>82.44</td></tr><tr><td>CLIPping[17]</td><td>99.85</td><td>94.03</td><td>91.27</td><td>95.24</td><td>99.05</td><td>89.52</td><td>79.87</td><td>80.33</td><td>73.89</td><td>56.62</td><td>58.28</td><td>57.85</td><td>77.25</td><td>77.93</td><td>52.43</td><td>60.65</td><td>80.00</td><td>77.89</td></tr><tr><td>Ours</td><td>100.00</td><td>99.80</td><td>99.29</td><td>98.79</td><td>99.72</td><td>99.99</td><td>90.89</td><td>96.83</td><td>92.64</td><td>91.62</td><td>87.24</td><td>86.98</td><td>87.88</td><td>88.53</td><td>75.68</td><td>90.92</td><td>96.51</td><td>93.14</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.09,
+                0.263,
+                0.907,
+                0.293
+            ],
+            "angle": 0,
+            "content": "Table 3. The \\(AP\\) values of different methods trained on GAN source images. Generative models other than those in the GANs column can be considered as the generalization domain."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.095,
+                0.305,
+                0.907,
+                0.481
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Detection method</td><td colspan=\"6\">Generative Adversarial Networks</td><td colspan=\"2\">Perceptual loss</td><td colspan=\"2\">Others</td><td colspan=\"7\">Diffusion Models</td><td>ACC</td></tr><tr><td>Pro-GAN</td><td>Cycle-GAN</td><td>Big-GAN</td><td>Style-GAN</td><td>Star-GAN</td><td>Gau-GAN</td><td>CRN</td><td>IMLE</td><td>Deep-fake</td><td>SAN</td><td>SDV1.4</td><td>SDV1.5</td><td>ADM</td><td>GLIDE</td><td>Midjourney</td><td>Wukong</td><td>VQDM</td><td>Average</td></tr><tr><td>CNNSpot[39]</td><td>99.99</td><td>87.59</td><td>71.18</td><td>89.95</td><td>94.60</td><td>81.44</td><td>85.87</td><td>85.87</td><td>51.69</td><td>50.00</td><td>50.82</td><td>50.88</td><td>60.20</td><td>57.85</td><td>50.77</td><td>51.13</td><td>56.20</td><td>69.18</td></tr><tr><td>FreDect[8]</td><td>99.36</td><td>78.76</td><td>81.97</td><td>78.01</td><td>94.62</td><td>80.57</td><td>60.04</td><td>60.13</td><td>63.29</td><td>50.00</td><td>40.02</td><td>40.38</td><td>64.66</td><td>55.43</td><td>46.89</td><td>41.54</td><td>78.95</td><td>65.57</td></tr><tr><td>Fusing[14]</td><td>99.90</td><td>87.01</td><td>77.32</td><td>85.21</td><td>97.04</td><td>76.95</td><td>85.17</td><td>88.54</td><td>53.76</td><td>54.56</td><td>51.04</td><td>51.34</td><td>56.50</td><td>57.15</td><td>52.12</td><td>51.67</td><td>55.09</td><td>69.43</td></tr><tr><td>DIRE[40]</td><td>100.00</td><td>76.65</td><td>72.12</td><td>85.03</td><td>100.00</td><td>67.79</td><td>-</td><td>-</td><td>-</td><td>-</td><td>49.74</td><td>49.83</td><td>75.78</td><td>71.75</td><td>58.01</td><td>54.46</td><td>53.68</td><td>-</td></tr><tr><td>Lgrad[36]</td><td>99.80</td><td>86.94</td><td>85.63</td><td>91.08</td><td>99.27</td><td>72.49</td><td>50.49</td><td>50.50</td><td>56.42</td><td>44.47</td><td>63.03</td><td>63.67</td><td>67.10</td><td>66.10</td><td>56.20</td><td>63.60</td><td>67.02</td><td>69.64</td></tr><tr><td>Univfd[28]</td><td>99.90</td><td>98.50</td><td>94.50</td><td>84.40</td><td>95.85</td><td>99.50</td><td>59.50</td><td>72.00</td><td>67.40</td><td>56.50</td><td>63.10</td><td>63.57</td><td>66.90</td><td>61.70</td><td>57.85</td><td>71.06</td><td>85.00</td><td>76.31</td></tr><tr><td>NPR[37]</td><td>99.80</td><td>96.10</td><td>84.40</td><td>97.70</td><td>99.30</td><td>82.50</td><td>50.00</td><td>50.00</td><td>80.20</td><td>69.20</td><td>76.60</td><td>77.90</td><td>69.70</td><td>77.30</td><td>77.80</td><td>76.10</td><td>78.10</td><td>78.98</td></tr><tr><td>CLIPping[17]</td><td>99.88</td><td>96.74</td><td>94.77</td><td>94.87</td><td>99.47</td><td>94.15</td><td>85.23</td><td>87.73</td><td>76.48</td><td>60.27</td><td>61.26</td><td>60.78</td><td>80.77</td><td>80.69</td><td>53.95</td><td>63.94</td><td>84.76</td><td>80.93</td></tr><tr><td>Ours</td><td>99.99</td><td>99.00</td><td>95.75</td><td>91.25</td><td>98.95</td><td>99.70</td><td>71.25</td><td>86.75</td><td>83.20</td><td>70.50</td><td>71.55</td><td>70.00</td><td>71.45</td><td>69.40</td><td>61.25</td><td>75.90</td><td>86.65</td><td>82.50</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.09,
+                0.492,
+                0.907,
+                0.521
+            ],
+            "angle": 0,
+            "content": "Table 4. Accuracy of different methods trained on GAN source images. The data from [40] is obtained from [37], and ”-” indicates that the method did not show data for the corresponding model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.54,
+                0.483,
+                0.751
+            ],
+            "angle": 0,
+            "content": "tures. We also observed that these models performed better on generative models with distributions similar to the training data (e.g., SDV1.5, Wukong). This suggests that they may have learned superficial biases from irrelevant features, which are superficially correlated with labels and do not generalize well to images from different distributions. In contrast, NPR, which detects authenticity by analyzing pixel relationships, partially mitigates these biases, resulting in less performance decline. When testing on GANs, we observed significant performance drops, particularly on GauGAN, where the accuracy and precision were around \\(50 - 60\\%\\). This further highlights how even subtle relevant features and surface biases can severely impact the generalization performance of these models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.765,
+                0.483,
+                0.903
+            ],
+            "angle": 0,
+            "content": "In contrast, Our method outperforms the baseline model, Univfd, with improvements of \\(20.22\\%\\) in average precision and \\(20.46\\%\\) in accuracy. This gap is even more pronounced across generative model families (GANs), where our method surpasses Univfd by \\(34.17\\%\\) in precision and \\(26.41\\%\\) in accuracy. Compared to state-of-the-art methods, our approach improves average precision by \\(5.55\\%\\) and accuracy by \\(9.33\\%\\). For other generative models (GANs), VIB-Net achieves gains of \\(12.48\\%\\) in average precision and"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.54,
+                0.907,
+                0.601
+            ],
+            "angle": 0,
+            "content": "23.59% in accuracy. This is due to VIB-Net's ability to retain compact deep features while eliminating irrelevant surface features, enhancing generalization. Overall, these results highlight the advantages of our method."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.607,
+                0.905,
+                0.623
+            ],
+            "angle": 0,
+            "content": "5.3. Performance evaluation based on GAN sources"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.511,
+                0.629,
+                0.907,
+                0.901
+            ],
+            "angle": 0,
+            "content": "So far, we have explored the generalization performance of using images generated by specific diffusion models as training data. Next, we trained using images generated by ProGAN and then tested on 18 different generative models to explore the performance of the detector on unseen generative models. The results are shown in Tabs. 3 and 4. From the tables, we can see that almost all models perform well when detecting images generated by other GAN variants. This is because, compared to diffusion-based generation methods, images generated by these GAN variants are visually more distinguishable, making it easier for the model to learn the distinguishing features of real and fake. However, when detecting images generated by models other than GAN variants, the performance of most models sharply declines. This suggests that even when the training data contains rich features related to real/fake classification, it does not necessarily ensure better generalization performance. This is because the model may overfit the incidental correla"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "23834"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.092,
+                0.089,
+                0.905,
+                0.189
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.092,
+                0.196,
+                0.905,
+                0.297
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.09,
+                0.308,
+                0.908,
+                0.337
+            ],
+            "angle": 0,
+            "content": "Figure 5. Ablation study on the network architecture. The \\(AP\\) and \\(ACC\\) values of different VIB-Net variants on various models. CLIP:ViT-L/14 shows better performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.355,
+                0.483,
+                0.431
+            ],
+            "angle": 0,
+            "content": "tions between irrelevant features and the task, which harms generalization performance. Univfd performs better across model families, but since it extracts general features, many of these features are unrelated to the real/fake classification task, which naturally affects the decision boundary."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.431,
+                0.484,
+                0.567
+            ],
+            "angle": 0,
+            "content": "In contrast, our method shows better performance in detecting real and fake images. Compared to the baseline Univfd, our method achieved improvements of \\(2.26\\%\\) and \\(6.19\\%\\) in average precision and accuracy. In OOD scenarios (excluding GANs), our method outperforms the baseline Univfd by \\(3.22\\%\\) and \\(8.48\\%\\). Compared with advanced methods, our method achieved improvements of \\(10.7\\%\\) to \\(23.08\\%\\) in average precision. These results further validate the effectiveness of our proposed method."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.576,
+                0.465,
+                0.591
+            ],
+            "angle": 0,
+            "content": "5.4. Effect of VIB-Net across different backbone"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.598,
+                0.483,
+                0.825
+            ],
+            "angle": 0,
+            "content": "So far, we have seen the impressive performance of VIB-Net with the CLIP:ViT-L/14 backbone. In this section, we will explore the effect of transferring the VIB component to other multimodal models. We consider the following backbone variants to form VIB-Net: (1) FLAVA-FULL[35] (2) CLIP:ResNet50 (3) CLIP:ViT-B/16 (4) CLIP:ViT-L/14[30]. For each VIB-Net variant, we train using images generated by Stable Diffusion v1.4 and corresponding real images. Fig. 5 show the AP and ACC results of these variants on the same generative model. We observe that among these multimodal models, ViT-L/14 achieves the best performance, which can be attributed to its larger architecture and attention-based global receptive field. More importantly, all these variants exhibit strong performance, which can be attributed to the unique capabilities of our method."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.834,
+                0.327,
+                0.85
+            ],
+            "angle": 0,
+            "content": "5.5. Analysis of latent features"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.856,
+                0.484,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Finally, we apply our method to address the issues analyzed in Sec. 3.2. In Sec. 3.2, we identified that some irrelevant features in CLIP's general features could affect the classi-"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.355,
+                0.907,
+                0.537
+            ],
+            "angle": 0,
+            "content": "fier's decision boundary, causing the model to learn superficial biases, which in turn impacts its generalization performance. As a solution, we proposed VIB-Net. Following the same experimental setup as in Sec. 3.2, we visualized the latent features extracted by VIB-Net using UMAP, and obtained encouraging results, as shown in Fig. 3b. We found that the latent features no longer exhibit obvious irrelevant features related to the real/fake image classification task, as the two separated clusters now belong to the same category (real or generated), unlike in Fig. 3a, where features of the same category were significantly separated. This demonstrates the advantage of our method."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.547,
+                0.634,
+                0.563
+            ],
+            "angle": 0,
+            "content": "6. Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.573,
+                0.909,
+                0.753
+            ],
+            "angle": 0,
+            "content": "We propose a universal real-fake image detection method, VIB-Net. First, we found that the general features extracted by current methods based on large-scale pre-trained models contain irrelevant features that are unrelated to the task of distinguishing real from fake images. Experiments have shown that these features not only affect classification decisions, but may also cause the model to learn superficial biases, thereby damaging its generalization performance. As an alternative, we use VIB to remove task-irrelevant features, aiming to obtain the most concise image representations. These simplified features are maximally correlated with the real-fake classification task."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.765,
+                0.682,
+                0.782
+            ],
+            "angle": 0,
+            "content": "Acknowledgements"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.789,
+                0.909,
+                0.888
+            ],
+            "angle": 0,
+            "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62376046 and Grant 62172067, and in part by the Natural Science Foundation of Chongqing under Grant CSTB2023NSCQ-MSX0341, the Science and Technology Research Program of Chongqing Municipal Education Commission (Grant No. KJQN202200635)."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "23835"
+        }
+    ],
+    [
+        {
+            "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,
+                0.157
+            ],
+            "angle": 0,
+            "content": "[1] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.101,
+                0.159,
+                0.484,
+                0.2
+            ],
+            "angle": 0,
+            "content": "[2] Alexander A Alemi, Ian Fischer, and Joshua V Dillon. Uncertainty in the variational information bottleneck. arXiv preprint arXiv:1807.00906, 2018. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.203,
+                0.483,
+                0.257
+            ],
+            "angle": 0,
+            "content": "[3] Xiuli Bi, Bo Liu, Fan Yang, Bin Xiao, Weisheng Li, Gao Huang, and Pamela C Cosman. Detecting generated images by real images only. arXiv preprint arXiv:2311.00962, 2023. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.26,
+                0.483,
+                0.342
+            ],
+            "angle": 0,
+            "content": "[4] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 103-120. Springer, 2020."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.345,
+                0.483,
+                0.426
+            ],
+            "angle": 0,
+            "content": "[5] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.431,
+                0.484,
+                0.499
+            ],
+            "angle": 0,
+            "content": "[6] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of ai-generated image detection with clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4356-4366, 2024. 1, 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.502,
+                0.483,
+                0.543
+            ],
+            "angle": 0,
+            "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.545,
+                0.483,
+                0.613
+            ],
+            "angle": 0,
+            "content": "[8] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.617,
+                0.483,
+                0.684
+            ],
+            "angle": 0,
+            "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.688,
+                0.483,
+                0.757
+            ],
+            "angle": 0,
+            "content": "[10] 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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.759,
+                0.483,
+                0.814
+            ],
+            "angle": 0,
+            "content": "[11] 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. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.817,
+                0.483,
+                0.857
+            ],
+            "angle": 0,
+            "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.86,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[13] Shizhe Hu, Zhengzheng Lou, Xiaoqiang Yan, and Yangdong Ye. A survey on information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3"
+        },
+        {
+            "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,
+                0.161
+            ],
+            "angle": 0,
+            "content": "[14] Yan Ju, Shan Jia, Lipeng Ke, Hongfei Xue, Koki Nagano, and Siwei Lyu. Fusing global and local features for generalized ai-synthesized image detection. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3465-3469. IEEE, 2022. 2, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.164,
+                0.905,
+                0.204
+            ],
+            "angle": 0,
+            "content": "[15] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 1, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.207,
+                0.905,
+                0.262
+            ],
+            "angle": 0,
+            "content": "[16] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.264,
+                0.905,
+                0.331
+            ],
+            "angle": 0,
+            "content": "[17] Sohail Ahmed Khan and Duc-Tien Dang-Nguyen. Clipping the deception: Adapting vision-language models for universal deepfake detection. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 1006-1015, 2024. 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.334,
+                0.905,
+                0.362
+            ],
+            "angle": 0,
+            "content": "[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.364,
+                0.905,
+                0.404
+            ],
+            "angle": 0,
+            "content": "[19] Li Lin, Irene Amerini, Xin Wang, Shu Hu, et al. Robust clip-based detector for exposing diffusion model-generated images. arXiv preprint arXiv:2404.12908, 2024. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.407,
+                0.905,
+                0.49
+            ],
+            "angle": 0,
+            "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, 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": [
+                0.518,
+                0.492,
+                0.905,
+                0.545
+            ],
+            "angle": 0,
+            "content": "[21] Bo Liu, Fan Yang, Xiuli Bi, Bin Xiao, Weisheng Li, and Xinbo Gao. Detecting generated images by real images. In European Conference on Computer Vision, pages 95-110. Springer, 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.548,
+                0.907,
+                0.602
+            ],
+            "angle": 0,
+            "content": "[22] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.605,
+                0.907,
+                0.673
+            ],
+            "angle": 0,
+            "content": "[23] Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. Lare^ 2: Latent reconstruction error based method for diffusion-generated image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17006-17015, 2024. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.676,
+                0.905,
+                0.729
+            ],
+            "angle": 0,
+            "content": "[24] Rabeeh Karimi Mahabadi, Yonatan Belinkov, and James Henderson. Variational information bottleneck for effective low-resource fine-tuning. arXiv preprint arXiv:2106.05469, 2021.3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.733,
+                0.905,
+                0.788
+            ],
+            "angle": 0,
+            "content": "[25] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.79,
+                0.905,
+                0.83
+            ],
+            "angle": 0,
+            "content": "[26] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.832,
+                0.905,
+                0.9
+            ],
+            "angle": 0,
+            "content": "[27] 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 preprint arXiv:2112.10741, 2021. 5"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.517,
+                0.093,
+                0.907,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.52,
+                0.957
+            ],
+            "angle": 0,
+            "content": "23836"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.092,
+                0.482,
+                0.161
+            ],
+            "angle": 0,
+            "content": "[28] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 2, 3, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.164,
+                0.482,
+                0.233
+            ],
+            "angle": 0,
+            "content": "[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.235,
+                0.482,
+                0.318
+            ],
+            "angle": 0,
+            "content": "[30] 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, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.321,
+                0.482,
+                0.389
+            ],
+            "angle": 0,
+            "content": "[31] 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"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.391,
+                0.482,
+                0.461
+            ],
+            "angle": 0,
+            "content": "[32] 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,
+                0.462,
+                0.482,
+                0.532
+            ],
+            "angle": 0,
+            "content": "[33] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.534,
+                0.482,
+                0.603
+            ],
+            "angle": 0,
+            "content": "[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.604,
+                0.482,
+                0.687
+            ],
+            "angle": 0,
+            "content": "[35] 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, pages 15638-15650, 2022. 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.689,
+                0.482,
+                0.772
+            ],
+            "angle": 0,
+            "content": "[36] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.774,
+                0.482,
+                0.857
+            ],
+            "angle": 0,
+            "content": "[37] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.859,
+                0.482,
+                0.902
+            ],
+            "angle": 0,
+            "content": "[38] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.093,
+                0.092,
+                0.482,
+                0.902
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.092,
+                0.907,
+                0.163
+            ],
+            "angle": 0,
+            "content": "[39] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 1, 2, 5, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.164,
+                0.906,
+                0.233
+            ],
+            "angle": 0,
+            "content": "[40] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 1, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.234,
+                0.906,
+                0.275
+            ],
+            "angle": 0,
+            "content": "[41] Among Wu and Cheng Deng. Tib: Detecting unknown objects via two-stream information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.277,
+                0.906,
+                0.332
+            ],
+            "angle": 0,
+            "content": "[42] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.334,
+                0.906,
+                0.389
+            ],
+            "angle": 0,
+            "content": "[43] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.39,
+                0.906,
+                0.459
+            ],
+            "angle": 0,
+            "content": "[44] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.46,
+                0.906,
+                0.529
+            ],
+            "angle": 0,
+            "content": "[45] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.531,
+                0.906,
+                0.6
+            ],
+            "angle": 0,
+            "content": "[46] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. Advances in Neural Information Processing Systems, 36, 2024. 5"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.517,
+                0.092,
+                0.907,
+                0.6
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "23837"
+        }
+    ]
+]

2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/c23bce64-f04c-4e24-b700-3d6473e34043_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9b4059a94666b77757e3602a8008a712e66b7716ebb6bc8373dac6cc32eb7ab
+size 3264067

2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/full.md

@@ -0,0 +1,262 @@
+# Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network
+
+Haifeng Zhang, Qinghui He, Xiuli Bi, Weisheng Li, Bo Liu*, Bin Xiao Chongqing University of Posts and Telecommunications
+
+s220331143@stu.cqupt.edu.cn, s220232014@stu.cqupt.edu.cn
+
+{bix1, liws, boliu, xiaobin}@cqupt.edu.cn
+
+# Abstract
+
+The rapid advancement of generative models has significantly improved the quality of generated images. Meanwhile, it challenges information authenticity and credibility. Current generated image detection methods based on large-scale pre-trained multimodal models have achieved impressive results. Although these models provide abundant features, the authentication task-related features are often submerged. Consequently, those authentication task-irrelated features cause models to learn superficial biases, thereby harming their generalization performance across different model genera (e.g., GANs and Diffusion Models). To this end, we proposed VIB-Net, which uses Variational Information Bottlenecks to enforce authentication task-related feature learning. We tested and analyzed the proposed method and existing methods on samples generated by 17 different generative models. Compared to SOTA methods, VIB-Net achieved a  $5.55\%$  improvement in mAP and a  $9.33\%$  increase in accuracy. Notably, in generalization tests on unseen generative models from different series, VIB-Net improved mAP by  $12.48\%$  and accuracy by  $23.59\%$  over SOTA methods. The code is available at https://github.com/oceanzhf/VIBAIGCDetect.
+
+# 1. Introduction
+
+The continuous advancements in deep learning have led to the emergence of numerous well-known image generation models, such as GANs[9][15] and Diffusion[12][31]. Images generated by these models have reached a level where observers can hardly distinguish them from real images, and the flexibility of tools based on these models has also greatly improved [6]. With a few simple steps, people can easily obtain high-quality images as needed. While this presents a tremendous opportunity for fields like visual arts, it also serves as a paradise for individuals looking to fab
+
+![](images/432848555de54ec22b2ec2160073a539e39993a3a52b1f214925a0ea15c4b742.jpg)  
+Figure 1. Each method's overall performance('P') on test data and generalization performance ('G') on unseen data are linked by dashed lines. Our detector demonstrates a clear advantage over all SOTA methods in both metrics.
+
+ricate misinformation. Given the ease of creating manipulated images today and the potential risk of using them for malicious purposes, there is an urgent need to develop detectors that help distinguish between real and generated images.
+
+Existing detection methods [4][39][25][43] focus on identifying specific traces left by generators in either the spatial or frequency domain of generated images, thus establishing a decision boundary between real and generated images. These detection methods are highly effective in detecting images generated by GANs and earlier diffusion models. However, their performance drops significantly when facing images generated by recent, popular models that visually resemble real images. Although some error reconstruction-based methods [40][23] have been developed specifically for detecting images generated by diffusion models, they similarly experience performance declines when applied to GAN-generated images. Consequently, current research increasingly focuses on improving the cross-model genera detection capabilities of detectors.
+
+Generalization across different domains is a primary re
+
+search challenge in deep learning. In this context, the emergence of large-scale pre-trained multimodal models has introduced new approaches and promising results. For instance, by using the image encoder from the CLIP [30] model to extract image features and training a classifier with linear probing, there is a notable improvement in generalization performance over models like ResNet50[28][6]. However, compared to traditional classification tasks, real and generated images share subtle, harder-to-capture distinguishing features. This implies that the features extracted by large-scale pre-trained models for the task of distinguishing real from fake images may contain a considerable amount of task-irrelevant features. Task-irrelevant features can either have no impact on the realizness of the image or, though effective in isolation, become redundant when combined with other features. These features not only interfere with the classifier's ability to accurately learn decision boundaries but also hinder the model's ability to capture and distinguish task-relevant characteristics. More critically, task-irrelevant features may lead the model to learn superficial biases. For example, if a fake image contains a tree while the real image does not, the model might erroneously associate the presence of the tree with being a fake image. Consequently, when detecting other images, the model may mistakenly classify a real image as fake if it contains a tree. Therefore, when dealing with high-quality generated images with subtle distinguishing features, it is challenging to fine-tune the model to separate essential from unnecessary information. This can lead to an overfitting of the incidental statistical correlations between unnecessary information and target images, ultimately undermining the model's generalization and discrimination capabilities.
+
+How can we harness the powerful feature extraction capability of large models while minimizing the interference of unnecessary information? By approaching this from a novel perspective, we propose addressing the issue using Information Bottleneck Theory. This theory aims to create latent encodings that not only effectively represent the target but also maximally compress the original input, thereby avoiding the impact of task-irrelevant features on the model. Specifically, we introduce a method called VIB-Net, which couples a large-scale pre-trained multimodal model with Variational Information Bottleneck (VIB) [1]. Considering that feature representations extracted by the model's image encoder are more task-specific, we decouple and filter these features using latent variables. This method maximizes the mutual information between the latent feature and the classification output, while minimizing the mutual information between the latent feature and the input, resulting in a concise feature that should only contain the effective information needed for the task. This concise latent feature helps the model avoid overfitting to irrelevant features and the incidental correlations with the task, preventing the model
+
+from learning superficial biases, which is beneficial to the model's generalization performance.
+
+In summary, our contributions are as follows: (1) To the best of our knowledge, we are the first to apply Information Bottleneck theory and techniques to the detection of generated images, demonstrating the potential of this theory and approach in this field. (2) We analyze the limitations of methods based on large-scale pre-trained models in detecting images generated by unseen generative models, and present our theory of what could be wrong with the existing paradigm. (3) We propose VIB-Net for the detection of generated images, which uses VIB to decouple and filter the general features extracted by large-scale pre-trained multimodal models, removing task-irrelevant information to enhance detection performance. (4) A large number of experiments validate the advantages of our method (see Fig. 1), especially in the improvement of generalization performance  $(+23.59\%$  accuracy).
+
+# 2. Related work
+
+Generalization Performance. The continuous advancement of generated image models presents ongoing challenges for generated image detection, making the generalization performance of detection models on unseen generators an increasing focus for researchers[14][21]. [39] explored the effects of data preprocessing and augmentation methods on the generalization performance of a ResNet50[11]-based detector. [44] argues that the differences between real and fake images are subtle and localized, framing forgery detection as a fine-grained task that could positively impact generalization performance. [3] takes a frequency-domain approach, mapping real images to a dense subspace and synthetic images outside this subspace to enhance generalization. While these methods have achieved certain successes on different models within the same generation series, they show limitations across different series. This is due to the fact that, with the development of generative algorithms, subtle traces in generated images are becoming increasingly difficult to capture. Additionally, artifacts in the frequency domain differ between model series and are gradually decreasing in visibility.
+
+The impressive results of large-scale pre-trained models in terms of transferability and generalization have sparked significant interest among researchers in the field of generated image detection. [28] argues that existing detectors perform asymmetric learning, where they use CLIP's image encoder to extract image features, exploring the potential of large models in detecting different model families. [6] suggests that detectors based on large models rely on higher-level semantic features, which enhances the generalization ability of the detectors. [19] uses both image and text features extracted by the CLIP model, and then employs a multilayer perceptron to detect images generated by
+
+diffusion models, improving generalization performance. Information Bottleneck. The Information Bottleneck (IB) [38] is an innovative information-theoretic method used for pattern analysis and representation learning. It provides an elegant balance between data compression and information retention[13]. VIB [1] is a variational approximation of IB, aimed at parameterizing IB using neural networks. [2] demonstrated VIB's capability in improving classification calibration and out-of-distribution data. [24] explored VIB as a regularization technique to address overfitting issues in NLP models. [41] proposed a dual-stream information bottleneck method to utilize in-distribution data to enhance out-of-distribution target detection tasks.
+
+Based on the above analysis, current methods for detecting generated images using large-scale pre-trained models simply utilize the general representations extracted from images, without thoroughly exploring the relationship between these general features and downstream tasks. Additionally, Information Bottleneck is an effective technique for improving performance in cross-domain and cross-distribution tasks. Therefore, unlike existing methods, our approach leverages Information Bottleneck techniques to explore the potential of large-scale pre-trained models in generated image detection from a completely new perspective, achieving outstanding performance.
+
+# 3. Preliminaries
+
+# 3.1. Problem background
+
+A common paradigm for using large-scale pretrained models is to leverage a large-scale pretrained multimodal model (e.g., CLIP) to learn the general representation of real and generated images, and then apply this feature representation to downstream tasks[28][6]. Building on their findings, we further consider the following points: (1) It is well-known that the features extracted by large-scale pre-trained visual encoders are general-purpose, meaning they contain many general features, such as semantic information. (2) Another widely accepted notion is that some critical features for distinguishing real from generated images are subtle and local.
+
+Further combining these two points, we propose the following hypotheses: (i) The general features extracted by large-scale pre-trained multimodal models are more beneficial for real-fake image classification tasks than those extracted by pre-trained visual models. (ii) These general features contain task-irrelevant features for real-fake classification. (iii) These task-irrelevant features affect the classification of real and fake images. Next, we will delve into these three hypotheses.
+
+# 3.2. Analysis of general feature representation
+
+For the first hypothesis, We use image pairs of real and generated images to aid our analysis[6], generating textual de
+
+![](images/e1410341985ff647e745d20a946c655965b19f0cbbf303dfb911f10c27cde35e.jpg)  
+Figure 2. UMAP visualization of features extracted from Dalle2 and Dalle3 generated images using large-scale pretrained multimodal(CLIP) and unimodal models(ResNet).
+
+scriptions for 1,000 real images and using these descriptions to produce 1,000 corresponding images with DALL-E 2 and DALL-E 3. We then compare features extracted by large-scale pre-trained multimodal models(CLIP[30]) and pre-trained visual models(ResNet[11]) through UMAP [26] visualization of 3,000 images. The results show that features extracted by multimodal models (e.g., CLIP) exhibit greater distribution differences for real images and DALL-E 2-generated images, suggesting multimodal models may be better at distinguishing real from generated images. However, for DALL-E 3-generated images, both model types show similar feature distributions, possibly due to advancements in image generation tools, which complicate the classification task and require further fine-tuning.
+
+To further explore hypotheses (ii) and (iii), we visualized the extracted features using both category and authenticity as labels to observe their distribution differences. Specifically, we prepared images of two categories (car and cat), each containing 500 real images and 500 generated images. We then used CLIP to extract their features and performed UMAP dimensionality reduction for visualization, with the results shown in Fig. 3a. We can see that the features extracted by CLIP clearly form two clusters: the left cluster mainly consists of light blue and light red points, while the right cluster is mainly composed of light pink and light purple points. These two clusters correspond to category 1 and category 2, respectively. This indicates that CLIP can significantly distinguish category information. However, this clear separation of category information may affect the decision-making process for classifying real vs. generated images. When examining the distinction between real and generated images, we notice that the right cluster shows a noticeable separation of real and generated features, with
+
+![](images/fe8f885df46bbd6b886ae2693b1dc8ce8b2ffa83bcc61143cf31da8d0c17917d.jpg)  
+(a) Visualization of features extracted by CLIP.  
+Figure 3. UMAP visualization of features extracted from real and generated images of two category using CLIP and VIB-Net.
+
+![](images/1ab822a33b6db285e7ab7f352b8aedae2588e0ed8278793c8c97bcbb3a4451d0.jpg)  
+(b) Visualization of features extracted by VIB-Net.
+
+two different colors observed within the cluster. This suggests that CLIP is capable of distinguishing real and generated features within category 2. However, in the left cluster, CLIP does not show significant ability to differentiate between real and generated features. We observe that in the left cluster (category 1), the two colors are mixed together, showing poor distinguishability. From these observations, it is easy to infer that the internal separation of real or generated images caused by semantic category information makes it more difficult for the classifier to learn the decision boundary. More importantly, the strong sensitivity of large-scale pre-trained multimodal models like CLIP to semantic information leads the model to tend to learn superficial biases.
+
+Based on this viewpoint, we consider whether it is possible to minimize the irrelevant features within these multimodal features while retaining the ones that are relevant to the classification task. This would help the model avoid learning superficial biases, allow the classifier to more easily learn decision boundaries, and thereby enhance the model's discrimination ability and generalization capability. This is the motivation of this paper. As shown in Fig. 3b, the method(VIB-Net) proposed in this paper achieves significant results.
+
+# 4. Approach
+
+Based on the issues analyzed in Sec. 3, an effective approach is to filter the multimodal features in order to obtain the representations most beneficial for the real-fake classification task, which serves as the starting point of this paper. Overall, our method is illustrated in Fig. 4. VIB attempts to filter out irrelevant features and to find the most concise representation, making it easier for the classifier to learn the features needed for distinguishing between real and fake tasks, thereby significantly improving the accuracy of the classification task. Below, we will provide a detailed description of our method.
+
+# 4.1. Information bottleneck constraint
+
+To specifically optimize the filtering of irrelevant information in the input features while retaining task-relevant information, we apply the information bottleneck (IB) principle as a constraint on the features to be optimized. Let  $X$ ,  $Z$ , and  $Y$  represent the input, the latent compressed representation, and the output target, respectively. The goal of IB is to find the maximum compressed representation  $Z$  of the input  $X$ , while preserving as much information about the output  $Y$  as possible. The maximization objective function is defined as follows:
+
+$$
+L _ {I B} = I (Z, Y) - \beta I (Z, X) \tag {1}
+$$
+
+Where  $I(.,.)$  denotes mutual information and  $\beta$  is a trade-off parameter. The first term constrains the latent compressed representation to retain sufficient information for downstream tasks, while the second term ensures that irrelevant and redundant information is filtered out.
+
+# 4.2. Variational information bottleneck
+
+To parameterize the information bottleneck (IB) using deep neural networks, we perform variational inference on Equation (1). First, based on the definition of mutual information, Equation (1) can be rewritten as:
+
+$$
+\begin{array}{l} L _ {I B} = \int d y d z p (y, z) \log {\frac {p (y | z)}{p (y)}} \\ - \beta \int d x d z p (x, z) \log \frac {p (z \mid x)}{p (z)} \tag {2} \\ \end{array}
+$$
+
+Where  $p(y|z)$  is defined by the encoder and the Markov chain as follows:
+
+$$
+p (y \mid z) = \int d x \frac {p (y \mid x) p (z \mid x) p (x)}{p (z)} \tag {3}
+$$
+
+Since it is not computable here, we consider setting  $q(y|z)$  as a parameterized approximation of  $p(y|z)$ , serving as a "decoder," which can be approximated by a linear layer. We introduce the Kullback-Leibler (KL) divergence to obtain an estimizable representation. As the KL divergence satisfies  $\mathrm{KL}(p(y|z),q(y|z))\geq 0$ , and by applying the Markov assumption,  $p(y,z)$  can be rewritten as  $p(y,z) = \int dx p(x)p(y|x)p(z|x)$ , thereby providing a lower bound for the first term of the objective function:
+
+$$
+I (Z, Y) \geq \int d x d y d z p (x) p (y | x) p (z | x) \log q (y | z) \tag {4}
+$$
+
+At the same time, let  $r(z)$  be an estimate of the prior probability  $p(z)$ , used to approximate the marginal distribution, thereby obtaining an upper bound for the second term:
+
+$$
+I (Z, X) \leq \int d x d z p (x) p (z | x) \log \frac {p (z | x)}{r (z)} \tag {5}
+$$
+
+![](images/f2654d25fbfe695a4e4ecf0ae41380f6fa32b64f9234168d03dc86580330d688.jpg)  
+Figure 4. Overview of VIB-Net. We utilize Variational Information Bottleneck (VIB) techniques to map the image features extracted by CLIP's visual encoder into latent features  $z$ , where the information in  $z$  is filtered based on the Information Bottleneck (IB). Subsequently, the latent features  $z$  are used as input to an MLP to learn the decision boundary.
+
+The encoder we use is defined as  $p(z|x) = p_{\theta}(z|x) = \mathcal{N}\left(z|f^{\mu}(x), f^{\Sigma}(x)\right)$ , where  $f(.)$  represents a shallow MLP,  $\mu$  denotes the K-dimensional mean vector of the output  $z$ , and  $\Sigma$  represents the  $K \times K$  covariance matrix. We select a parameterized Gaussian distribution for both  $p_{\theta}(z|x)$  and  $r(z)$  to allow for the analytical computation of KL divergence. Additionally, we employ the reparameterization trick to estimate gradients [18]. Combining all the components above yields the following objective function:
+
+$$
+\begin{array}{l} L _ {V I B} = \frac {1}{N} \sum_ {n = 1} ^ {N} z \sim p _ {\theta} (z | x) [ - \log q (y _ {\mathrm {n}} | z _ {n}) ] \\ + \beta \mathrm {K L} [ p _ {\theta} (z | x _ {n}), r (z) ] \tag {6} \\ \end{array}
+$$
+
+# 4.3. Workflow of the overall method
+
+As shown in Fig. 4, the parts in VIB-Net work in the following way: First, the real and generated images  $\{x_{1}, x_{2}, x_{3}, \dots, x_{n}\}$  are input into the CLIP image encoder, yielding the corresponding image features  $\{t_{1}, t_{2}, t_{3}, \dots, t_{n}\}$ . Next, these feature vectors are passed into the Variational Information Bottleneck (VIB). Specifically, the image features are processed through a shallow MLP, then a linear layer  $f(.)$  calculates  $\mu$  and  $\Sigma$ , which are transformed using a softmax activation. This layer has  $K$  dimensions, representing the bottleneck size. At this point, we obtain the latent compressed feature  $z$ , which is then fed into a linear layer  $q(y|z)$  to output the target task label  $y$ .
+
+# 5. Experiments
+
+# 5.1. Settings
+
+To better evaluate the model's performance, we utilize two mainstream public datasets: ForenSynths[39] and
+
+Genimage[46]. The ForenSynths dataset primarily consists of images generated by GANs, while the Genimage dataset mainly includes images generated by diffusion models. Consequently, we set up two training and testing schemes: (1) Training on ForenSynths and then testing on both ForenSynths and Genimage, and (2) similarly, training on Genimage and testing on both Genimage and ForenSynths. This setup allows us to more effectively validate the generalization performance of the detector on cross model genera scenarios. The details are as follows:
+
+# Training set:
+
+The ForenSynths training set consists of 20 different categories, each containing 18,000 real images from the LSUN dataset [42] and an equal number of images generated using ProGAN[15]. Following the setup in Genimage, we use images generated by Stable Diffusion v1.4 [32] in the Genimage dataset, along with their corresponding real images, as the training set. In total, these include 32,000 real and generated images.
+
+# Testing set:
+
+11 models from ForenSynths The test set includes images generated by 11 different models, primarily from the GAN family[15][45][29][5][16]. The real images are sourced from six datasets, including ImageNet[34], LSUN[42], FaceForensics++[33], CelebA-HQ[15], CelebA[22], and COCO[20].
+
+7 models from Genimage To validate the method across different generation models, we use images generated by seven advanced generators along with their corresponding real images as part of the test set. These seven generation models include Stable Diffusion v1.4/1.5[32], VQDM[10], ADM[7], GLIDE[27], Wukong, and Midjourney.
+
+Real-vs-Fake classification baselines: We compare VIB-Net with existing state-of-the-art detection methods, in
+
+<table><tr><td rowspan="2">Detection method</td><td colspan="7">Diffusion Models</td><td colspan="5">Generative Adversarial Networks</td><td colspan="2">Others</td><td>AP</td><td></td></tr><tr><td>SDV1.4</td><td>SDV1.5</td><td>ADM</td><td>GLIDE</td><td>Midjourney</td><td>Wukong</td><td>VQDM</td><td>Pro-GAN</td><td>Cycle-Big-GAN</td><td>Style-GAN</td><td>Star-GAN</td><td>Gau-GAN</td><td colspan="2">Deep-fake</td><td>Average</td><td></td></tr><tr><td>CNNSpot[39]</td><td>99.98</td><td>99.83</td><td>51.10</td><td>58.80</td><td>67.93</td><td>99.80</td><td>49.92</td><td>53.15</td><td>50.23</td><td>49.79</td><td>55.98</td><td>47.07</td><td>56.08</td><td>54.86</td><td>54.03</td><td>63.24</td></tr><tr><td>Fusing[14]</td><td>99.90</td><td>97.98</td><td>69.30</td><td>94.20</td><td>81.20</td><td>99.90</td><td>84.60</td><td>67.63</td><td>87.79</td><td>69.37</td><td>67.90</td><td>91.20</td><td>43.08</td><td>72.56</td><td>89.42</td><td>81.07</td></tr><tr><td>Lgrad[36]</td><td>99.94</td><td>99.92</td><td>58.52</td><td>84.00</td><td>91.06</td><td>99.72</td><td>56.34</td><td>83.59</td><td>90.24</td><td>47.51</td><td>82.74</td><td>99.19</td><td>49.25</td><td>66.49</td><td>65.09</td><td>78.24</td></tr><tr><td>Univfd[28]</td><td>96.04</td><td>96.26</td><td>66.34</td><td>93.73</td><td>92.08</td><td>90.98</td><td>74.53</td><td>51.77</td><td>63.42</td><td>75.81</td><td>54.12</td><td>54.93</td><td>65.99</td><td>70.24</td><td>83.34</td><td>75.31</td></tr><tr><td>NPR[37]</td><td>100.00</td><td>99.97</td><td>94.70</td><td>95.80</td><td>95.50</td><td>100.00</td><td>86.30</td><td>83.30</td><td>94.90</td><td>72.00</td><td>82.70</td><td>97.30</td><td>66.00</td><td>85.30</td><td>95.90</td><td>89.98</td></tr><tr><td>CLIPping[17]</td><td>93.97</td><td>93.10</td><td>68.00</td><td>87.44</td><td>77.34</td><td>86.52</td><td>77.17</td><td>88.54</td><td>88.44</td><td>85.33</td><td>77.23</td><td>89.82</td><td>81.56</td><td>61.19</td><td>57.37</td><td>80.87</td></tr><tr><td>Ours</td><td>100.00</td><td>99.97</td><td>95.49</td><td>97.13</td><td>97.81</td><td>99.93</td><td>97.00</td><td>96.59</td><td>98.44</td><td>97.17</td><td>84.31</td><td>97.60</td><td>96.94</td><td>81.32</td><td>93.27</td><td>95.53</td></tr></table>
+
+Table 1. The  $AP$  values of different methods based on diffusion sources. Models outside the Diffusion Models (DMs) column are considered as the generalization domain across generative model series.  
+
+<table><tr><td rowspan="2">Detection method</td><td colspan="7">Diffusion Models</td><td colspan="5">Generative Adversarial Networks</td><td>Others</td><td>ACC</td><td></td></tr><tr><td>SDV1.4</td><td>SDV1.5</td><td>ADM</td><td>GLIDE</td><td>Midjourney</td><td>Wukong</td><td>VQDM</td><td>Pro-GAN</td><td>Cycle-Big-GAN</td><td>Style-GAN</td><td>Star-GAN</td><td>Gau-GAN</td><td>Deep-SAN</td><td>Average</td><td></td></tr><tr><td>CNNSpot[39]</td><td>99.48</td><td>99.35</td><td>50.10</td><td>50.90</td><td>56.42</td><td>97.90</td><td>50.04</td><td>50.27</td><td>49.81</td><td>50.10</td><td>50.98</td><td>49.77</td><td>50.38</td><td>51.98</td><td>50.22</td></tr><tr><td>Fusing[14]</td><td>99.90</td><td>99.91</td><td>51.30</td><td>57.50</td><td>52.30</td><td>99.90</td><td>64.20</td><td>51.20</td><td>52.40</td><td>53.50</td><td>50.20</td><td>58.20</td><td>49.32</td><td>51.02</td><td>64.84</td></tr><tr><td>Lgrad[36]</td><td>99.12</td><td>99.05</td><td>53.00</td><td>64.24</td><td>76.34</td><td>97.53</td><td>50.93</td><td>61.61</td><td>60.74</td><td>48.82</td><td>61.43</td><td>50.17</td><td>49.70</td><td>50.17</td><td>56.49</td></tr><tr><td>Univfd[28]</td><td>83.55</td><td>84.80</td><td>53.35</td><td>75.30</td><td>71.60</td><td>73.55</td><td>55.10</td><td>58.65</td><td>59.30</td><td>61.45</td><td>56.80</td><td>61.45</td><td>55.30</td><td>58.40</td><td>72.00</td></tr><tr><td>NPR[37]</td><td>100.00</td><td>99.90</td><td>73.00</td><td>89.70</td><td>82.30</td><td>100.00</td><td>68.30</td><td>60.30</td><td>67.20</td><td>59.20</td><td>58.00</td><td>73.20</td><td>52.00</td><td>74.80</td><td>89.60</td></tr><tr><td>CLIPping[17]</td><td>96.07</td><td>95.48</td><td>70.14</td><td>85.00</td><td>77.66</td><td>88.86</td><td>79.35</td><td>88.78</td><td>88.48</td><td>89.57</td><td>80.69</td><td>90.82</td><td>85.94</td><td>66.91</td><td>61.64</td></tr><tr><td>Ours</td><td>99.55</td><td>99.20</td><td>73.85</td><td>74.25</td><td>88.05</td><td>98.25</td><td>89.35</td><td>89.70</td><td>88.60</td><td>91.20</td><td>74.10</td><td>80.70</td><td>87.15</td><td>72.00</td><td>81.50</td></tr></table>
+
+Table 2. Similar to Tab. 1, we compare these methods using classification accuracy, which is influenced by manually set thresholds. For a fair comparison, the classification threshold for all models is set to 0.5.
+
+cluding CNNDetection (CVPR 2020) [39], FreDect (ICML 2020) [8], Fusing (ICIP 2022) [14], LGrad (CVPR 2023) [36], DIRE (ICCV 2023) [40], Ojha (CVPR 2023) [28], NPR (CVPR 2024) [37], CLIPping(ICMR 2024) [17]. To ensure a fair comparison, we re-train and predict using the official code on ForenSynths and Genimage.
+
+Evaluation metrics We follow existing research and consider two metrics: Average Precision (AP) and Accuracy (ACC). Average Precision (AP) is threshold-independent and evaluates the overall performance of the model. Accuracy is defined as the number of correct predictions divided by the number of test images, and this metric is influenced by the decision threshold. Given that there is no prior information in real-world scenarios, we standardize the decision threshold to 0.5.
+
+Implementation Details Our method is implemented using the PyTorch framework on the NVIDIA Tesla V100. The model is trained with the Adam optimizer, a learning rate of  $2 \times 10^{-4}$ , and a batch size of 256. Following the same data augmentation strategy as in [39][28], which includes Gaussian blur and JPEG compression, experiments have shown that this is critical for the model's generalization ability.
+
+# 5.2. Performance evaluation based on Diffusion sources
+
+It is well-known that images generated by the current state-of-the-art diffusion models are more realistic than those generated by other types of generative models. To validate the effectiveness of the proposed method, we designed a challenging experimental scenario. Specifically, we trained various detector using Stable Diffusion v1.4 and then tested them on images generated by 15 different models, which include the latest diffusion-based generative models as well as classic GANs. The results are presented in Tabs. 1 and 2. Bold indicates the highest value in the column, and underline indicates the second highest value in the column. This applies to the other tables as well.
+
+We first evaluated the performance of each model on the latest generative models. As shown in Tabs. 1 and 2, advanced diffusion models significantly challenged most detection models, causing a sharp drop in performance. For example, Univfd and Lgrad experienced a  $40\%$  decrease in accuracy and precision when tested on ADM, Midjourney, and VQDM. This decline is likely due to the high-quality images produced by these models, which make it difficult for detection methods to extract distinguishing fea
+
+<table><tr><td rowspan="2">Detection method</td><td colspan="6">Generative Adversarial Networks</td><td colspan="2">Perceptual loss</td><td colspan="2">Others</td><td colspan="7">Diffusion Models</td><td>AP</td></tr><tr><td>Pro-GAN</td><td>Cycle-GAN</td><td>Big-GAN</td><td>Style-GAN</td><td>Star-GAN</td><td>Gau-GAN</td><td>CRN</td><td>IMLE</td><td>Deep-fake</td><td>SAN</td><td>SDV1.4</td><td>SDV1.5</td><td>ADM</td><td>GLIDE</td><td>Midjourney</td><td>Wukong</td><td>VQDM</td><td>Average</td></tr><tr><td>CNNSpot[39]</td><td>99.99</td><td>96.40</td><td>87.50</td><td>96.94</td><td>94.24</td><td>98.28</td><td>96.61</td><td>98.76</td><td>64.42</td><td>55.89</td><td>52.86</td><td>53.25</td><td>65.14</td><td>68.10</td><td>56.60</td><td>51.15</td><td>69.49</td><td>76.80</td></tr><tr><td>FreDect[8]</td><td>99.99</td><td>84.77</td><td>93.62</td><td>88.97</td><td>99.48</td><td>82.85</td><td>82.27</td><td>69.77</td><td>70.77</td><td>49.50</td><td>38.50</td><td>38.41</td><td>63.72</td><td>54.73</td><td>47.25</td><td>40.44</td><td>86.01</td><td>70.06</td></tr><tr><td>Fusing[14]</td><td>100.00</td><td>95.50</td><td>90.76</td><td>99.48</td><td>99.82</td><td>88.32</td><td>90.32</td><td>96.02</td><td>71.12</td><td>77.33</td><td>65.30</td><td>65.62</td><td>74.85</td><td>77.44</td><td>69.91</td><td>64.53</td><td>75.42</td><td>82.46</td></tr><tr><td>Lgrad[36]</td><td>99.90</td><td>94.01</td><td>90.75</td><td>99.80</td><td>99.98</td><td>79.29</td><td>61.64</td><td>67.05</td><td>71.71</td><td>45.09</td><td>70.90</td><td>71.72</td><td>71.83</td><td>75.96</td><td>71.42</td><td>66.51</td><td>70.23</td><td>76.93</td></tr><tr><td>Univfd[28]</td><td>100.00</td><td>99.21</td><td>98.31</td><td>97.98</td><td>99.35</td><td>99.80</td><td>96.72</td><td>99.00</td><td>82.04</td><td>82.18</td><td>85.48</td><td>82.30</td><td>84.34</td><td>84.04</td><td>69.10</td><td>90.13</td><td>94.96</td><td>90.88</td></tr><tr><td>NPR[37]</td><td>100.00</td><td>98.50</td><td>87.80</td><td>99.80</td><td>99.90</td><td>85.50</td><td>50.00</td><td>50.00</td><td>82.40</td><td>71.60</td><td>84.00</td><td>84.60</td><td>74.60</td><td>85.70</td><td>85.40</td><td>80.50</td><td>81.20</td><td>82.44</td></tr><tr><td>CLIPping[17]</td><td>99.85</td><td>94.03</td><td>91.27</td><td>95.24</td><td>99.05</td><td>89.52</td><td>79.87</td><td>80.33</td><td>73.89</td><td>56.62</td><td>58.28</td><td>57.85</td><td>77.25</td><td>77.93</td><td>52.43</td><td>60.65</td><td>80.00</td><td>77.89</td></tr><tr><td>Ours</td><td>100.00</td><td>99.80</td><td>99.29</td><td>98.79</td><td>99.72</td><td>99.99</td><td>90.89</td><td>96.83</td><td>92.64</td><td>91.62</td><td>87.24</td><td>86.98</td><td>87.88</td><td>88.53</td><td>75.68</td><td>90.92</td><td>96.51</td><td>93.14</td></tr></table>
+
+Table 3. The  $AP$  values of different methods trained on GAN source images. Generative models other than those in the GANs column can be considered as the generalization domain.  
+
+<table><tr><td rowspan="2">Detection method</td><td colspan="6">Generative Adversarial Networks</td><td colspan="2">Perceptual loss</td><td colspan="2">Others</td><td colspan="7">Diffusion Models</td><td>ACC</td></tr><tr><td>Pro-GAN</td><td>Cycle-GAN</td><td>Big-GAN</td><td>Style-GAN</td><td>Star-GAN</td><td>Gau-GAN</td><td>CRN</td><td>IMLE</td><td>Deep-fake</td><td>SAN</td><td>SDV1.4</td><td>SDV1.5</td><td>ADM</td><td>GLIDE</td><td>Midjourney</td><td>Wukong</td><td>VQDM</td><td>Average</td></tr><tr><td>CNNSpot[39]</td><td>99.99</td><td>87.59</td><td>71.18</td><td>89.95</td><td>94.60</td><td>81.44</td><td>85.87</td><td>85.87</td><td>51.69</td><td>50.00</td><td>50.82</td><td>50.88</td><td>60.20</td><td>57.85</td><td>50.77</td><td>51.13</td><td>56.20</td><td>69.18</td></tr><tr><td>FreDect[8]</td><td>99.36</td><td>78.76</td><td>81.97</td><td>78.01</td><td>94.62</td><td>80.57</td><td>60.04</td><td>60.13</td><td>63.29</td><td>50.00</td><td>40.02</td><td>40.38</td><td>64.66</td><td>55.43</td><td>46.89</td><td>41.54</td><td>78.95</td><td>65.57</td></tr><tr><td>Fusing[14]</td><td>99.90</td><td>87.01</td><td>77.32</td><td>85.21</td><td>97.04</td><td>76.95</td><td>85.17</td><td>88.54</td><td>53.76</td><td>54.56</td><td>51.04</td><td>51.34</td><td>56.50</td><td>57.15</td><td>52.12</td><td>51.67</td><td>55.09</td><td>69.43</td></tr><tr><td>DIRE[40]</td><td>100.00</td><td>76.65</td><td>72.12</td><td>85.03</td><td>100.00</td><td>67.79</td><td>-</td><td>-</td><td>-</td><td>-</td><td>49.74</td><td>49.83</td><td>75.78</td><td>71.75</td><td>58.01</td><td>54.46</td><td>53.68</td><td>-</td></tr><tr><td>Lgrad[36]</td><td>99.80</td><td>86.94</td><td>85.63</td><td>91.08</td><td>99.27</td><td>72.49</td><td>50.49</td><td>50.50</td><td>56.42</td><td>44.47</td><td>63.03</td><td>63.67</td><td>67.10</td><td>66.10</td><td>56.20</td><td>63.60</td><td>67.02</td><td>69.64</td></tr><tr><td>Univfd[28]</td><td>99.90</td><td>98.50</td><td>94.50</td><td>84.40</td><td>95.85</td><td>99.50</td><td>59.50</td><td>72.00</td><td>67.40</td><td>56.50</td><td>63.10</td><td>63.57</td><td>66.90</td><td>61.70</td><td>57.85</td><td>71.06</td><td>85.00</td><td>76.31</td></tr><tr><td>NPR[37]</td><td>99.80</td><td>96.10</td><td>84.40</td><td>97.70</td><td>99.30</td><td>82.50</td><td>50.00</td><td>50.00</td><td>80.20</td><td>69.20</td><td>76.60</td><td>77.90</td><td>69.70</td><td>77.30</td><td>77.80</td><td>76.10</td><td>78.10</td><td>78.98</td></tr><tr><td>CLIPping[17]</td><td>99.88</td><td>96.74</td><td>94.77</td><td>94.87</td><td>99.47</td><td>94.15</td><td>85.23</td><td>87.73</td><td>76.48</td><td>60.27</td><td>61.26</td><td>60.78</td><td>80.77</td><td>80.69</td><td>53.95</td><td>63.94</td><td>84.76</td><td>80.93</td></tr><tr><td>Ours</td><td>99.99</td><td>99.00</td><td>95.75</td><td>91.25</td><td>98.95</td><td>99.70</td><td>71.25</td><td>86.75</td><td>83.20</td><td>70.50</td><td>71.55</td><td>70.00</td><td>71.45</td><td>69.40</td><td>61.25</td><td>75.90</td><td>86.65</td><td>82.50</td></tr></table>
+
+Table 4. Accuracy of different methods trained on GAN source images. The data from [40] is obtained from [37], and ”-” indicates that the method did not show data for the corresponding model.
+
+tures. We also observed that these models performed better on generative models with distributions similar to the training data (e.g., SDV1.5, Wukong). This suggests that they may have learned superficial biases from irrelevant features, which are superficially correlated with labels and do not generalize well to images from different distributions. In contrast, NPR, which detects authenticity by analyzing pixel relationships, partially mitigates these biases, resulting in less performance decline. When testing on GANs, we observed significant performance drops, particularly on GauGAN, where the accuracy and precision were around  $50 - 60\%$ . This further highlights how even subtle relevant features and surface biases can severely impact the generalization performance of these models.
+
+In contrast, Our method outperforms the baseline model, Univfd, with improvements of  $20.22\%$  in average precision and  $20.46\%$  in accuracy. This gap is even more pronounced across generative model families (GANs), where our method surpasses Univfd by  $34.17\%$  in precision and  $26.41\%$  in accuracy. Compared to state-of-the-art methods, our approach improves average precision by  $5.55\%$  and accuracy by  $9.33\%$ . For other generative models (GANs), VIB-Net achieves gains of  $12.48\%$  in average precision and
+
+23.59% in accuracy. This is due to VIB-Net's ability to retain compact deep features while eliminating irrelevant surface features, enhancing generalization. Overall, these results highlight the advantages of our method.
+
+# 5.3. Performance evaluation based on GAN sources
+
+So far, we have explored the generalization performance of using images generated by specific diffusion models as training data. Next, we trained using images generated by ProGAN and then tested on 18 different generative models to explore the performance of the detector on unseen generative models. The results are shown in Tabs. 3 and 4. From the tables, we can see that almost all models perform well when detecting images generated by other GAN variants. This is because, compared to diffusion-based generation methods, images generated by these GAN variants are visually more distinguishable, making it easier for the model to learn the distinguishing features of real and fake. However, when detecting images generated by models other than GAN variants, the performance of most models sharply declines. This suggests that even when the training data contains rich features related to real/fake classification, it does not necessarily ensure better generalization performance. This is because the model may overfit the incidental correla
+
+![](images/03c6eff0437e82330a33e46b3b5553ef28669794ee5a81cf45ccb08c1d454389.jpg)
+
+![](images/446f4f542fe26f5297875262a6a6efa70423d50c08a5ff47202b7b01c2126714.jpg)  
+Figure 5. Ablation study on the network architecture. The  $AP$  and  $ACC$  values of different VIB-Net variants on various models. CLIP:ViT-L/14 shows better performance.
+
+tions between irrelevant features and the task, which harms generalization performance. Univfd performs better across model families, but since it extracts general features, many of these features are unrelated to the real/fake classification task, which naturally affects the decision boundary.
+
+In contrast, our method shows better performance in detecting real and fake images. Compared to the baseline Univfd, our method achieved improvements of  $2.26\%$  and  $6.19\%$  in average precision and accuracy. In OOD scenarios (excluding GANs), our method outperforms the baseline Univfd by  $3.22\%$  and  $8.48\%$ . Compared with advanced methods, our method achieved improvements of  $10.7\%$  to  $23.08\%$  in average precision. These results further validate the effectiveness of our proposed method.
+
+# 5.4. Effect of VIB-Net across different backbone
+
+So far, we have seen the impressive performance of VIB-Net with the CLIP:ViT-L/14 backbone. In this section, we will explore the effect of transferring the VIB component to other multimodal models. We consider the following backbone variants to form VIB-Net: (1) FLAVA-FULL[35] (2) CLIP:ResNet50 (3) CLIP:ViT-B/16 (4) CLIP:ViT-L/14[30]. For each VIB-Net variant, we train using images generated by Stable Diffusion v1.4 and corresponding real images. Fig. 5 show the AP and ACC results of these variants on the same generative model. We observe that among these multimodal models, ViT-L/14 achieves the best performance, which can be attributed to its larger architecture and attention-based global receptive field. More importantly, all these variants exhibit strong performance, which can be attributed to the unique capabilities of our method.
+
+# 5.5. Analysis of latent features
+
+Finally, we apply our method to address the issues analyzed in Sec. 3.2. In Sec. 3.2, we identified that some irrelevant features in CLIP's general features could affect the classi-
+
+fier's decision boundary, causing the model to learn superficial biases, which in turn impacts its generalization performance. As a solution, we proposed VIB-Net. Following the same experimental setup as in Sec. 3.2, we visualized the latent features extracted by VIB-Net using UMAP, and obtained encouraging results, as shown in Fig. 3b. We found that the latent features no longer exhibit obvious irrelevant features related to the real/fake image classification task, as the two separated clusters now belong to the same category (real or generated), unlike in Fig. 3a, where features of the same category were significantly separated. This demonstrates the advantage of our method.
+
+# 6. Conclusion
+
+We propose a universal real-fake image detection method, VIB-Net. First, we found that the general features extracted by current methods based on large-scale pre-trained models contain irrelevant features that are unrelated to the task of distinguishing real from fake images. Experiments have shown that these features not only affect classification decisions, but may also cause the model to learn superficial biases, thereby damaging its generalization performance. As an alternative, we use VIB to remove task-irrelevant features, aiming to obtain the most concise image representations. These simplified features are maximally correlated with the real-fake classification task.
+
+# Acknowledgements
+
+This work was supported in part by the National Natural Science Foundation of China under Grant 62376046 and Grant 62172067, and in part by the Natural Science Foundation of Chongqing under Grant CSTB2023NSCQ-MSX0341, the Science and Technology Research Program of Chongqing Municipal Education Commission (Grant No. KJQN202200635).
+
+# References
+
+[1] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 2, 3  
+[2] Alexander A Alemi, Ian Fischer, and Joshua V Dillon. Uncertainty in the variational information bottleneck. arXiv preprint arXiv:1807.00906, 2018. 3  
+[3] Xiuli Bi, Bo Liu, Fan Yang, Bin Xiao, Weisheng Li, Gao Huang, and Pamela C Cosman. Detecting generated images by real images only. arXiv preprint arXiv:2311.00962, 2023. 2  
+[4] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 103-120. Springer, 2020.  
+[5] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 5  
+[6] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of ai-generated image detection with clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4356-4366, 2024. 1, 2, 3  
+[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5  
+[8] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 6, 7  
+[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. 1  
+[10] 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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 5  
+[11] 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. 2, 3  
+[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1  
+[13] Shizhe Hu, Zhengzheng Lou, Xiaoqiang Yan, and Yangdong Ye. A survey on information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3
+
+[14] Yan Ju, Shan Jia, Lipeng Ke, Hongfei Xue, Koki Nagano, and Siwei Lyu. Fusing global and local features for generalized ai-synthesized image detection. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3465-3469. IEEE, 2022. 2, 6, 7  
+[15] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 1, 5  
+[16] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5  
+[17] Sohail Ahmed Khan and Duc-Tien Dang-Nguyen. Clipping the deception: Adapting vision-language models for universal deepfake detection. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 1006-1015, 2024. 6, 7  
+[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5  
+[19] Li Lin, Irene Amerini, Xin Wang, Shu Hu, et al. Robust clip-based detector for exposing diffusion model-generated images. arXiv preprint arXiv:2404.12908, 2024. 2  
+[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, 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  
+[21] Bo Liu, Fan Yang, Xiuli Bi, Bin Xiao, Weisheng Li, and Xinbo Gao. Detecting generated images by real images. In European Conference on Computer Vision, pages 95-110. Springer, 2022. 2  
+[22] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5  
+[23] Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. Lare^ 2: Latent reconstruction error based method for diffusion-generated image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17006-17015, 2024. 1  
+[24] Rabeeh Karimi Mahabadi, Yonatan Belinkov, and James Henderson. Variational information bottleneck for effective low-resource fine-tuning. arXiv preprint arXiv:2106.05469, 2021.3  
+[25] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 1  
+[26] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 3  
+[27] 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 preprint arXiv:2112.10741, 2021. 5
+
+[28] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 2, 3, 6, 7  
+[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 5  
+[30] 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, 8  
+[31] 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  
+[32] 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  
+[33] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 5  
+[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5  
+[35] 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, pages 15638-15650, 2022. 8  
+[36] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 6, 7  
+[37] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 6, 7  
+[38] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3
+
+[39] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 1, 2, 5, 6, 7  
+[40] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 1, 6, 7  
+[41] Among Wu and Cheng Deng. Tib: Detecting unknown objects via two-stream information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3  
+[42] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5  
+[43] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 1  
+[44] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021. 2  
+[45] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 5  
+[46] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. Advances in Neural Information Processing Systems, 36, 2024. 5

2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:612e7c18bec565f4ea639888abf9975e5f914852f40adb3b0a504ea9c8157ceb
+size 714411

2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/cd8b4c9e-1814-423f-bdad-2b631dc6a75a_content_list.json

@@ -0,0 +1,1622 @@
+[
+    {
+        "type": "text",
+        "text": "Towards Universal Dataset Distillation via Task-Driven Diffusion",
+        "text_level": 1,
+        "bbox": [
+            169,
+            130,
+            826,
+            151
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ding Qi $^{1*}$  Jian Li $^{2\\dagger}$  Junyao Gao $^{1}$  Shuguang Dou $^{1}$  Ying Tai $^{5}$   \nJianlong Hu $^{2}$  Bo Zhao $^{4}$  Yabiao Wang $^{3,2\\dagger}$  Chengjie Wang $^{4,2}$  Cairong Zhao $^{1\\dagger}$ $^{1}$ Tongji University  $^{2}$ Tencent YouTu Lab  $^{3}$ Zhejiang University  \n $^{4}$ Shanghai Jiao Tong University  $^{5}$ Nanjing University",
+        "bbox": [
+            161,
+            178,
+            833,
+            253
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/273080066be2f11e90e6d1d4791effb0af3c1dd2d9a9a35dc7bff1253d4caa6d.jpg",
+        "image_caption": [
+            "(a) One Universal Dataset Distillation Framework for Mutli-Tasks"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            268,
+            598,
+            455
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/29377090049304128148076e6e350a1f9d66a781866c0624fbfc3fec438bfb75.jpg",
+        "image_caption": [
+            "(b) Improved Performance by UniDD",
+            "Figure 1. (a) We introduce the Universal Dataset Distillation Framework (UniDD), a novel dataset distillation framework supporting a range of vision tasks, including classification, object detection, and segmentation. UniDD incorporates diverse label types, such as class labels, bounding box coordinates, and pixel-level masks, facilitating the synthesis of task-specific datasets. (b) UniDD improves the performance of dataset distillation across diverse tasks on ImageNet [8], Pascal VOC [11, 12], and MS COCO [28]."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            629,
+            279,
+            867,
+            438
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            563,
+            326,
+            578
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Dataset distillation (DD) condenses key information from large-scale datasets into smaller synthetic datasets, reducing storage and computational costs for training networks. However, most recent research has primarily focused on image classification tasks, with limited exploration in detection and segmentation. Two key challenges remain: (i) Task Optimization Heterogeneity, where existing methods focus on class-level information but fail to address the diverse needs of detection and segmentation, and (ii) Inflexible Image Generation, where current generation methods rely on global updates for single-class targets and lack localized optimization for specific object regions. To address these challenges, we propose UniDD, a universal dataset distillation framework built on a task-driven diffusion model for diverse DD tasks, as shown in Fig. 1. Our approach operates in two stages: Universal Task Knowledge Mining, which captures task-relevant information through task",
+        "bbox": [
+            88,
+            595,
+            485,
+            853
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "specific proxy model training, and Universal Task-Driven Diffusion, where these proxies guide the diffusion process to generate task-specific synthetic images. Extensive experiments across ImageNet-1K, Pascal VOC, and MS COCO demonstrate that UniDD consistently outperforms state-of-the-art methods. In particular, on ImageNet-1K with IPC-10, UniDD surpasses previous diffusion-based methods by  $6.1\\%$ , while also reducing deployment costs.",
+        "bbox": [
+            511,
+            564,
+            908,
+            686
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            513,
+            720,
+            645,
+            736
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Dataset distillation (DD) [44], as an innovative data optimization technique in machine learning [13, 14, 20, 22, 45], condenses essential information from large datasets into smaller proxy sets. Training on these highly ensemble proxy sets enables the model to use much fewer resources while providing comparable performance. Prior DD methods excelled in image classification using various techniques like Gradient Matching [26, 48, 50], Distribution Matching [49, 51], Training Trajectory Matching [3, 7], and Decoupled Methods [42, 46].",
+        "bbox": [
+            509,
+            750,
+            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": "*Work done when Ding Qi is an intern in Tencent YouTu Lab, and the advisor is Jian Li.",
+        "bbox": [
+            89,
+            862,
+            482,
+            886
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "† Corresponding author",
+        "bbox": [
+            107,
+            887,
+            246,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "10557",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite advancements in DD, research on key visual tasks—detection and segmentation—remains limited. We identify the following key challenges: (i) Task Optimization Heterogeneity. Existing DD methods are primarily focused on image classification, optimizing distillation by compressing class-level information, but they fall short in addressing the diverse needs of tasks like detection [27] and segmentation, such as object localization and pixel-level information. (ii) Inflexible Image Generation. Current image generation methods rely on global updates for single-class targets and lack localized optimization for specific object regions. These limitations become especially evident in images with multiple objects or complex structures, hindering the ability of existing DD to generate distilled data that meets the requirements of diverse tasks. Consequently, the applicability of DD in the visual domain remains constrained.",
+        "bbox": [
+            89,
+            90,
+            480,
+            345
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address these limitations, we propose a Universal Dataset Distillation (UniDD) framework that optimizes for multi-task learning, including classification, detection, and segmentation, overcoming the shortcomings in task adaptation and image generation. Our approach consists of two stages: Universal Task Knowledge Mining and Universal Task-Driven Diffusion. In the first stage, we introduce Task-Specific Proxy  $(\\mathcal{T}_{\\mathcal{SP}})$  models, trained on real datasets to extract task-specific information. This allows the  $\\mathcal{T}_{\\mathcal{SP}}$  models to capture and store category information for classification, bounding box positions for detection, and mask data for segmentation within their parameters. Additionally, previous research has often overlooked the importance of contextual information, particularly in high-resolution datasets like ImageNet [8], Pascal VOC [11, 12] and MS COCO [28], where rich object-context relationships exist. To address this, we introduce Task-Aligned Contextual Prompting, which uses a Vision-Language model [33] to generate contextual prompts for real images. In the second stage, we introduce a Task-Driven Diffusion for image synthesis. Specifically, the  $\\mathcal{T}_{\\mathcal{SP}}$  models trained in the first stage supervise the denoising process of the diffusion model by calculating the task loss on the predicted clean images. This guides the diffusion model in generating images that match the given labels. Notably, since the  $\\mathcal{T}_{\\mathcal{SP}}$  models already align with the target domain, there is no need to retrain the diffusion model on the real set, significantly reducing the cost in practical applications.",
+        "bbox": [
+            91,
+            351,
+            483,
+            773
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In summary, our contributions are threefold:",
+        "bbox": [
+            109,
+            777,
+            403,
+            791
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We identify key challenges in extending dataset distillation beyond classification and propose the first universal framework for DD across classification, detection, and segmentation tasks.",
+            "- We develop a two-stage approach, Universal Task Knowledge Mining and Universal Task-Driven Diffusion for Data Synthesis, which flexibly synthesizes images for"
+        ],
+        "bbox": [
+            91,
+            794,
+            483,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "multi-tasks. Additionally, we address the previously overlooked need to align potential contextual information within datasets.",
+        "bbox": [
+            526,
+            90,
+            903,
+            135
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- Our method achieves state-of-the-art performance in classification tasks, on the ImageNet-1K dataset, our approach surpasses the previous state-of-the-art method by over  $6.1\\%$ . Additionally, we establish new benchmarks for detection and segmentation on VOC and COCO, showing significant improvements over coreset methods.",
+        "bbox": [
+            513,
+            136,
+            903,
+            227
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related works",
+        "text_level": 1,
+        "bbox": [
+            513,
+            242,
+            660,
+            258
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Dataset Distillation",
+        "text_level": 1,
+        "bbox": [
+            513,
+            268,
+            697,
+            282
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Dataset distillation [44], a dataset reduction method, addresses the challenge of handling substantial data by synthesizing a smaller, representative dataset. Existing dataset distillation methods can be categorized into meta-learning and data matching frameworks based on whether they explicitly mimic the performance of target data. In the meta-learning framework, the distilled data are treated as hyperparameters and optimized in a nested loop fashion according to the trained model's risk. For example, KIP [32] performs kernel ridge regression using the Neural Tangent Kernel (NTK) [25]. The data matching framework includes gradient matching, distribution matching, and trajectory matching. These methods update distilled data by imitating the influence of target data. For example, DC [50] and DM [49] enforce synthetic and real images to have similar gradients or distribution at every training step. MTT [3] proposes to imitate long-range trajectories of optimization steps for real images, thereby better mimicking the training dynamics of real images for synthetic images. However, the aforementioned gradient matching methods require a large number of iterations. To address this, DREAM [31] and IDC [26] focus on more efficient dataset distillation training. Due to the unsatisfactory performance of DM, CAFE [43] and IDM [51] propose improved distribution matching between the real and synthetic data. For the trajectory matching method, DATM [19] scales trajectory matching-based methods to larger synthetic datasets, achieving lossless dataset distillation.",
+        "bbox": [
+            511,
+            291,
+            906,
+            714
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Despite significant progress in previous work, there have been few breakthroughs in dataset distillation for detection and segmentation tasks. In this work, we introduce the first universal framework that simultaneously handles classification, object detection, and semantic segmentation, significantly expanding the generality and scope of dataset distillation research.",
+        "bbox": [
+            511,
+            715,
+            905,
+            819
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Diffusion Models",
+        "text_level": 1,
+        "bbox": [
+            513,
+            832,
+            681,
+            847
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Diffusion models [15, 16, 37, 47] are generative models that show superior fidelity and diversity compared to GANs [17]. It aims to learn the mapping from a Gaussian",
+        "bbox": [
+            511,
+            854,
+            905,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "10558",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/8c94ce46423189613163e117cc42deebceb0535c1c09b9fdc51f7a33944a2e6a.jpg",
+        "image_caption": [
+            "(a) Synthetic images for classification."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            90,
+            354,
+            287
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/6a727fa95fd69b053327cd236578d7e039dc29ffce88b311fa0bedf52568970d.jpg",
+        "image_caption": [
+            "(b) Synthetic images for detection."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            367,
+            90,
+            629,
+            287
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/b749891040b76d77d0cb4ecba985295e6e5ca13b40b01681cb87fd578d07b433.jpg",
+        "image_caption": [
+            "(c) Synthetic images for segmentation.",
+            "Figure 2. Visualization of synthetic images generated by UniDD. (a) For classification, images are synthesized on the ImageWoof [8] dataset. (b) For detection, images are synthesized on the Pascal VOC dataset with specified bounding box coordinates. (c) For segmentation, images are synthesized on the Pascal VOC [11, 12] dataset with pixel-level masks. Our UniDD is the first approach to enable high-realism data distillation across classification, detection, and segmentation tasks."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            642,
+            90,
+            903,
+            287
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "distribution to the real dataset distribution by adding noise in the diffusion process and reversing the noise to the raw image in the denoising process. In the forward process, diffusion models progressively add noise to a clean sample  $z_0$ , resulting in a noisy sample  $z_t$ :",
+        "bbox": [
+            89,
+            378,
+            483,
+            455
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nz _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\quad \\epsilon \\sim \\mathcal {N} (0, I). \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            147,
+            460,
+            482,
+            479
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In the reverse process, a denoising network  $\\epsilon_{\\theta}$  estimates the noise and iteratively recovers the clean sample  $z_0$ :",
+        "bbox": [
+            89,
+            484,
+            483,
+            513
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\epsilon_ {\\theta} (z _ {t}, t) \\approx \\frac {z _ {t} - \\sqrt {\\alpha_ {t}} z _ {0}}{\\sqrt {1 - \\alpha_ {t}}}. \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            200,
+            518,
+            482,
+            554
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "DDIM [40] provides the following formula for predicting a clean data sample:",
+        "bbox": [
+            89,
+            556,
+            483,
+            587
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {z} _ {0} = \\frac {z _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} (z _ {t} , t)}{\\sqrt {\\alpha_ {t}}}. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            187,
+            590,
+            482,
+            627
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This formula predicts the clean data point  $\\hat{z}_0$  from the noisy sample  $z_{t}$ , enabling deterministic sampling for efficient and high-quality image synthesis.",
+        "bbox": [
+            89,
+            631,
+            483,
+            675
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To introduce additional conditions into the diffusion model, classifier guidance [9] has been proposed to train an additional classifier on the noise to guide the model to generate images that are closer to the specified condition. Additionally, [23] introduces classifier-free guidance that trains both an unconditioned and conditioned model simultaneously, achieving more efficient guidance. Following these, recent methods [1, 10, 30] have been proposed to generate high-quality images from diverse conditions.",
+        "bbox": [
+            89,
+            676,
+            483,
+            811
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Universal Dataset Distillation Framework",
+        "text_level": 1,
+        "bbox": [
+            89,
+            823,
+            465,
+            839
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Problem Definition",
+        "text_level": 1,
+        "bbox": [
+            89,
+            848,
+            272,
+            863
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Universal Dataset Distillation. Current dataset distillation research primarily targets classification tasks and lacks",
+        "bbox": [
+            89,
+            869,
+            483,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "a generalizable approach for detection and segmentation. To address this, we propose a universal dataset distillation standard that constructs a compact synthetic dataset  $\\mathcal{S} = (\\hat{x}_1, \\hat{y}_1), \\dots, (\\hat{x}_{|\\mathcal{S}|}, \\hat{y}_{|\\mathcal{S}|})$  from a larger real dataset  $\\mathcal{T} = (x_1, y_1), \\dots, (x_{|\\mathcal{T}|}, y_{|\\mathcal{T}|})$ , with  $|S| \\ll |\\mathcal{T}|$ . The primary difference between tasks lies in label structure: classification uses single-class labels, detection includes bounding boxes with class information, and segmentation requires pixel-level masks. Our goal is for models trained on the distilled dataset  $\\mathcal{S}$  to perform comparably to those trained on the full dataset  $\\mathcal{T}$  across classification, detection, and segmentation tasks, as formalized by:",
+        "bbox": [
+            511,
+            378,
+            906,
+            561
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbb {E} _ {(x, y) \\sim P _ {D}} \\left| \\ell \\left(\\theta_ {\\mathcal {T}} ^ {t a s k} (x), y\\right) - \\ell \\left(\\theta_ {\\mathcal {S}} ^ {t a s k} (x), y\\right) \\right| \\leq \\epsilon , \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            537,
+            570,
+            906,
+            590
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $P_{D}$  denotes the distribution of test data,  $\\ell$  represents the performance metric specific to each task,  $task\\in$  {class,detect,segment}, and  $\\epsilon$  is a small tolerance value. For each task, the parameter set  $\\theta_{\\mathcal{T}}^{task}$  on the real dataset  $\\mathcal{T}$  is optimized as follows:",
+        "bbox": [
+            511,
+            598,
+            906,
+            674
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\theta_ {\\mathcal {T}} ^ {t a s k} = \\arg \\min _ {\\theta} \\mathbb {E} _ {(x, y) \\in \\mathcal {T}} \\left[ \\ell \\left(\\phi_ {\\theta} ^ {t a s k} (x), y\\right) \\right], \\qquad (5)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            560,
+            683,
+            906,
+            708
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Similarly,  $\\theta_S^{task}$  is optimized using the synthetic dataset  $S$ .",
+        "bbox": [
+            511,
+            717,
+            898,
+            734
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Universal Task Knowledge Mining",
+        "text_level": 1,
+        "bbox": [
+            511,
+            742,
+            815,
+            758
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Task-Specific Proxy Training. Building a universal DD framework is challenging due to the diverse optimization objectives in classification, detection, and segmentation tasks. Recent decoupled DD studies have reached a consensus [34, 42, 46]: a fully trained network can capture and retain essential dataset information. Inspired by this, we extend to general tasks, training a classifier, detector, and segmenter on task-specific datasets to extract critical information—class labels for classification, bounding boxes for",
+        "bbox": [
+            511,
+            763,
+            906,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "10559",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/697bfc9651c9aab0d9e97c88dab77dad9859a6862ec3e7f7180dc844b1d8405e.jpg",
+        "image_caption": [
+            "Figure 3. Universal Dataset Distillation Framework. This framework consists of two stages: in the Universal Task Knowledge Mining stage, we introduce an easily deployable task-specific proxy  $(\\mathcal{T}_{\\mathcal{SP}})$  model trained on the target dataset and a vision-language model to extract task-aligned context. In the Universal Task-Driven Diffusion stage, we apply the gradient of the supervision loss, derived from the  $\\mathcal{T}_{\\mathcal{SP}}$  model's prediction of the clean image  $\\hat{z}_0$ , at each sampling step."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            93,
+            88,
+            906,
+            375
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "detection, and pixel-level masks for segmentation. We call this Task-Specific Proxy  $(\\mathcal{T}_{\\mathcal{SP}})$  Model, defined as:",
+        "bbox": [
+            89,
+            446,
+            482,
+            477
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\theta = \\arg \\min  _ {\\theta} \\mathcal {L} _ {\\text {t a s k}} (\\mathcal {F} (x), y), \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            187,
+            489,
+            482,
+            510
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{F}$  represents the  $\\mathcal{T}_{\\mathcal{SP}}$  model trained on the target data,  $\\theta$  denotes the model's parameters,  $\\mathcal{L}_{task}$  represents the loss function specific to the task.",
+        "bbox": [
+            89,
+            522,
+            482,
+            566
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The  $\\mathcal{T}_{\\mathcal{SP}}$  model choice is flexible, allowing adaptation for optimal performance on target datasets. For example, ResNet-18 [21] can serve as a lightweight, fast-deploy option for simple classification, while Faster R-CNN [36] may be used for complex detection tasks to capture key detection information. Unlike previous diffusion-based dataset distillation methods, our approach relies on task-driven learning to align synthetic and real datasets, avoiding costly diffusion model fine-tuning on the target dataset and significantly reducing deployment overhead.",
+        "bbox": [
+            89,
+            568,
+            482,
+            718
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Task-Aligned Contextual Prompting. Previous diffusion-based dataset distillation has largely focused on classification tasks and overlooked the role of natural language in guiding image generation. For diverse tasks, textual prompts must go beyond describing the primary objects in images. As inputs to Stable Diffusion models, prompts can also guide the generation process. To address this, we use a vision-language model to extract natural language descriptions and design task-specific questions to generate task-aligned prompts, a method we call Task-Aligned Contextual Prompting. Following GPT-4o [33], we create tailored prompts for different tasks. For classification, prompts de",
+        "bbox": [
+            89,
+            719,
+            482,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "scribe object classes, while for detection and segmentation, they also include object relationships and the interaction between objects and the background, providing richer contextual guidance.",
+        "bbox": [
+            511,
+            446,
+            906,
+            507
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Universal Task-Driven Diffusion.",
+        "text_level": 1,
+        "bbox": [
+            511,
+            516,
+            802,
+            531
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Inspired by recent advancements in guidance diffusion techniques [1, 9], we propose a Task-Driven Diffusion approach for Dataset Distillation. As illustrated in Fig. 3, during each sampling step  $S(z_{t}, t)$  of the diffusion process, the trained  $\\mathcal{T}_{\\mathcal{SP}}$  model  $\\mathcal{F}$  provides task-specific guidance. Given an label  $\\tilde{a}$ , the original noise prediction  $\\epsilon_{\\theta}(z_{t}, t, \\tilde{c})$  is adjusted using the output of  $\\mathcal{F}$ . The adjusted noise prediction  $\\hat{\\epsilon}$  is formulated as follows:",
+        "bbox": [
+            511,
+            537,
+            906,
+            657
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\epsilon} _ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right) = \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right) + \\sqrt {1 - \\alpha_ {t}} \\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} \\left(\\tilde {a}, \\mathcal {F} \\left(z _ {t}\\right)\\right), \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            524,
+            667,
+            906,
+            685
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{L}_{\\mathrm{task}}$  represents the task-specific loss function and  $\\tilde{a} = \\{a_{cl}, a_{bbox}, a_{mask}\\}$  may denote a class label  $a_{cl}$ , a bounding box  $a_{bbox}$ , or a pixel-level segmentation mask  $a_{mask}$ . This adjustment mechanism effectively guides the diffusion process to produce data closely aligned with the intended distribution for each task.  $\\tilde{c}$  is a textual prompt.",
+        "bbox": [
+            511,
+            694,
+            905,
+            784
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Because  $\\mathcal{F}$  is trained on clean images, it may not perform well directly on noisy inputs. Therefore, we first use Equation 3 to estimate the clean image  $\\hat{z}_0$ . Once  $\\hat{z}_0$  is obtained, we apply  $\\mathcal{F}$  to compute the task-driven loss as follows:",
+        "bbox": [
+            511,
+            785,
+            905,
+            845
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\epsilon} _ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right)) = \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right) + s (t) \\cdot \\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} (\\tilde {a}, \\mathcal {F} (\\hat {z} _ {0})), \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            519,
+            857,
+            906,
+            875
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $s(t)$  is the task-driven strength at each sampling step,",
+        "bbox": [
+            511,
+            885,
+            905,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "10560",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "defined as:",
+        "bbox": [
+            89,
+            90,
+            166,
+            104
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ns (t) = \\lambda_ {s} \\cdot \\sqrt {1 - \\alpha_ {t}}, \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            215,
+            104,
+            480,
+            122
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "with  $\\lambda_{s}$  being a constant coefficient controlling the overall intensity of the task-driven. The term  $\\nabla_{z_t}\\mathcal{L}_{task}(\\tilde{a},\\mathcal{F}(\\hat{z}_0))$  represents the gradient of the loss function computed based on the  $\\mathcal{T}_{\\mathcal{SP}}$  model's prediction of the clean image  $\\hat{z}_0$ . Specifically, this gradient is computed as:",
+        "bbox": [
+            89,
+            128,
+            483,
+            204
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} (\\tilde {a}, \\mathcal {F} (\\hat {z} _ {0})) = \\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} \\left(\\tilde {a}, \\mathcal {F} \\left(\\frac {z _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta}}{\\sqrt {\\alpha_ {t}}}\\right)\\right). \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            89,
+            213,
+            483,
+            258
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Unlike previous approaches [18, 41] that required finetuning the diffusion model, such as those involving SD [37], to align with new data domains, our method leverages a  $\\mathcal{T}_{\\mathcal{SP}}$  model trained specifically on the target dataset. This model encapsulates critical information about the target task, effectively driving the diffusion process to generate data that is well-aligned with the target task without the need for additional finetuning. This not only conserves time and resources but also mitigates the risk of overfitting the diffusion model to the specific task.",
+        "bbox": [
+            89,
+            267,
+            483,
+            417
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Proxy-Driven High-Realism Refinement. To ensure the high realism and accuracy of the generated images, we employ a trained  $\\mathcal{T}_{\\mathcal{SP}}$  model, to score each image. By setting a well-defined threshold, we filter out low-quality images based on their scores. The remaining images are then relabeled using the same  $\\mathcal{T}_{\\mathcal{SP}}$  model. This combined process of filtering and relabeling can be expressed as:",
+        "bbox": [
+            89,
+            417,
+            483,
+            525
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\{x, y \\} = \\left\\{x _ {i}, \\mathcal {F} _ {\\text {l a b e l}} \\left(x _ {i}\\right) \\mid \\operatorname {s c o r e} \\left(\\mathcal {F} \\left(x _ {i}\\right)\\right) \\geq \\tau \\right\\} \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            122,
+            535,
+            482,
+            554
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{F}(x_i)$  represents the model's output score for image  $x_{i}$ ,  $\\tau$  is the quality threshold, and  $\\mathcal{F}_{\\mathrm{label}}$  generates the accurate label  $y_{i}$  for the filtered image. This ensures that both the generated images and their labels are of high quality and closely match the desired data distribution.",
+        "bbox": [
+            89,
+            564,
+            483,
+            638
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            642,
+            223,
+            659
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Experimental Setups",
+        "text_level": 1,
+        "bbox": [
+            89,
+            667,
+            287,
+            683
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Datasets and Evaluation Metric. To enhance practicality, we prioritize high-resolution and large-scale datasets. (1) Classification task: our experiments are primarily conducted on the full ImageNet 1K dataset [8], with IPC settings commonly set to 10 and 50, following the configurations in [41]. Additionally, based on the dataset partitioning settings from previous work [3], we also use the ImageNet 10-class subsets, such as ImageWoof, ImageFruit, ImageNette, and ImageMeow. These 10-class subsets contain very similar categories, making them challenging for dataset distillation. Top-1 accuracy  $(\\%)$  is reported in the experimental tables. (2) Object detection task: we conduct experiments on the widely-used Pascal VOC [11, 12] and MS COCO [28] datasets, using mAP (Mean Average",
+        "bbox": [
+            89,
+            688,
+            483,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Algorithm 1 Universal Dataset Distillation Approach",
+        "text_level": 1,
+        "bbox": [
+            514,
+            90,
+            870,
+            106
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Input: Real data  $\\mathcal{T}$ , Diffusion model  $\\epsilon_{\\theta}$ ,  $\\mathcal{T}_{\\mathcal{SP}}$  model  $\\mathcal{F}$ , Sampling steps  $T$ , Task-Driven strength  $s(t)$",
+        "bbox": [
+            513,
+            107,
+            905,
+            137
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Output: Synthetic dataset  $S$ .",
+        "bbox": [
+            514,
+            138,
+            710,
+            152
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1: Step I: Universal Task Knowledge Mining",
+            "2: Train  $\\mathcal{T}_{\\mathcal{SP}}$  model  $\\mathcal{F}$  on  $\\mathcal{T}$  using Equ. 6",
+            "3: Use GPT-4o to generate task-aligned context for classification, detection, and segmentation datasets"
+        ],
+        "bbox": [
+            522,
+            155,
+            903,
+            215
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4: Step II: Universal Task-Driven Diffusion for Data Synthetic",
+        "bbox": [
+            522,
+            215,
+            903,
+            244
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "5: for each sampling step  $t = 1$  to  $T$  do",
+            "6: Generate noisy sample  $z_{t}$ .",
+            "7: Predict clean image  $\\hat{z}_0$  using Equ. 3",
+            "8: Calculate guidance loss  $\\nabla_{z_t}\\mathcal{L}_{task}(c,\\mathcal{F}(\\hat{z}_0))$",
+            "9: Adjust noise prediction using Equ. 8",
+            "10: Update  $z_{t-1}$  using adjusted noise prediction.",
+            "11: end for",
+            "12: Score each generated image  $x_{i}$  using  $\\mathcal{F}(x_i)$ .",
+            "13: Filter and relabel synthetic set based on threshold  $\\tau$  using Equ. 11."
+        ],
+        "bbox": [
+            519,
+            246,
+            903,
+            396
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Precision) and  $AP_{50}$  (Average Precision at IoU threshold of 0.50) as evaluation metric. (3) Semantic segmentation task: we utilize the Pascal VOC dataset, with mIoU (Mean Intersection over Union) as evaluation metrics.",
+        "bbox": [
+            511,
+            428,
+            905,
+            488
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. Unless otherwise specified, the Task-Specific Proxy  $\\mathcal{T}_{\\mathcal{SP}}$  model used for the classification task is ResNet-18 [21], the detector is Faster R-CNN [36], and the segmentation model is LR-ASPP [24], all utilizing the official pre-trained weights from PyTorch. For the diffusion model, we use SD v1.5 [37] as the baseline. For quantitative comparisons, we generate  $256 \\times 256$  images for classification tasks and  $512 \\times 512$  images for detection and segmentation tasks. Synthetic images are created using 250 denoising steps based on a predefined number of random noise samples, aligned with the IPC settings. The threshold  $\\tau$  for Proxy-Driven High-Realism Refinement is set to 0.9 for classification, 0.85 for detection, and 0.75 for segmentation. Generally, this operation results in a data discard rate of approximately  $5 - 10\\%$ . All experiments are conducted on NVIDIA V100 GPUs, with further details provided in the special experiments section. We generate the distilled dataset 3 times, train the network from scratch 5 times per dataset (with pre-trained backbone weights for the detector and segmenter), and evaluate on the original test set, reporting results as  $\\bar{x} \\pm std$ .",
+        "bbox": [
+            511,
+            489,
+            906,
+            806
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Results on Classification",
+        "text_level": 1,
+        "bbox": [
+            511,
+            816,
+            736,
+            832
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We perform comprehensive comparisons across datasets of various scales. For large-scale datasets ImageNet-1k, we compare TESLA [7], SRe $^2$ L [46], D $^4$ M [41], RDED [42], and MiniMax Diffusion [18] (abbreviated as MiMxDiff in",
+        "bbox": [
+            511,
+            839,
+            906,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "10561",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/af2b02f0dee62e5c55deefb8ab940f3c5d2775ef44dd83859d3426b594a6e0c0.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 1. Performance comparison with state-of-the-art methods on large-scale dataset ImageNet-1K. All methods employ ResNet18 as test model. All standard deviations in this table are less than one. The results of Full dataset come from the official PyTorch. The best results are marked in bold, while the second-best results are underlined."
+        ],
+        "table_body": "<table><tr><td>Dataset</td><td>IPC</td><td>Method</td><td>Accuracy(%)</td></tr><tr><td rowspan=\"12\">ImageNet-1K</td><td>1000</td><td>Full dataset</td><td>69.8</td></tr><tr><td rowspan=\"6\">10</td><td>TESLA</td><td>7.7</td></tr><tr><td>SRe2L</td><td>21.3</td></tr><tr><td>D4M</td><td>27.9</td></tr><tr><td>RDED</td><td>42.0</td></tr><tr><td>MiMxDiff</td><td>44.3</td></tr><tr><td>UniDD (Ours)</td><td>50.4 (↑6.1)</td></tr><tr><td rowspan=\"5\">50</td><td>SRe2L</td><td>46.8</td></tr><tr><td>D4M</td><td>55.2</td></tr><tr><td>RDED</td><td>56.5</td></tr><tr><td>MiMxDiff</td><td>58.6</td></tr><tr><td>UniDD (Ours)</td><td>62.8 (↑4.2)</td></tr></table>",
+        "bbox": [
+            93,
+            88,
+            480,
+            311
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "the table). For small-scale dataset ImageNet-Subset, we compare Random selection, DM [49], MTT [3], HaBa [29], DataDAM [38], and MiniMax Diffusion.",
+        "bbox": [
+            89,
+            431,
+            482,
+            477
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Quantitative comparison on ImageNet-1K and ImageNet-Subset. As shown in the Tab 1, we utilize ResNet-18 as the  $\\mathcal{T}_{\\mathcal{SP}}$  model, trained on ImageNet-1K. We report results for IPC-10 and IPC-50, evaluating all outcomes with the ResNet-18 model. To ensure a fair comparison, our training and testing pipeline is aligned with that of RDED and  $\\mathrm{D^4M}$ . Our method, UniDD, achieves state-of-the-art performance on this large-scale dataset. Specifically, at IPC-10, UniDD surpasses optimization-based methods, delivering substantial gains over TESLA and  $\\mathrm{SRe}^2\\mathrm{L}$ . Additionally, it demonstrates significant advantages over diffusion-based methods, outperforming  $\\mathrm{D^4M}$  by  $22.5\\%$  and MiniMax Diffusion by  $6.1\\%$ . These results underscore UniDD's effectiveness by using a lightweight  $\\mathcal{T}_{\\mathcal{SP}}$  model to guide SD denoising, aligning generative and target domains without requiring diffusion model finetuning, thus advancing performance in dataset distillation.",
+        "bbox": [
+            89,
+            478,
+            482,
+            748
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As shown in the Tab 2, we conduct experiments on the ImageNet-subset under the IPC-10 setting. For consistency, the test model is standardized to a 5-layer ConvNet across all methods, except for MiniMax Diffusion, which uses a 6-layer ConvNet. Our method stands out among the comparison methods, outperforming the second-best by  $8.2\\%$  on ImageWoof,  $9.1\\%$  on ImageFruit, and  $6.6\\%$  on ImageMeow. Additionally, our method surpasses the diffusion-based MiniMax Diffusion by  $9.8\\%$  on ImageWoof, further demonstrating the effectiveness of UniDD approach in ag",
+        "bbox": [
+            89,
+            750,
+            482,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "gregating information from the original dataset.",
+        "bbox": [
+            511,
+            90,
+            828,
+            106
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Qualitative Result. In Fig. 2 (a), we present the synthetic images generated by UniDD on ImageWoof. Benefiting from the powerful generative capability of SD and the use of a  $\\mathcal{T}_{\\mathcal{SP}}$  model to align the real and synthetic data domains, our synthetic images for the classification task exhibit significantly higher realism compared to optimization-based methods [3, 46, 50]. However, unlike other diffusion-based dataset distillation methods [18, 41], we do not finetune SD, greatly reducing the deployment time of UniDD.",
+        "bbox": [
+            511,
+            107,
+            903,
+            243
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Results on Detection and Segmentation",
+        "text_level": 1,
+        "bbox": [
+            511,
+            256,
+            848,
+            272
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We deploy our UniDD framework for detection and segmentation tasks and conduct experiments on the Pascal VOC and MS COCO datasets. Since dataset distillation is rarely applied to these tasks, we reference earlier dataset distillation works (DD [44], DC [50]) and compare UniDD with core-set selection methods. Specifically, we construct the following baseline methods: Random [35], K-center [39], and Herding [2, 5]. Additionally, we develop a method that selects a subset with balanced class representation, which we refer to as Uniform.",
+        "bbox": [
+            511,
+            279,
+            903,
+            429
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Quantitative Result. In Table 3, we compare coreset selection methods with UniDD across different compression ratios on the object detection task. According to the official MMDetection implementation [4], the detector Faster R-CNN [36] achieves  $51.4\\%$  mAP and  $80.3\\%$ $\\mathrm{AP}_{50}$  on Pascal VOC with the full dataset, and  $32.6\\%$  mAP and  $51.4\\%$ $\\mathrm{AP}_{50}$  on MS COCO, providing an approximate performance upper bound. On Pascal VOC, UniDD demonstrates significant performance gains, achieving  $7.7\\%$  mAP and  $19.2\\%$ $\\mathrm{AP}_{50}$  higher than the Random method at an extremely low compression ratio of  $0.5\\%$ . Additionally, UniDD surpasses Uniform by  $11.1\\%$  mAP and  $21.2\\%$ $\\mathrm{AP}_{50}$  at a  $1\\%$  compression ratio. On the more challenging MS COCO dataset, UniDD also shows a clear advantage, surpassing the best coreset selection method by  $3.7\\%$ ,  $3.4\\%$ , and  $3.4\\%$  mAP at compression ratios of  $0.25\\%$ ,  $0.5\\%$ , and  $1\\%$ , respectively. This indicates UniDD's robustness, achieving strong results even on more complex datasets, which can be attributed to the Task Knowledge Mining in the first stage that extracts sufficient task-relevant information.",
+        "bbox": [
+            511,
+            431,
+            903,
+            733
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In Table 4, we present experimental results on the VOC dataset for the semantic segmentation task. Based on the MMSegmentation [6] implementation, LR-ASPP with a MobileNetV3 backbone achieves  $32.9\\%$  mIoU on the full dataset. UniDD outperforms coreset selection methods at all compression ratios, demonstrating its effectiveness in dataset distillation for segmentation tasks. This result further underscores UniDD's versatility across different tasks. By leveraging task-driven diffusion synthesis alongside task knowledge mining, UniDD generates synthetic data well-suited to complex, pixel-level predictions, illustrating its",
+        "bbox": [
+            511,
+            734,
+            906,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "10562",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/9ac17681a769dc58ea4f70fb98a6be3e550ad19b2c698138fcc146c167facb21.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Random</td><td>DM</td><td>MTT</td><td>HaBa</td><td>DataDAM</td><td>MiMxDiff</td><td>UniDD (Ours)</td><td>Full</td></tr><tr><td>ImageNette</td><td>47.7±2.4</td><td>58.1±0.3</td><td>63.0 ±1.3</td><td>64.7 ±1.6</td><td>59.4±0.4</td><td>62.0±0.2</td><td>66.0±0.7</td><td>87.4 ±1.0</td></tr><tr><td>ImageWoof</td><td>27.0±1.9</td><td>31.4±0.5</td><td>35.8 ±1.8</td><td>38.6 ±1.2</td><td>34.4±0.4</td><td>37.0±1.0</td><td>46.8±0.9</td><td>67.0 ±1.3</td></tr><tr><td>ImageFruit</td><td>21.4±1.2</td><td>-</td><td>40.3 ±1.3</td><td>42.5 ±1.6</td><td>-</td><td>-</td><td>51.6±0.5</td><td>63.9 ±2.0</td></tr><tr><td>ImageMeow</td><td>29.0±1.1</td><td>-</td><td>40.4 ±2.2</td><td>42.9 ±0.9</td><td>-</td><td>-</td><td>49.5±0.7</td><td>66.7 ±1.1</td></tr></table>",
+        "bbox": [
+            99,
+            88,
+            895,
+            189
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/8ffe4f26155fc76f3ae3557195cd552a06d0dcbf656e869a985da23a576b5bcc.jpg",
+        "table_caption": [
+            "Table 2. Performance comparison with other state-of-the-art methods on ImageNet-Subset. IPC is set to 10. The best results are marked in bold, while the second-best results are underlined. “-” were not officially implemented in the original papers."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"3\">Methods</td><td colspan=\"4\">Object Detection</td></tr><tr><td colspan=\"2\">Pascal VOC</td><td colspan=\"2\">MS COCO</td></tr><tr><td>mAP</td><td>AP50</td><td>mAP</td><td>AP50</td></tr><tr><td>Ratio</td><td colspan=\"2\">0.5%</td><td colspan=\"2\">0.25%</td></tr><tr><td>Random</td><td>0.8±0.2</td><td>3.1±0.4</td><td>0.5±0.1</td><td>1.7±0.3</td></tr><tr><td>Uniform</td><td>0.9±0.1</td><td>3.4±0.3</td><td>0.8±0.2</td><td>2.4±0.5</td></tr><tr><td>K-Center</td><td>0.5±0.1</td><td>2.1±0.3</td><td>0.4±0.1</td><td>1.5±0.2</td></tr><tr><td>Herding</td><td>0.6±0.2</td><td>2.4±0.2</td><td>0.5±0.1</td><td>1.8±0.4</td></tr><tr><td>UniDD (Ours)</td><td>8.5±0.4</td><td>22.3±0.6</td><td>4.5±0.3</td><td>10.3±0.4</td></tr><tr><td>Ratio</td><td colspan=\"2\">1%</td><td colspan=\"2\">0.5%</td></tr><tr><td>Random</td><td>4.2±0.5</td><td>13.7±0.6</td><td>3.7±0.2</td><td>10.1±0.3</td></tr><tr><td>Uniform</td><td>5.7±0.2</td><td>17.7±0.4</td><td>3.4±0.4</td><td>9.5±0.6</td></tr><tr><td>K-Center</td><td>3.6±0.6</td><td>12.3±0.3</td><td>3.2±0.5</td><td>9.3±0.5</td></tr><tr><td>Herding</td><td>3.5±0.5</td><td>11.9±0.5</td><td>3.5±0.3</td><td>9.7±0.3</td></tr><tr><td>UniDD (Ours)</td><td>16.8±0.5</td><td>38.9±0.7</td><td>7.1±0.4</td><td>16.9±0.3</td></tr><tr><td>Ratio</td><td colspan=\"2\">2%</td><td colspan=\"2\">1%</td></tr><tr><td>Random</td><td>12.4±0.4</td><td>34.3±0.5</td><td>7.2±0.8</td><td>17.3±0.9</td></tr><tr><td>Uniform</td><td>13.8±0.3</td><td>36.2±0.4</td><td>7.4±0.5</td><td>17.6±0.5</td></tr><tr><td>K-Center</td><td>10.9±0.6</td><td>29.3±0.6</td><td>6.1±0.3</td><td>15.4±0.6</td></tr><tr><td>Herding</td><td>10.4±0.4</td><td>28.7±0.7</td><td>6.7±0.4</td><td>16.3±0.7</td></tr><tr><td>UniDD (Ours)</td><td>23.9±0.5</td><td>48.5±0.6</td><td>10.8±0.4</td><td>22.5±0.5</td></tr><tr><td>Full</td><td>51.4±0.8</td><td>80.3±0.4</td><td>32.6±0.7</td><td>51.4±0.8</td></tr></table>",
+        "bbox": [
+            94,
+            253,
+            480,
+            563
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/18f5cb838b2dda3a4d1303afa850090e13a56caf7efc736833c2f0c0ee600c6c.jpg",
+        "table_caption": [
+            "Table 3. Performance comparison with coreset selection methods on Pascal VOC and MS COCO for the object detection task, using Faster R-CNN as the detector. \"Ratio\" denotes the compression ratio."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"4\">Semantic Segmentation (mIoU)</td></tr><tr><td>Ratio: 3.5%</td><td>Ratio: 7%</td><td>Ratio: 14%</td><td>Full</td></tr><tr><td>Random</td><td>6.7±0.4</td><td>9.6±0.5</td><td>11.3±0.4</td><td></td></tr><tr><td>Uniform</td><td>6.5±0.6</td><td>9.3±0.4</td><td>12.1±0.5</td><td></td></tr><tr><td>K-Center</td><td>6.1±0.4</td><td>8.5±0.6</td><td>10.4±0.7</td><td>32.9±0.6</td></tr><tr><td>Herding</td><td>6.2±0.7</td><td>8.7±0.4</td><td>10.6±0.3</td><td></td></tr><tr><td>UniDD (Ours)</td><td>10.3±0.5</td><td>12.3±0.3</td><td>14.9±0.7</td><td></td></tr></table>",
+        "bbox": [
+            94,
+            646,
+            480,
+            742
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "versatility across a range of vision tasks.",
+        "bbox": [
+            89,
+            838,
+            359,
+            853
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Qualitative Result. As shown in Fig. 2 (b) and 2 (c), we present visualizations of detection and segmentation data synthesized using the UniDD framework. These images",
+        "bbox": [
+            89,
+            854,
+            483,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "are generated with  $\\mathcal{T}_{\\mathcal{SP}}$  models, Faster R-CNN for detection and LR-ASPP for segmentation, which effectively steer the diffusion process toward the specified labels, bounding boxes, and masks. Previous research in dataset distillation has predominantly focused on single-category label synthesis, where each image aligns with a single class. In contrast, our UniDD framework overcomes this limitation, enabling dataset distillation for more complex, multi-class tasks. This advancement not only broadens the applicability of dataset distillation but also offers valuable insights for future research in this domain.",
+        "bbox": [
+            511,
+            256,
+            906,
+            421
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            511,
+            433,
+            666,
+            450
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Component Analysis. We conduct a comprehensive ablation study on the proposed UniDD, analyzing its components: (1) Baseline (a standard SD v1.5 checkpoint), (2) Task-driven Diffusion with the  $\\mathcal{T}_{\\mathcal{SP}}$  model, (3) Task-Aligned Contextual Prompting (abbreviated as Context), and (4) Proxy-Driven High-Realism Refinement (abbreviated as Refine). As shown in Table 5, the proposed task-driven approach significantly improves the performance of the baseline, demonstrating that this guidance effectively bridges the gap between real and synthetic datasets. The contextual information is more effective for more complex detection and segmentation datasets, as it reveals underlying target relationships. The refinement module also proves to be effective, as the randomness of the SD model can lead to the generation of images that deviate from the annotations or result in failed generations.",
+        "bbox": [
+            511,
+            455,
+            908,
+            699
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/7f22ca9639b56d5f84fd761931c9b64315b1de352de9424b41fd7c3c733dfecb.jpg",
+        "table_caption": [
+            "Table 4. Performance comparison with coreset selection methods on Pascal VOC for the semantic segmentation task, using LR-ASPP as the segmenter. The metric is mIoU, and \"Ratio\" denotes the compression ratio."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"3\">Methods</td><td colspan=\"4\">Dataset</td></tr><tr><td colspan=\"2\">Cls. / Accuracy (%)</td><td>Det. / mAP (%)</td><td>Seg. / mIoU (%)</td></tr><tr><td>ImageNette</td><td>ImageWoof</td><td>VOC</td><td>VOC</td></tr><tr><td>Baseline</td><td>60.4±0.9</td><td>39.1±1.1</td><td>11.5±0.2</td><td>7.2±0.3</td></tr><tr><td>Baseline+TSP</td><td>64.2±0.8</td><td>45.3±0.6</td><td>14.7±0.4</td><td>8.9±0.5</td></tr><tr><td>Baseline+TSRP+Context</td><td>64.7±0.4</td><td>46.1±0.7</td><td>15.9±0.3</td><td>9.6±0.4</td></tr><tr><td>Baseline+TSP+Context+Refine</td><td>66.0±0.7</td><td>46.8±0.9</td><td>16.8±0.5</td><td>10.3±0.5</td></tr></table>",
+        "bbox": [
+            514,
+            712,
+            903,
+            787
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 5. Ablation study on UniDD. For classification datasets, IPC is set to 10 with ResNet-18 as the test model. For detection on VOC, the compression ratio is set to  $1\\%$  with Faster R-CNN as the test model. For segmentation on VOC, the ratio is set to  $3.5\\%$  with LR-ASPP as the test model.",
+        "bbox": [
+            511,
+            797,
+            906,
+            867
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Cross-Architecture Analysis. Cross-Architecture experi",
+        "bbox": [
+            511,
+            885,
+            903,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "10563",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/8167e876d94a483cab78fcb5c4737e0d923e8288bd77a481dbabd652b184c829.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Dataset</td><td rowspan=\"2\">Test Model</td><td colspan=\"3\">\\( T_{SP} \\)Model</td></tr><tr><td>ResNet-18</td><td>ResNet-50</td><td>ResNet-101</td></tr><tr><td rowspan=\"3\">ImageNette</td><td>ResNet-18</td><td>66.0±0.7</td><td>68.8±0.7</td><td>69.5±0.7</td></tr><tr><td>ResNet-50</td><td>67.2±0.6</td><td>69.5±1.1</td><td>70.7±0.9</td></tr><tr><td>ResNet-101</td><td>69.3±0.9</td><td>70.9±1.2</td><td>72.1±0.6</td></tr><tr><td rowspan=\"3\">ImageWoof</td><td>ResNet-18</td><td>46.8±0.9</td><td>49.2±0.5</td><td>50.1±0.8</td></tr><tr><td>ResNet-50</td><td>47.3±0.7</td><td>50.0±0.8</td><td>50.9±0.6</td></tr><tr><td>ResNet-101</td><td>48.9±1.1</td><td>51.5±0.6</td><td>52.4±0.7</td></tr></table>",
+        "bbox": [
+            91,
+            88,
+            480,
+            207
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6.  $\\mathcal{T}_{\\mathcal{SP}}$  Model Ablation and Cross-Architecture Testing: We use ResNet-18, ResNet-50, and ResNet-101 as  $\\mathcal{T}_{\\mathcal{SP}}$  models to generate synthetic images. During testing, the same three models are used. IPC is set to 10.",
+        "bbox": [
+            89,
+            217,
+            482,
+            272
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/710da0b5dfd0a15b8f27cf0cd423dfc27eb92886c2f2843ff03c914356ef0257.jpg",
+        "image_caption": [
+            "(a) Accuracy vs. Guidance strength"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            102,
+            297,
+            272,
+            428
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/28c224eb7f4d1179d114e2528c56f2684e8881d26b53783ef767363b2beb0521.jpg",
+        "image_caption": [
+            "(b) Accuracy vs. Sample Step"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            290,
+            297,
+            460,
+            426
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/56268c4bf81de374ccba9e17d01f26bc4e2c049a1a2d33703cbd143c4e5f24bf.jpg",
+        "image_caption": [
+            "Figure 4. Hyper-parameter analysis on (a) guidance strength  $\\lambda_{s}$  and (b) sampling step  $t$ . The star symbol indicates the parameters chosen for trade-off.",
+            "(a) Training Time overhead",
+            "Figure 5. Comparison of distillation cost on ImageNet-1k using our UniDD with comparison methods (MTT [3],  $\\mathrm{SRe}^2\\mathrm{L}$  [46], and  $\\mathrm{D^4M}$  [41]), tested on V100 GPUs."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            99,
+            518,
+            277,
+            652
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/3980bfe19f23ab16a4f8b59e65acfff76214a08a92e6d61e9526d8bf476c7cb1.jpg",
+        "image_caption": [
+            "(b) Training GPU overhead"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            287,
+            518,
+            464,
+            652
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "ments are crucial for validating the generalizability of synthetic data. We compare different combinations of  $\\mathcal{T}_{\\mathcal{SP}}$  models and testing models. In this study, we set up three  $\\mathcal{T}_{\\mathcal{SP}}$  models: ResNet-18, ResNet-50, and ResNet-101. These models are also used during testing to evaluate the cross-architecture performance of the synthetic datasets. As shown in Table. 6, using more complex  $\\mathcal{T}_{\\mathcal{SP}}$  models improves the guidance effectiveness. Moreover, when testing across different architectures, the performance does not degrade as seen in previous optimization-based meth",
+        "bbox": [
+            89,
+            750,
+            482,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "ods [3, 49, 50], demonstrating the generalizability of our synthetic data.",
+        "bbox": [
+            511,
+            90,
+            903,
+            119
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Guidance Strength Analysis. As shown in Fig. 4 (a), we analyze the performance variations of ImageWoof and ImageNette at IPC-10 under different guidance strength settings  $\\lambda_{s}$ , with Conv5 as the test model. The best results are achieved at  $\\lambda_{s} = 100$ , but further increasing the guidance strength interferes with the quality of diffusion-generated images, leading to a decline in test performance.",
+        "bbox": [
+            511,
+            121,
+            903,
+            226
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Sample Step Analysis. As shown in Fig. 4 (b), we analyze the impact of different SD sample steps on the performance of synthetic datasets under the IPC-10 setting, with Image-Woof and ImageNette resolutions set to  $256 \\times 256$ . Performance improves significantly with sample steps below 250, but gains are minimal beyond this point while time consumption increases. Thus, we select 250 steps as the optimal trade-off.",
+        "bbox": [
+            511,
+            227,
+            903,
+            345
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Distillation Cost Analysis. As shown in Fig. 5, we evaluate the time and GPU cost of various DD methods. Previous methods primarily focus on the time required to synthesize images, often neglecting the time needed to train or deploy the DD model. For MTT [3], the deployment time corresponds to the expert trajectory training time, while for  $\\mathrm{D^4M}$  [41], it is the time required to finetune the SD model. The slight increase in UniDD's distillation time and GPU cost is mainly due to the added supervision process of the  $\\mathcal{T}_{\\mathcal{SP}}$  model and the refinement stage, where approximately  $5\\% -10\\%$  of low-quality generated data is discarded. Combining the deployment and distillation costs for comparison, UniDD demonstrates the best advantage, indicating its superior suitability for practical deployment.",
+        "bbox": [
+            511,
+            348,
+            906,
+            560
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            511,
+            573,
+            633,
+            588
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this study, we introduce the first Universal Dataset Distillation (UniDD) framework, capable of handling classification, detection, and segmentation tasks within a unified approach. UniDD follows a two-stage process: in the first stage, Task-Specific Proxy Training and Task-Aligned Contextual Prompting extract critical information from real datasets. In the second stage, Universal Taskdriven Diffusion synthesizes images with specific categories, bounding boxes, and segmentation masks tailored to each task. We conduct extensive experiments to validate UniDD's effectiveness across all three tasks. For classification, UniDD achieves state-of-the-art results on ImageNet1k and ImageNet-Subset. In detection and segmentation, we establish the first dataset distillation benchmark on Pascal VOC and MS COCO, comparing UniDD to coreset selection methods. Encouraging results demonstrate that our UniDD consistently outperforms these methods, especially under extremely low compression ratios.",
+        "bbox": [
+            511,
+            598,
+            906,
+            869
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments. This work was supported by National Natural Science Fund of China (62473286).",
+        "bbox": [
+            511,
+            869,
+            903,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "10564",
+        "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] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. 3, 4",
+            "[2] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 6",
+            "[3] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4750-4759, 2022. 1, 2, 5, 6, 8",
+            "[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6",
+            "[5] Yutian Chen, Max Welling, and Alex Smola. Super-samples from kernel herding. arXiv preprint arXiv:1203.3472, 2012. 6",
+            "[6] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.6",
+            "[7] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation toImagenet-1k with constant memory. In International Conference on Machine Learning, pages 6565–6590. PMLR, 2023. 1, 5",
+            "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1, 2, 3, 5",
+            "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3, 4",
+            "[10] 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. 3",
+            "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 1, 2, 3, 5",
+            "[12] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 1, 2, 3, 5",
+            "[13] Junyao Gao, Xinyang Jiang, Huishuai Zhang, Yifan Yang, Shuguang Dou, Dongsheng Li, Duoqian Miao, Cheng Deng,"
+        ],
+        "bbox": [
+            93,
+            114,
+            483,
+            902
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "and Cairong Zhao. Similarity distribution based membership inference attack on person re-identification. In Proceedings of the AAAI conference on artificial intelligence, pages 14820-14828, 2023. 1",
+            "[14] Junyao Gao, Xinyang Jiang, Shuguang Dou, Dongsheng Li, Duoqian Miao, and Cairong Zhao. Re-id-leak: Membership inference attacks against person re-identification. International Journal of Computer Vision, 132(10):4673-4687, 2024. 1",
+            "[15] Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. Styleshot: A snapshot on any style. arXiv preprint arXiv:2407.01414, 2024. 2",
+            "[16] Junyao Gao, Yanan Sun, Fei Shen, Xin Jiang, Zhening Xing, Kai Chen, and Cairong Zhao. Faceshot: Bring any character into life. arXiv preprint arXiv:2503.00740, 2025. 2",
+            "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2",
+            "[18] Jianyang Gu, Saeed Vahidian, Vyacheslav Kungurtsev, Haonan Wang, Wei Jiang, Yang You, and Yiran Chen. Efficient dataset distillation via minimax diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15793-15803, 2024. 5, 6",
+            "[19] Ziyao Guo, Kai Wang, George Cazenavette, Hui Li, Kaipeng Zhang, and Yang You. Towards lossless dataset distillation via difficulty-aligned trajectory matching. arXiv preprint arXiv:2310.05773, 2023. 2",
+            "[20] 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. 1",
+            "[21] 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. 4, 5",
+            "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1",
+            "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3",
+            "[24] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 5",
+            "[25] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 2",
+            "[26] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangwoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In International Conference on Machine Learning, pages 11102-11118. PMLR, 2022. 1, 2"
+        ],
+        "bbox": [
+            516,
+            92,
+            906,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "10565",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[27] Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd: dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060-5069, 2019. 2",
+            "[28] 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. 1, 2, 5",
+            "[29] Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xin chao Wang. Dataset distillation via factorization. Advances in Neural Information Processing Systems, 35:1100-1113, 2022. 6",
+            "[30] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 3",
+            "[31] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. arXiv preprint arXiv:2302.14416, 2023. 2",
+            "[32] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. In International Conference on Learning Representations, 2020. 2",
+            "[33] OpenAI. Gpt-4 technical report. https://openai.com/research/gpt-4, 2023. Accessed: 2024-11-13. 2, 4",
+            "[34] Ding Qi, Jian Li, Jinlong Peng, Bo Zhao, Shuguang Dou, Jialin Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cairong Zhao. Fetch and forge: Efficient dataset condensation for object detection. Advances in Neural Information Processing Systems, 37:119283-119300, 2024. 3",
+            "[35] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 6",
+            "[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 4, 5, 6",
+            "[37] 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. 2, 5",
+            "[38] Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstantinos N Plataniotis. Datadam: Efficient dataset distillation with attention matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17097-17107, 2023. 6",
+            "[39] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489, 2017. 6"
+        ],
+        "bbox": [
+            91,
+            92,
+            485,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3",
+            "[41] Duo Su, Junjie Hou, Weizhi Gao, Yingjie Tian, and Bowen Tang. D^4: Dataset distillation via disentangled diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5809-5818, 2024. 5, 6, 8",
+            "[42] Peng Sun, Bei Shi, Daiwei Yu, and Tao Lin. On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9390-9399, 2024. 1, 3, 5",
+            "[43] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12196-12205, 2022. 2",
+            "[44] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2, 6",
+            "[45] Yichao Yan, Zanwei Zhou, Zi Wang, Jingnan Gao, and Xiaokang Yang. Dialoguenerf: Towards realistic avatar face-to-face conversation video generation. Visual Intelligence, 2 (1):24, 2024. 1",
+            "[46] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. arXiv preprint arXiv:2306.13092, 2023. 1, 3, 5, 6, 8",
+            "[47] 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. 2",
+            "[48] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In International Conference on Machine Learning, pages 12674-12685. PMLR, 2021. 1",
+            "[49] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6514-6523, 2023. 1, 2, 6, 8",
+            "[50] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. In International Conference on Learning Representations, 2020. 1, 2, 6, 8",
+            "[51] Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved distribution matching for dataset condensation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7856-7865, 2023. 1, 2"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            781
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "10566",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/cd8b4c9e-1814-423f-bdad-2b631dc6a75a_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84e0686f3b43bafc7fb6a1a9330f093437069b096dd172aebee39099a1f7099d
+size 4585278

2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/full.md

@@ -0,0 +1,338 @@
+# Towards Universal Dataset Distillation via Task-Driven Diffusion
+
+Ding Qi $^{1*}$  Jian Li $^{2\dagger}$  Junyao Gao $^{1}$  Shuguang Dou $^{1}$  Ying Tai $^{5}$   
+Jianlong Hu $^{2}$  Bo Zhao $^{4}$  Yabiao Wang $^{3,2\dagger}$  Chengjie Wang $^{4,2}$  Cairong Zhao $^{1\dagger}$ $^{1}$ Tongji University  $^{2}$ Tencent YouTu Lab  $^{3}$ Zhejiang University  
+ $^{4}$ Shanghai Jiao Tong University  $^{5}$ Nanjing University
+
+![](images/273080066be2f11e90e6d1d4791effb0af3c1dd2d9a9a35dc7bff1253d4caa6d.jpg)  
+(a) One Universal Dataset Distillation Framework for Mutli-Tasks
+
+![](images/29377090049304128148076e6e350a1f9d66a781866c0624fbfc3fec438bfb75.jpg)  
+(b) Improved Performance by UniDD  
+Figure 1. (a) We introduce the Universal Dataset Distillation Framework (UniDD), a novel dataset distillation framework supporting a range of vision tasks, including classification, object detection, and segmentation. UniDD incorporates diverse label types, such as class labels, bounding box coordinates, and pixel-level masks, facilitating the synthesis of task-specific datasets. (b) UniDD improves the performance of dataset distillation across diverse tasks on ImageNet [8], Pascal VOC [11, 12], and MS COCO [28].
+
+# Abstract
+
+Dataset distillation (DD) condenses key information from large-scale datasets into smaller synthetic datasets, reducing storage and computational costs for training networks. However, most recent research has primarily focused on image classification tasks, with limited exploration in detection and segmentation. Two key challenges remain: (i) Task Optimization Heterogeneity, where existing methods focus on class-level information but fail to address the diverse needs of detection and segmentation, and (ii) Inflexible Image Generation, where current generation methods rely on global updates for single-class targets and lack localized optimization for specific object regions. To address these challenges, we propose UniDD, a universal dataset distillation framework built on a task-driven diffusion model for diverse DD tasks, as shown in Fig. 1. Our approach operates in two stages: Universal Task Knowledge Mining, which captures task-relevant information through task
+
+specific proxy model training, and Universal Task-Driven Diffusion, where these proxies guide the diffusion process to generate task-specific synthetic images. Extensive experiments across ImageNet-1K, Pascal VOC, and MS COCO demonstrate that UniDD consistently outperforms state-of-the-art methods. In particular, on ImageNet-1K with IPC-10, UniDD surpasses previous diffusion-based methods by  $6.1\%$ , while also reducing deployment costs.
+
+# 1. Introduction
+
+Dataset distillation (DD) [44], as an innovative data optimization technique in machine learning [13, 14, 20, 22, 45], condenses essential information from large datasets into smaller proxy sets. Training on these highly ensemble proxy sets enables the model to use much fewer resources while providing comparable performance. Prior DD methods excelled in image classification using various techniques like Gradient Matching [26, 48, 50], Distribution Matching [49, 51], Training Trajectory Matching [3, 7], and Decoupled Methods [42, 46].
+
+Despite advancements in DD, research on key visual tasks—detection and segmentation—remains limited. We identify the following key challenges: (i) Task Optimization Heterogeneity. Existing DD methods are primarily focused on image classification, optimizing distillation by compressing class-level information, but they fall short in addressing the diverse needs of tasks like detection [27] and segmentation, such as object localization and pixel-level information. (ii) Inflexible Image Generation. Current image generation methods rely on global updates for single-class targets and lack localized optimization for specific object regions. These limitations become especially evident in images with multiple objects or complex structures, hindering the ability of existing DD to generate distilled data that meets the requirements of diverse tasks. Consequently, the applicability of DD in the visual domain remains constrained.
+
+To address these limitations, we propose a Universal Dataset Distillation (UniDD) framework that optimizes for multi-task learning, including classification, detection, and segmentation, overcoming the shortcomings in task adaptation and image generation. Our approach consists of two stages: Universal Task Knowledge Mining and Universal Task-Driven Diffusion. In the first stage, we introduce Task-Specific Proxy  $(\mathcal{T}_{\mathcal{SP}})$  models, trained on real datasets to extract task-specific information. This allows the  $\mathcal{T}_{\mathcal{SP}}$  models to capture and store category information for classification, bounding box positions for detection, and mask data for segmentation within their parameters. Additionally, previous research has often overlooked the importance of contextual information, particularly in high-resolution datasets like ImageNet [8], Pascal VOC [11, 12] and MS COCO [28], where rich object-context relationships exist. To address this, we introduce Task-Aligned Contextual Prompting, which uses a Vision-Language model [33] to generate contextual prompts for real images. In the second stage, we introduce a Task-Driven Diffusion for image synthesis. Specifically, the  $\mathcal{T}_{\mathcal{SP}}$  models trained in the first stage supervise the denoising process of the diffusion model by calculating the task loss on the predicted clean images. This guides the diffusion model in generating images that match the given labels. Notably, since the  $\mathcal{T}_{\mathcal{SP}}$  models already align with the target domain, there is no need to retrain the diffusion model on the real set, significantly reducing the cost in practical applications.
+
+In summary, our contributions are threefold:
+
+- We identify key challenges in extending dataset distillation beyond classification and propose the first universal framework for DD across classification, detection, and segmentation tasks.  
+- We develop a two-stage approach, Universal Task Knowledge Mining and Universal Task-Driven Diffusion for Data Synthesis, which flexibly synthesizes images for
+
+multi-tasks. Additionally, we address the previously overlooked need to align potential contextual information within datasets.
+
+- Our method achieves state-of-the-art performance in classification tasks, on the ImageNet-1K dataset, our approach surpasses the previous state-of-the-art method by over  $6.1\%$ . Additionally, we establish new benchmarks for detection and segmentation on VOC and COCO, showing significant improvements over coreset methods.
+
+# 2. Related works
+
+# 2.1. Dataset Distillation
+
+Dataset distillation [44], a dataset reduction method, addresses the challenge of handling substantial data by synthesizing a smaller, representative dataset. Existing dataset distillation methods can be categorized into meta-learning and data matching frameworks based on whether they explicitly mimic the performance of target data. In the meta-learning framework, the distilled data are treated as hyperparameters and optimized in a nested loop fashion according to the trained model's risk. For example, KIP [32] performs kernel ridge regression using the Neural Tangent Kernel (NTK) [25]. The data matching framework includes gradient matching, distribution matching, and trajectory matching. These methods update distilled data by imitating the influence of target data. For example, DC [50] and DM [49] enforce synthetic and real images to have similar gradients or distribution at every training step. MTT [3] proposes to imitate long-range trajectories of optimization steps for real images, thereby better mimicking the training dynamics of real images for synthetic images. However, the aforementioned gradient matching methods require a large number of iterations. To address this, DREAM [31] and IDC [26] focus on more efficient dataset distillation training. Due to the unsatisfactory performance of DM, CAFE [43] and IDM [51] propose improved distribution matching between the real and synthetic data. For the trajectory matching method, DATM [19] scales trajectory matching-based methods to larger synthetic datasets, achieving lossless dataset distillation.
+
+Despite significant progress in previous work, there have been few breakthroughs in dataset distillation for detection and segmentation tasks. In this work, we introduce the first universal framework that simultaneously handles classification, object detection, and semantic segmentation, significantly expanding the generality and scope of dataset distillation research.
+
+# 2.2. Diffusion Models
+
+Diffusion models [15, 16, 37, 47] are generative models that show superior fidelity and diversity compared to GANs [17]. It aims to learn the mapping from a Gaussian
+
+![](images/8c94ce46423189613163e117cc42deebceb0535c1c09b9fdc51f7a33944a2e6a.jpg)  
+(a) Synthetic images for classification.
+
+![](images/6a727fa95fd69b053327cd236578d7e039dc29ffce88b311fa0bedf52568970d.jpg)  
+(b) Synthetic images for detection.
+
+![](images/b749891040b76d77d0cb4ecba985295e6e5ca13b40b01681cb87fd578d07b433.jpg)  
+(c) Synthetic images for segmentation.  
+Figure 2. Visualization of synthetic images generated by UniDD. (a) For classification, images are synthesized on the ImageWoof [8] dataset. (b) For detection, images are synthesized on the Pascal VOC dataset with specified bounding box coordinates. (c) For segmentation, images are synthesized on the Pascal VOC [11, 12] dataset with pixel-level masks. Our UniDD is the first approach to enable high-realism data distillation across classification, detection, and segmentation tasks.
+
+distribution to the real dataset distribution by adding noise in the diffusion process and reversing the noise to the raw image in the denoising process. In the forward process, diffusion models progressively add noise to a clean sample  $z_0$ , resulting in a noisy sample  $z_t$ :
+
+$$
+z _ {t} = \sqrt {\alpha_ {t}} z _ {0} + \sqrt {1 - \alpha_ {t}} \epsilon , \quad \epsilon \sim \mathcal {N} (0, I). \tag {1}
+$$
+
+In the reverse process, a denoising network  $\epsilon_{\theta}$  estimates the noise and iteratively recovers the clean sample  $z_0$ :
+
+$$
+\epsilon_ {\theta} (z _ {t}, t) \approx \frac {z _ {t} - \sqrt {\alpha_ {t}} z _ {0}}{\sqrt {1 - \alpha_ {t}}}. \tag {2}
+$$
+
+DDIM [40] provides the following formula for predicting a clean data sample:
+
+$$
+\hat {z} _ {0} = \frac {z _ {t} - \sqrt {1 - \alpha_ {t}} \epsilon_ {\theta} (z _ {t} , t)}{\sqrt {\alpha_ {t}}}. \tag {3}
+$$
+
+This formula predicts the clean data point  $\hat{z}_0$  from the noisy sample  $z_{t}$ , enabling deterministic sampling for efficient and high-quality image synthesis.
+
+To introduce additional conditions into the diffusion model, classifier guidance [9] has been proposed to train an additional classifier on the noise to guide the model to generate images that are closer to the specified condition. Additionally, [23] introduces classifier-free guidance that trains both an unconditioned and conditioned model simultaneously, achieving more efficient guidance. Following these, recent methods [1, 10, 30] have been proposed to generate high-quality images from diverse conditions.
+
+# 3. Universal Dataset Distillation Framework
+
+# 3.1. Problem Definition
+
+Universal Dataset Distillation. Current dataset distillation research primarily targets classification tasks and lacks
+
+a generalizable approach for detection and segmentation. To address this, we propose a universal dataset distillation standard that constructs a compact synthetic dataset  $\mathcal{S} = (\hat{x}_1, \hat{y}_1), \dots, (\hat{x}_{|\mathcal{S}|}, \hat{y}_{|\mathcal{S}|})$  from a larger real dataset  $\mathcal{T} = (x_1, y_1), \dots, (x_{|\mathcal{T}|}, y_{|\mathcal{T}|})$ , with  $|S| \ll |\mathcal{T}|$ . The primary difference between tasks lies in label structure: classification uses single-class labels, detection includes bounding boxes with class information, and segmentation requires pixel-level masks. Our goal is for models trained on the distilled dataset  $\mathcal{S}$  to perform comparably to those trained on the full dataset  $\mathcal{T}$  across classification, detection, and segmentation tasks, as formalized by:
+
+$$
+\mathbb {E} _ {(x, y) \sim P _ {D}} \left| \ell \left(\theta_ {\mathcal {T}} ^ {t a s k} (x), y\right) - \ell \left(\theta_ {\mathcal {S}} ^ {t a s k} (x), y\right) \right| \leq \epsilon , \tag {4}
+$$
+
+where  $P_{D}$  denotes the distribution of test data,  $\ell$  represents the performance metric specific to each task,  $task\in$  {class,detect,segment}, and  $\epsilon$  is a small tolerance value. For each task, the parameter set  $\theta_{\mathcal{T}}^{task}$  on the real dataset  $\mathcal{T}$  is optimized as follows:
+
+$$
+\theta_ {\mathcal {T}} ^ {t a s k} = \arg \min _ {\theta} \mathbb {E} _ {(x, y) \in \mathcal {T}} \left[ \ell \left(\phi_ {\theta} ^ {t a s k} (x), y\right) \right], \qquad (5)
+$$
+
+Similarly,  $\theta_S^{task}$  is optimized using the synthetic dataset  $S$ .
+
+# 3.2. Universal Task Knowledge Mining
+
+Task-Specific Proxy Training. Building a universal DD framework is challenging due to the diverse optimization objectives in classification, detection, and segmentation tasks. Recent decoupled DD studies have reached a consensus [34, 42, 46]: a fully trained network can capture and retain essential dataset information. Inspired by this, we extend to general tasks, training a classifier, detector, and segmenter on task-specific datasets to extract critical information—class labels for classification, bounding boxes for
+
+![](images/697bfc9651c9aab0d9e97c88dab77dad9859a6862ec3e7f7180dc844b1d8405e.jpg)  
+Figure 3. Universal Dataset Distillation Framework. This framework consists of two stages: in the Universal Task Knowledge Mining stage, we introduce an easily deployable task-specific proxy  $(\mathcal{T}_{\mathcal{SP}})$  model trained on the target dataset and a vision-language model to extract task-aligned context. In the Universal Task-Driven Diffusion stage, we apply the gradient of the supervision loss, derived from the  $\mathcal{T}_{\mathcal{SP}}$  model's prediction of the clean image  $\hat{z}_0$ , at each sampling step.
+
+detection, and pixel-level masks for segmentation. We call this Task-Specific Proxy  $(\mathcal{T}_{\mathcal{SP}})$  Model, defined as:
+
+$$
+\theta = \arg \min  _ {\theta} \mathcal {L} _ {\text {t a s k}} (\mathcal {F} (x), y), \tag {6}
+$$
+
+where  $\mathcal{F}$  represents the  $\mathcal{T}_{\mathcal{SP}}$  model trained on the target data,  $\theta$  denotes the model's parameters,  $\mathcal{L}_{task}$  represents the loss function specific to the task.
+
+The  $\mathcal{T}_{\mathcal{SP}}$  model choice is flexible, allowing adaptation for optimal performance on target datasets. For example, ResNet-18 [21] can serve as a lightweight, fast-deploy option for simple classification, while Faster R-CNN [36] may be used for complex detection tasks to capture key detection information. Unlike previous diffusion-based dataset distillation methods, our approach relies on task-driven learning to align synthetic and real datasets, avoiding costly diffusion model fine-tuning on the target dataset and significantly reducing deployment overhead.
+
+Task-Aligned Contextual Prompting. Previous diffusion-based dataset distillation has largely focused on classification tasks and overlooked the role of natural language in guiding image generation. For diverse tasks, textual prompts must go beyond describing the primary objects in images. As inputs to Stable Diffusion models, prompts can also guide the generation process. To address this, we use a vision-language model to extract natural language descriptions and design task-specific questions to generate task-aligned prompts, a method we call Task-Aligned Contextual Prompting. Following GPT-4o [33], we create tailored prompts for different tasks. For classification, prompts de
+
+scribe object classes, while for detection and segmentation, they also include object relationships and the interaction between objects and the background, providing richer contextual guidance.
+
+# 3.3. Universal Task-Driven Diffusion.
+
+Inspired by recent advancements in guidance diffusion techniques [1, 9], we propose a Task-Driven Diffusion approach for Dataset Distillation. As illustrated in Fig. 3, during each sampling step  $S(z_{t}, t)$  of the diffusion process, the trained  $\mathcal{T}_{\mathcal{SP}}$  model  $\mathcal{F}$  provides task-specific guidance. Given an label  $\tilde{a}$ , the original noise prediction  $\epsilon_{\theta}(z_{t}, t, \tilde{c})$  is adjusted using the output of  $\mathcal{F}$ . The adjusted noise prediction  $\hat{\epsilon}$  is formulated as follows:
+
+$$
+\hat {\epsilon} _ {\theta} \left(z _ {t}, t, \tilde {c}\right) = \epsilon_ {\theta} \left(z _ {t}, t, \tilde {c}\right) + \sqrt {1 - \alpha_ {t}} \nabla_ {z _ {t}} \mathcal {L} _ {\text {t a s k}} \left(\tilde {a}, \mathcal {F} \left(z _ {t}\right)\right), \tag {7}
+$$
+
+where  $\mathcal{L}_{\mathrm{task}}$  represents the task-specific loss function and  $\tilde{a} = \{a_{cl}, a_{bbox}, a_{mask}\}$  may denote a class label  $a_{cl}$ , a bounding box  $a_{bbox}$ , or a pixel-level segmentation mask  $a_{mask}$ . This adjustment mechanism effectively guides the diffusion process to produce data closely aligned with the intended distribution for each task.  $\tilde{c}$  is a textual prompt.
+
+Because  $\mathcal{F}$  is trained on clean images, it may not perform well directly on noisy inputs. Therefore, we first use Equation 3 to estimate the clean image  $\hat{z}_0$ . Once  $\hat{z}_0$  is obtained, we apply  $\mathcal{F}$  to compute the task-driven loss as follows:
+
+$$
+\hat {\epsilon} _ {\theta} \left(z _ {t}, t, \tilde {c}\right)) = \epsilon_ {\theta} \left(z _ {t}, t, \tilde {c}\right) + s (t) \cdot \nabla_ {z _ {t}} \mathcal {L} _ {\text {t a s k}} (\tilde {a}, \mathcal {F} (\hat {z} _ {0})), \tag {8}
+$$
+
+where  $s(t)$  is the task-driven strength at each sampling step,
+
+defined as:
+
+$$
+s (t) = \lambda_ {s} \cdot \sqrt {1 - \alpha_ {t}}, \tag {9}
+$$
+
+with  $\lambda_{s}$  being a constant coefficient controlling the overall intensity of the task-driven. The term  $\nabla_{z_t}\mathcal{L}_{task}(\tilde{a},\mathcal{F}(\hat{z}_0))$  represents the gradient of the loss function computed based on the  $\mathcal{T}_{\mathcal{SP}}$  model's prediction of the clean image  $\hat{z}_0$ . Specifically, this gradient is computed as:
+
+$$
+\nabla_ {z _ {t}} \mathcal {L} _ {\text {t a s k}} (\tilde {a}, \mathcal {F} (\hat {z} _ {0})) = \nabla_ {z _ {t}} \mathcal {L} _ {\text {t a s k}} \left(\tilde {a}, \mathcal {F} \left(\frac {z _ {t} - \sqrt {1 - \alpha_ {t}} \epsilon_ {\theta}}{\sqrt {\alpha_ {t}}}\right)\right). \tag {10}
+$$
+
+Unlike previous approaches [18, 41] that required finetuning the diffusion model, such as those involving SD [37], to align with new data domains, our method leverages a  $\mathcal{T}_{\mathcal{SP}}$  model trained specifically on the target dataset. This model encapsulates critical information about the target task, effectively driving the diffusion process to generate data that is well-aligned with the target task without the need for additional finetuning. This not only conserves time and resources but also mitigates the risk of overfitting the diffusion model to the specific task.
+
+Proxy-Driven High-Realism Refinement. To ensure the high realism and accuracy of the generated images, we employ a trained  $\mathcal{T}_{\mathcal{SP}}$  model, to score each image. By setting a well-defined threshold, we filter out low-quality images based on their scores. The remaining images are then relabeled using the same  $\mathcal{T}_{\mathcal{SP}}$  model. This combined process of filtering and relabeling can be expressed as:
+
+$$
+\{x, y \} = \left\{x _ {i}, \mathcal {F} _ {\text {l a b e l}} \left(x _ {i}\right) \mid \operatorname {s c o r e} \left(\mathcal {F} \left(x _ {i}\right)\right) \geq \tau \right\} \tag {11}
+$$
+
+where  $\mathcal{F}(x_i)$  represents the model's output score for image  $x_{i}$ ,  $\tau$  is the quality threshold, and  $\mathcal{F}_{\mathrm{label}}$  generates the accurate label  $y_{i}$  for the filtered image. This ensures that both the generated images and their labels are of high quality and closely match the desired data distribution.
+
+# 4. Experiments
+
+# 4.1. Experimental Setups
+
+Datasets and Evaluation Metric. To enhance practicality, we prioritize high-resolution and large-scale datasets. (1) Classification task: our experiments are primarily conducted on the full ImageNet 1K dataset [8], with IPC settings commonly set to 10 and 50, following the configurations in [41]. Additionally, based on the dataset partitioning settings from previous work [3], we also use the ImageNet 10-class subsets, such as ImageWoof, ImageFruit, ImageNette, and ImageMeow. These 10-class subsets contain very similar categories, making them challenging for dataset distillation. Top-1 accuracy  $(\%)$  is reported in the experimental tables. (2) Object detection task: we conduct experiments on the widely-used Pascal VOC [11, 12] and MS COCO [28] datasets, using mAP (Mean Average
+
+# Algorithm 1 Universal Dataset Distillation Approach
+
+Input: Real data  $\mathcal{T}$ , Diffusion model  $\epsilon_{\theta}$ ,  $\mathcal{T}_{\mathcal{SP}}$  model  $\mathcal{F}$ , Sampling steps  $T$ , Task-Driven strength  $s(t)$
+
+Output: Synthetic dataset  $S$ .
+
+1: Step I: Universal Task Knowledge Mining  
+2: Train  $\mathcal{T}_{\mathcal{SP}}$  model  $\mathcal{F}$  on  $\mathcal{T}$  using Equ. 6  
+3: Use GPT-4o to generate task-aligned context for classification, detection, and segmentation datasets
+
+4: Step II: Universal Task-Driven Diffusion for Data Synthetic
+
+5: for each sampling step  $t = 1$  to  $T$  do  
+6: Generate noisy sample  $z_{t}$ .  
+7: Predict clean image  $\hat{z}_0$  using Equ. 3  
+8: Calculate guidance loss  $\nabla_{z_t}\mathcal{L}_{task}(c,\mathcal{F}(\hat{z}_0))$  
+9: Adjust noise prediction using Equ. 8  
+10: Update  $z_{t-1}$  using adjusted noise prediction.  
+11: end for  
+12: Score each generated image  $x_{i}$  using  $\mathcal{F}(x_i)$ .  
+13: Filter and relabel synthetic set based on threshold  $\tau$  using Equ. 11.
+
+Precision) and  $AP_{50}$  (Average Precision at IoU threshold of 0.50) as evaluation metric. (3) Semantic segmentation task: we utilize the Pascal VOC dataset, with mIoU (Mean Intersection over Union) as evaluation metrics.
+
+Implementation Details. Unless otherwise specified, the Task-Specific Proxy  $\mathcal{T}_{\mathcal{SP}}$  model used for the classification task is ResNet-18 [21], the detector is Faster R-CNN [36], and the segmentation model is LR-ASPP [24], all utilizing the official pre-trained weights from PyTorch. For the diffusion model, we use SD v1.5 [37] as the baseline. For quantitative comparisons, we generate  $256 \times 256$  images for classification tasks and  $512 \times 512$  images for detection and segmentation tasks. Synthetic images are created using 250 denoising steps based on a predefined number of random noise samples, aligned with the IPC settings. The threshold  $\tau$  for Proxy-Driven High-Realism Refinement is set to 0.9 for classification, 0.85 for detection, and 0.75 for segmentation. Generally, this operation results in a data discard rate of approximately  $5 - 10\%$ . All experiments are conducted on NVIDIA V100 GPUs, with further details provided in the special experiments section. We generate the distilled dataset 3 times, train the network from scratch 5 times per dataset (with pre-trained backbone weights for the detector and segmenter), and evaluate on the original test set, reporting results as  $\bar{x} \pm std$ .
+
+# 4.2. Results on Classification
+
+We perform comprehensive comparisons across datasets of various scales. For large-scale datasets ImageNet-1k, we compare TESLA [7], SRe $^2$ L [46], D $^4$ M [41], RDED [42], and MiniMax Diffusion [18] (abbreviated as MiMxDiff in
+
+<table><tr><td>Dataset</td><td>IPC</td><td>Method</td><td>Accuracy(%)</td></tr><tr><td rowspan="12">ImageNet-1K</td><td>1000</td><td>Full dataset</td><td>69.8</td></tr><tr><td rowspan="6">10</td><td>TESLA</td><td>7.7</td></tr><tr><td>SRe2L</td><td>21.3</td></tr><tr><td>D4M</td><td>27.9</td></tr><tr><td>RDED</td><td>42.0</td></tr><tr><td>MiMxDiff</td><td>44.3</td></tr><tr><td>UniDD (Ours)</td><td>50.4 (↑6.1)</td></tr><tr><td rowspan="5">50</td><td>SRe2L</td><td>46.8</td></tr><tr><td>D4M</td><td>55.2</td></tr><tr><td>RDED</td><td>56.5</td></tr><tr><td>MiMxDiff</td><td>58.6</td></tr><tr><td>UniDD (Ours)</td><td>62.8 (↑4.2)</td></tr></table>
+
+Table 1. Performance comparison with state-of-the-art methods on large-scale dataset ImageNet-1K. All methods employ ResNet18 as test model. All standard deviations in this table are less than one. The results of Full dataset come from the official PyTorch. The best results are marked in bold, while the second-best results are underlined.
+
+the table). For small-scale dataset ImageNet-Subset, we compare Random selection, DM [49], MTT [3], HaBa [29], DataDAM [38], and MiniMax Diffusion.
+
+Quantitative comparison on ImageNet-1K and ImageNet-Subset. As shown in the Tab 1, we utilize ResNet-18 as the  $\mathcal{T}_{\mathcal{SP}}$  model, trained on ImageNet-1K. We report results for IPC-10 and IPC-50, evaluating all outcomes with the ResNet-18 model. To ensure a fair comparison, our training and testing pipeline is aligned with that of RDED and  $\mathrm{D^4M}$ . Our method, UniDD, achieves state-of-the-art performance on this large-scale dataset. Specifically, at IPC-10, UniDD surpasses optimization-based methods, delivering substantial gains over TESLA and  $\mathrm{SRe}^2\mathrm{L}$ . Additionally, it demonstrates significant advantages over diffusion-based methods, outperforming  $\mathrm{D^4M}$  by  $22.5\%$  and MiniMax Diffusion by  $6.1\%$ . These results underscore UniDD's effectiveness by using a lightweight  $\mathcal{T}_{\mathcal{SP}}$  model to guide SD denoising, aligning generative and target domains without requiring diffusion model finetuning, thus advancing performance in dataset distillation.
+
+As shown in the Tab 2, we conduct experiments on the ImageNet-subset under the IPC-10 setting. For consistency, the test model is standardized to a 5-layer ConvNet across all methods, except for MiniMax Diffusion, which uses a 6-layer ConvNet. Our method stands out among the comparison methods, outperforming the second-best by  $8.2\%$  on ImageWoof,  $9.1\%$  on ImageFruit, and  $6.6\%$  on ImageMeow. Additionally, our method surpasses the diffusion-based MiniMax Diffusion by  $9.8\%$  on ImageWoof, further demonstrating the effectiveness of UniDD approach in ag
+
+gregating information from the original dataset.
+
+Qualitative Result. In Fig. 2 (a), we present the synthetic images generated by UniDD on ImageWoof. Benefiting from the powerful generative capability of SD and the use of a  $\mathcal{T}_{\mathcal{SP}}$  model to align the real and synthetic data domains, our synthetic images for the classification task exhibit significantly higher realism compared to optimization-based methods [3, 46, 50]. However, unlike other diffusion-based dataset distillation methods [18, 41], we do not finetune SD, greatly reducing the deployment time of UniDD.
+
+# 4.3. Results on Detection and Segmentation
+
+We deploy our UniDD framework for detection and segmentation tasks and conduct experiments on the Pascal VOC and MS COCO datasets. Since dataset distillation is rarely applied to these tasks, we reference earlier dataset distillation works (DD [44], DC [50]) and compare UniDD with core-set selection methods. Specifically, we construct the following baseline methods: Random [35], K-center [39], and Herding [2, 5]. Additionally, we develop a method that selects a subset with balanced class representation, which we refer to as Uniform.
+
+Quantitative Result. In Table 3, we compare coreset selection methods with UniDD across different compression ratios on the object detection task. According to the official MMDetection implementation [4], the detector Faster R-CNN [36] achieves  $51.4\%$  mAP and  $80.3\%$ $\mathrm{AP}_{50}$  on Pascal VOC with the full dataset, and  $32.6\%$  mAP and  $51.4\%$ $\mathrm{AP}_{50}$  on MS COCO, providing an approximate performance upper bound. On Pascal VOC, UniDD demonstrates significant performance gains, achieving  $7.7\%$  mAP and  $19.2\%$ $\mathrm{AP}_{50}$  higher than the Random method at an extremely low compression ratio of  $0.5\%$ . Additionally, UniDD surpasses Uniform by  $11.1\%$  mAP and  $21.2\%$ $\mathrm{AP}_{50}$  at a  $1\%$  compression ratio. On the more challenging MS COCO dataset, UniDD also shows a clear advantage, surpassing the best coreset selection method by  $3.7\%$ ,  $3.4\%$ , and  $3.4\%$  mAP at compression ratios of  $0.25\%$ ,  $0.5\%$ , and  $1\%$ , respectively. This indicates UniDD's robustness, achieving strong results even on more complex datasets, which can be attributed to the Task Knowledge Mining in the first stage that extracts sufficient task-relevant information.
+
+In Table 4, we present experimental results on the VOC dataset for the semantic segmentation task. Based on the MMSegmentation [6] implementation, LR-ASPP with a MobileNetV3 backbone achieves  $32.9\%$  mIoU on the full dataset. UniDD outperforms coreset selection methods at all compression ratios, demonstrating its effectiveness in dataset distillation for segmentation tasks. This result further underscores UniDD's versatility across different tasks. By leveraging task-driven diffusion synthesis alongside task knowledge mining, UniDD generates synthetic data well-suited to complex, pixel-level predictions, illustrating its
+
+<table><tr><td>Dataset</td><td>Random</td><td>DM</td><td>MTT</td><td>HaBa</td><td>DataDAM</td><td>MiMxDiff</td><td>UniDD (Ours)</td><td>Full</td></tr><tr><td>ImageNette</td><td>47.7±2.4</td><td>58.1±0.3</td><td>63.0 ±1.3</td><td>64.7 ±1.6</td><td>59.4±0.4</td><td>62.0±0.2</td><td>66.0±0.7</td><td>87.4 ±1.0</td></tr><tr><td>ImageWoof</td><td>27.0±1.9</td><td>31.4±0.5</td><td>35.8 ±1.8</td><td>38.6 ±1.2</td><td>34.4±0.4</td><td>37.0±1.0</td><td>46.8±0.9</td><td>67.0 ±1.3</td></tr><tr><td>ImageFruit</td><td>21.4±1.2</td><td>-</td><td>40.3 ±1.3</td><td>42.5 ±1.6</td><td>-</td><td>-</td><td>51.6±0.5</td><td>63.9 ±2.0</td></tr><tr><td>ImageMeow</td><td>29.0±1.1</td><td>-</td><td>40.4 ±2.2</td><td>42.9 ±0.9</td><td>-</td><td>-</td><td>49.5±0.7</td><td>66.7 ±1.1</td></tr></table>
+
+Table 2. Performance comparison with other state-of-the-art methods on ImageNet-Subset. IPC is set to 10. The best results are marked in bold, while the second-best results are underlined. “-” were not officially implemented in the original papers.  
+
+<table><tr><td rowspan="3">Methods</td><td colspan="4">Object Detection</td></tr><tr><td colspan="2">Pascal VOC</td><td colspan="2">MS COCO</td></tr><tr><td>mAP</td><td>AP50</td><td>mAP</td><td>AP50</td></tr><tr><td>Ratio</td><td colspan="2">0.5%</td><td colspan="2">0.25%</td></tr><tr><td>Random</td><td>0.8±0.2</td><td>3.1±0.4</td><td>0.5±0.1</td><td>1.7±0.3</td></tr><tr><td>Uniform</td><td>0.9±0.1</td><td>3.4±0.3</td><td>0.8±0.2</td><td>2.4±0.5</td></tr><tr><td>K-Center</td><td>0.5±0.1</td><td>2.1±0.3</td><td>0.4±0.1</td><td>1.5±0.2</td></tr><tr><td>Herding</td><td>0.6±0.2</td><td>2.4±0.2</td><td>0.5±0.1</td><td>1.8±0.4</td></tr><tr><td>UniDD (Ours)</td><td>8.5±0.4</td><td>22.3±0.6</td><td>4.5±0.3</td><td>10.3±0.4</td></tr><tr><td>Ratio</td><td colspan="2">1%</td><td colspan="2">0.5%</td></tr><tr><td>Random</td><td>4.2±0.5</td><td>13.7±0.6</td><td>3.7±0.2</td><td>10.1±0.3</td></tr><tr><td>Uniform</td><td>5.7±0.2</td><td>17.7±0.4</td><td>3.4±0.4</td><td>9.5±0.6</td></tr><tr><td>K-Center</td><td>3.6±0.6</td><td>12.3±0.3</td><td>3.2±0.5</td><td>9.3±0.5</td></tr><tr><td>Herding</td><td>3.5±0.5</td><td>11.9±0.5</td><td>3.5±0.3</td><td>9.7±0.3</td></tr><tr><td>UniDD (Ours)</td><td>16.8±0.5</td><td>38.9±0.7</td><td>7.1±0.4</td><td>16.9±0.3</td></tr><tr><td>Ratio</td><td colspan="2">2%</td><td colspan="2">1%</td></tr><tr><td>Random</td><td>12.4±0.4</td><td>34.3±0.5</td><td>7.2±0.8</td><td>17.3±0.9</td></tr><tr><td>Uniform</td><td>13.8±0.3</td><td>36.2±0.4</td><td>7.4±0.5</td><td>17.6±0.5</td></tr><tr><td>K-Center</td><td>10.9±0.6</td><td>29.3±0.6</td><td>6.1±0.3</td><td>15.4±0.6</td></tr><tr><td>Herding</td><td>10.4±0.4</td><td>28.7±0.7</td><td>6.7±0.4</td><td>16.3±0.7</td></tr><tr><td>UniDD (Ours)</td><td>23.9±0.5</td><td>48.5±0.6</td><td>10.8±0.4</td><td>22.5±0.5</td></tr><tr><td>Full</td><td>51.4±0.8</td><td>80.3±0.4</td><td>32.6±0.7</td><td>51.4±0.8</td></tr></table>
+
+Table 3. Performance comparison with coreset selection methods on Pascal VOC and MS COCO for the object detection task, using Faster R-CNN as the detector. "Ratio" denotes the compression ratio.  
+
+<table><tr><td rowspan="2">Methods</td><td colspan="4">Semantic Segmentation (mIoU)</td></tr><tr><td>Ratio: 3.5%</td><td>Ratio: 7%</td><td>Ratio: 14%</td><td>Full</td></tr><tr><td>Random</td><td>6.7±0.4</td><td>9.6±0.5</td><td>11.3±0.4</td><td></td></tr><tr><td>Uniform</td><td>6.5±0.6</td><td>9.3±0.4</td><td>12.1±0.5</td><td></td></tr><tr><td>K-Center</td><td>6.1±0.4</td><td>8.5±0.6</td><td>10.4±0.7</td><td>32.9±0.6</td></tr><tr><td>Herding</td><td>6.2±0.7</td><td>8.7±0.4</td><td>10.6±0.3</td><td></td></tr><tr><td>UniDD (Ours)</td><td>10.3±0.5</td><td>12.3±0.3</td><td>14.9±0.7</td><td></td></tr></table>
+
+versatility across a range of vision tasks.
+
+Qualitative Result. As shown in Fig. 2 (b) and 2 (c), we present visualizations of detection and segmentation data synthesized using the UniDD framework. These images
+
+are generated with  $\mathcal{T}_{\mathcal{SP}}$  models, Faster R-CNN for detection and LR-ASPP for segmentation, which effectively steer the diffusion process toward the specified labels, bounding boxes, and masks. Previous research in dataset distillation has predominantly focused on single-category label synthesis, where each image aligns with a single class. In contrast, our UniDD framework overcomes this limitation, enabling dataset distillation for more complex, multi-class tasks. This advancement not only broadens the applicability of dataset distillation but also offers valuable insights for future research in this domain.
+
+# 4.4. Ablation Study
+
+Component Analysis. We conduct a comprehensive ablation study on the proposed UniDD, analyzing its components: (1) Baseline (a standard SD v1.5 checkpoint), (2) Task-driven Diffusion with the  $\mathcal{T}_{\mathcal{SP}}$  model, (3) Task-Aligned Contextual Prompting (abbreviated as Context), and (4) Proxy-Driven High-Realism Refinement (abbreviated as Refine). As shown in Table 5, the proposed task-driven approach significantly improves the performance of the baseline, demonstrating that this guidance effectively bridges the gap between real and synthetic datasets. The contextual information is more effective for more complex detection and segmentation datasets, as it reveals underlying target relationships. The refinement module also proves to be effective, as the randomness of the SD model can lead to the generation of images that deviate from the annotations or result in failed generations.
+
+Table 4. Performance comparison with coreset selection methods on Pascal VOC for the semantic segmentation task, using LR-ASPP as the segmenter. The metric is mIoU, and "Ratio" denotes the compression ratio.  
+
+<table><tr><td rowspan="3">Methods</td><td colspan="4">Dataset</td></tr><tr><td colspan="2">Cls. / Accuracy (%)</td><td>Det. / mAP (%)</td><td>Seg. / mIoU (%)</td></tr><tr><td>ImageNette</td><td>ImageWoof</td><td>VOC</td><td>VOC</td></tr><tr><td>Baseline</td><td>60.4±0.9</td><td>39.1±1.1</td><td>11.5±0.2</td><td>7.2±0.3</td></tr><tr><td>Baseline+TSP</td><td>64.2±0.8</td><td>45.3±0.6</td><td>14.7±0.4</td><td>8.9±0.5</td></tr><tr><td>Baseline+TSRP+Context</td><td>64.7±0.4</td><td>46.1±0.7</td><td>15.9±0.3</td><td>9.6±0.4</td></tr><tr><td>Baseline+TSP+Context+Refine</td><td>66.0±0.7</td><td>46.8±0.9</td><td>16.8±0.5</td><td>10.3±0.5</td></tr></table>
+
+Table 5. Ablation study on UniDD. For classification datasets, IPC is set to 10 with ResNet-18 as the test model. For detection on VOC, the compression ratio is set to  $1\%$  with Faster R-CNN as the test model. For segmentation on VOC, the ratio is set to  $3.5\%$  with LR-ASPP as the test model.
+
+Cross-Architecture Analysis. Cross-Architecture experi
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">Test Model</td><td colspan="3">\( T_{SP} \)Model</td></tr><tr><td>ResNet-18</td><td>ResNet-50</td><td>ResNet-101</td></tr><tr><td rowspan="3">ImageNette</td><td>ResNet-18</td><td>66.0±0.7</td><td>68.8±0.7</td><td>69.5±0.7</td></tr><tr><td>ResNet-50</td><td>67.2±0.6</td><td>69.5±1.1</td><td>70.7±0.9</td></tr><tr><td>ResNet-101</td><td>69.3±0.9</td><td>70.9±1.2</td><td>72.1±0.6</td></tr><tr><td rowspan="3">ImageWoof</td><td>ResNet-18</td><td>46.8±0.9</td><td>49.2±0.5</td><td>50.1±0.8</td></tr><tr><td>ResNet-50</td><td>47.3±0.7</td><td>50.0±0.8</td><td>50.9±0.6</td></tr><tr><td>ResNet-101</td><td>48.9±1.1</td><td>51.5±0.6</td><td>52.4±0.7</td></tr></table>
+
+Table 6.  $\mathcal{T}_{\mathcal{SP}}$  Model Ablation and Cross-Architecture Testing: We use ResNet-18, ResNet-50, and ResNet-101 as  $\mathcal{T}_{\mathcal{SP}}$  models to generate synthetic images. During testing, the same three models are used. IPC is set to 10.
+
+![](images/710da0b5dfd0a15b8f27cf0cd423dfc27eb92886c2f2843ff03c914356ef0257.jpg)  
+(a) Accuracy vs. Guidance strength
+
+![](images/28c224eb7f4d1179d114e2528c56f2684e8881d26b53783ef767363b2beb0521.jpg)  
+(b) Accuracy vs. Sample Step
+
+![](images/56268c4bf81de374ccba9e17d01f26bc4e2c049a1a2d33703cbd143c4e5f24bf.jpg)  
+Figure 4. Hyper-parameter analysis on (a) guidance strength  $\lambda_{s}$  and (b) sampling step  $t$ . The star symbol indicates the parameters chosen for trade-off.  
+(a) Training Time overhead  
+Figure 5. Comparison of distillation cost on ImageNet-1k using our UniDD with comparison methods (MTT [3],  $\mathrm{SRe}^2\mathrm{L}$  [46], and  $\mathrm{D^4M}$  [41]), tested on V100 GPUs.
+
+![](images/3980bfe19f23ab16a4f8b59e65acfff76214a08a92e6d61e9526d8bf476c7cb1.jpg)  
+(b) Training GPU overhead
+
+ments are crucial for validating the generalizability of synthetic data. We compare different combinations of  $\mathcal{T}_{\mathcal{SP}}$  models and testing models. In this study, we set up three  $\mathcal{T}_{\mathcal{SP}}$  models: ResNet-18, ResNet-50, and ResNet-101. These models are also used during testing to evaluate the cross-architecture performance of the synthetic datasets. As shown in Table. 6, using more complex  $\mathcal{T}_{\mathcal{SP}}$  models improves the guidance effectiveness. Moreover, when testing across different architectures, the performance does not degrade as seen in previous optimization-based meth
+
+ods [3, 49, 50], demonstrating the generalizability of our synthetic data.
+
+Guidance Strength Analysis. As shown in Fig. 4 (a), we analyze the performance variations of ImageWoof and ImageNette at IPC-10 under different guidance strength settings  $\lambda_{s}$ , with Conv5 as the test model. The best results are achieved at  $\lambda_{s} = 100$ , but further increasing the guidance strength interferes with the quality of diffusion-generated images, leading to a decline in test performance.
+
+Sample Step Analysis. As shown in Fig. 4 (b), we analyze the impact of different SD sample steps on the performance of synthetic datasets under the IPC-10 setting, with Image-Woof and ImageNette resolutions set to  $256 \times 256$ . Performance improves significantly with sample steps below 250, but gains are minimal beyond this point while time consumption increases. Thus, we select 250 steps as the optimal trade-off.
+
+Distillation Cost Analysis. As shown in Fig. 5, we evaluate the time and GPU cost of various DD methods. Previous methods primarily focus on the time required to synthesize images, often neglecting the time needed to train or deploy the DD model. For MTT [3], the deployment time corresponds to the expert trajectory training time, while for  $\mathrm{D^4M}$  [41], it is the time required to finetune the SD model. The slight increase in UniDD's distillation time and GPU cost is mainly due to the added supervision process of the  $\mathcal{T}_{\mathcal{SP}}$  model and the refinement stage, where approximately  $5\% -10\%$  of low-quality generated data is discarded. Combining the deployment and distillation costs for comparison, UniDD demonstrates the best advantage, indicating its superior suitability for practical deployment.
+
+# 5. Conclusion
+
+In this study, we introduce the first Universal Dataset Distillation (UniDD) framework, capable of handling classification, detection, and segmentation tasks within a unified approach. UniDD follows a two-stage process: in the first stage, Task-Specific Proxy Training and Task-Aligned Contextual Prompting extract critical information from real datasets. In the second stage, Universal Taskdriven Diffusion synthesizes images with specific categories, bounding boxes, and segmentation masks tailored to each task. We conduct extensive experiments to validate UniDD's effectiveness across all three tasks. For classification, UniDD achieves state-of-the-art results on ImageNet1k and ImageNet-Subset. In detection and segmentation, we establish the first dataset distillation benchmark on Pascal VOC and MS COCO, comparing UniDD to coreset selection methods. Encouraging results demonstrate that our UniDD consistently outperforms these methods, especially under extremely low compression ratios.
+
+Acknowledgments. This work was supported by National Natural Science Fund of China (62473286).
+
+# References
+
+[1] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. 3, 4  
+[2] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 6  
+[3] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4750-4759, 2022. 1, 2, 5, 6, 8  
+[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6  
+[5] Yutian Chen, Max Welling, and Alex Smola. Super-samples from kernel herding. arXiv preprint arXiv:1203.3472, 2012. 6  
+[6] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.6  
+[7] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation toImagenet-1k with constant memory. In International Conference on Machine Learning, pages 6565–6590. PMLR, 2023. 1, 5  
+[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1, 2, 3, 5  
+[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3, 4  
+[10] 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. 3  
+[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 1, 2, 3, 5  
+[12] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 1, 2, 3, 5  
+[13] Junyao Gao, Xinyang Jiang, Huishuai Zhang, Yifan Yang, Shuguang Dou, Dongsheng Li, Duoqian Miao, Cheng Deng,
+
+and Cairong Zhao. Similarity distribution based membership inference attack on person re-identification. In Proceedings of the AAAI conference on artificial intelligence, pages 14820-14828, 2023. 1  
+[14] Junyao Gao, Xinyang Jiang, Shuguang Dou, Dongsheng Li, Duoqian Miao, and Cairong Zhao. Re-id-leak: Membership inference attacks against person re-identification. International Journal of Computer Vision, 132(10):4673-4687, 2024. 1  
+[15] Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. Styleshot: A snapshot on any style. arXiv preprint arXiv:2407.01414, 2024. 2  
+[16] Junyao Gao, Yanan Sun, Fei Shen, Xin Jiang, Zhening Xing, Kai Chen, and Cairong Zhao. Faceshot: Bring any character into life. arXiv preprint arXiv:2503.00740, 2025. 2  
+[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2  
+[18] Jianyang Gu, Saeed Vahidian, Vyacheslav Kungurtsev, Haonan Wang, Wei Jiang, Yang You, and Yiran Chen. Efficient dataset distillation via minimax diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15793-15803, 2024. 5, 6  
+[19] Ziyao Guo, Kai Wang, George Cazenavette, Hui Li, Kaipeng Zhang, and Yang You. Towards lossless dataset distillation via difficulty-aligned trajectory matching. arXiv preprint arXiv:2310.05773, 2023. 2  
+[20] 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. 1  
+[21] 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. 4, 5  
+[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1  
+[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3  
+[24] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 5  
+[25] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 2  
+[26] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangwoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In International Conference on Machine Learning, pages 11102-11118. PMLR, 2022. 1, 2
+
+[27] Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd: dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060-5069, 2019. 2  
+[28] 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. 1, 2, 5  
+[29] Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xin chao Wang. Dataset distillation via factorization. Advances in Neural Information Processing Systems, 35:1100-1113, 2022. 6  
+[30] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 3  
+[31] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. arXiv preprint arXiv:2302.14416, 2023. 2  
+[32] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. In International Conference on Learning Representations, 2020. 2  
+[33] OpenAI. Gpt-4 technical report. https://openai.com/research/gpt-4, 2023. Accessed: 2024-11-13. 2, 4  
+[34] Ding Qi, Jian Li, Jinlong Peng, Bo Zhao, Shuguang Dou, Jialin Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cairong Zhao. Fetch and forge: Efficient dataset condensation for object detection. Advances in Neural Information Processing Systems, 37:119283-119300, 2024. 3  
+[35] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 6  
+[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 4, 5, 6  
+[37] 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. 2, 5  
+[38] Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstantinos N Plataniotis. Datadam: Efficient dataset distillation with attention matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17097-17107, 2023. 6  
+[39] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489, 2017. 6
+
+[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3  
+[41] Duo Su, Junjie Hou, Weizhi Gao, Yingjie Tian, and Bowen Tang. D^4: Dataset distillation via disentangled diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5809-5818, 2024. 5, 6, 8  
+[42] Peng Sun, Bei Shi, Daiwei Yu, and Tao Lin. On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9390-9399, 2024. 1, 3, 5  
+[43] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12196-12205, 2022. 2  
+[44] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2, 6  
+[45] Yichao Yan, Zanwei Zhou, Zi Wang, Jingnan Gao, and Xiaokang Yang. Dialoguenerf: Towards realistic avatar face-to-face conversation video generation. Visual Intelligence, 2 (1):24, 2024. 1  
+[46] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. arXiv preprint arXiv:2306.13092, 2023. 1, 3, 5, 6, 8  
+[47] 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. 2  
+[48] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In International Conference on Machine Learning, pages 12674-12685. PMLR, 2021. 1  
+[49] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6514-6523, 2023. 1, 2, 6, 8  
+[50] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. In International Conference on Learning Representations, 2020. 1, 2, 6, 8  
+[51] Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved distribution matching for dataset condensation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7856-7865, 2023. 1, 2

2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b0686c7d02b839dd5d304a75e8e60ecdb9296f74202278a867ad64aaa6926e1
+size 647860

2025/Towards Universal Soccer Video Understanding/bd5f526c-bc30-45cc-bbcd-e151dfc21f1a_content_list.json

@@ -0,0 +1,1840 @@
+[
+    {
+        "type": "text",
+        "text": "Towards Universal Soccer Video Understanding",
+        "text_level": 1,
+        "bbox": [
+            256,
+            130,
+            741,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jiayuan Rao $^{1,2*}$ , Haoning Wu $^{1,2*}$ , Hao Jiang $^{3}$ , Ya Zhang $^{1}$ , Yanfeng Wang $^{1\\dagger}$ , Weidi Xie $^{1\\dagger}$ $^{1}$ School of Artificial Intelligence, Shanghai Jiao Tong University, China  \n $^{2}$ CMIC, Shanghai Jiao Tong University, China  \nhttps://jyrao.github.io/UniSoccer/",
+        "bbox": [
+            151,
+            179,
+            843,
+            262
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            248,
+            311,
+            326,
+            328
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "As a globally celebrated sport, soccer has attracted widespread interest from fans all over the world. This paper aims to develop a comprehensive multi-modal framework for soccer video understanding. Specifically, we make the following contributions in this paper: (i) we introduce SoccerReplay-1988, the largest multi-modal soccer dataset to date, featuring videos and detailed annotations from 1,988 complete matches, with an automated annotation pipeline; (ii) we present an advanced soccer-specific visual encoder, MatchVision, which leverages spatiotemporal information across soccer videos and excels in various downstream tasks; (iii) we conduct extensive experiments and ablation studies on event classification, commentary generation, and multi-view foul recognition. MatchVision demonstrates state-of-the-art performance on all of them, substantially outperforming existing models, which highlights the superiority of our proposed data and model. We believe that this work will offer a standard paradigm for sports understanding research.",
+        "bbox": [
+            88,
+            342,
+            483,
+            630
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "\"Football is one of the world's best means of communication. It is impartial, apolitical, and universal.\"",
+        "bbox": [
+            89,
+            637,
+            482,
+            667
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "— Franz Beckenbauer (1945 - 2024)",
+        "bbox": [
+            223,
+            667,
+            482,
+            683
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            91,
+            704,
+            220,
+            720
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Soccer, celebrated worldwide for its significant commercial value, has recently seen great research interest in integrating artificial intelligence (AI) for soccer video understanding. This is primarily motivated by the sport's complexity and the growing demand for enhanced analytics and improved viewing experiences. AI systems facilitate tactical analysis [49], allowing coaches to devise better strategies by uncovering patterns not apparent to the naked eye. In addition, it also supports automated content generation and enriches",
+        "bbox": [
+            89,
+            729,
+            483,
+            864
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/b768a20c91a1e5521d31cbbdecbc86cf5ce93b5b2297350c47f6ff106847ef5b.jpg",
+        "image_caption": [
+            "Videos in SoccerReplay-1988"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            297,
+            547,
+            315
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/d8fceea74ff9cd5dcca5ddf474b1fa93d9e505c697fe8845e096779cc9f8715b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            751,
+            300,
+            777,
+            315
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/ed384a51e59726eec53f918b94be181ccdfedb6766313c36bbcbed6f6b542de5.jpg",
+        "image_caption": [
+            "Figure 1. Overview. We present SoccerReplay-1988, the largest soccer dataset to date, and a powerful soccer-specific visual encoder, MatchVision, capable of excelling in various tasks such as event classification and commentary generation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            516,
+            319,
+            903,
+            547
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "fan engagement through interactive and personalized content [34, 36, 40]. These capabilities promote a deeper understanding of soccer, simplify content creation, and foster a more engaging experience for fans and professionals.",
+        "bbox": [
+            511,
+            626,
+            906,
+            686
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Existing research in soccer video analysis primarily revolves around the SoccerNet series datasets [7, 10, 15], which comprise 500 full-match videos for benchmarking various tasks, such as action spotting [10, 15] and commentary generation [34, 36, 40]. Despite this extensive coverage, the focus has predominantly been on designing specialized models for task-specific applications, leading to fragmented and incompatible solutions. Such fragmentation underscores the need for a unified framework capable of integrating diverse demands, enabling more holistic and scalable advancements in soccer video understanding.",
+        "bbox": [
+            511,
+            686,
+            908,
+            853
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we introduce SoccerReplay-1988, the largest and most comprehensive multi-modal soccer video dataset to date, featuring 1,988 complete match videos with",
+        "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_footnote",
+        "text": "*: These authors contribute equally to this work.",
+        "bbox": [
+            89,
+            875,
+            349,
+            887
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "†: Corresponding author.",
+        "bbox": [
+            91,
+            888,
+            225,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "8384",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "rich annotations, such as event labels and textual commentaries. This dataset offers a solid foundation for developing advanced soccer understanding models and establishes a challenging new benchmark for the field. Additionally, we have harmonized existing datasets to be compatible with ours, further expanding the available data resources.",
+        "bbox": [
+            89,
+            90,
+            480,
+            181
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Leveraging this dataset, we develop MatchVision, an advanced soccer-specific visual encoder tailored for diverse soccer understanding tasks. It employs the cutting-edge visual-language foundation model as the backbone, e.g., SigLIP [58]. We further extend framewise visual features into spatiotemporal representations with temporal attentions [3], by training on diverse visual-language tasks on SoccerReplay-1988, as depicted in Figure 1. As a result, MatchVision exhibits strong adaptability across various tasks, such as event classification and commentary generation, serving as a universal and unified framework for comprehensive soccer video understanding.",
+        "bbox": [
+            89,
+            181,
+            482,
+            362
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To summarize, we make the following contributions in this paper: (i) we construct SoccerReplay-1988, the largest and most diverse soccer video dataset to date, featuring videos of 1,988 soccer matches with rich annotations, supported by an automated curation pipeline. This provides a solid foundation for developing robust and comprehensive soccer understanding models; (ii) we present a powerful soccer-specific visual encoder, termed MatchVision, which effectively leverages spatiotemporal information in soccer videos, and can adapt to various tasks such as event classification and commentary generation, serving as a unified framework for soccer understanding; (iii) we establish more comprehensive and challenging benchmarks based on our dataset, enabling more professional evaluation of soccer understanding models; (iv) extensive experiments and ablation studies demonstrate the superiority of our data and model across various downstream tasks, achieving state-of-the-art performance on both existing benchmarks and our newly established ones. We believe this work offers a viable paradigm for future sports video understanding.",
+        "bbox": [
+            91,
+            363,
+            482,
+            664
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Works",
+        "text_level": 1,
+        "bbox": [
+            89,
+            676,
+            238,
+            691
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Sports Understanding [44] is an evolving field that encompasses multiple research topics and integrates diverse data modalities, covering various tasks such as action spotting [10, 15, 16], commentary generation [34, 37, 40, 51, 52, 57], athlete analysis [41, 55], tactical planning [49], sports health [39], and intelligent refereeing [22, 23]. Furthermore, with the rapid development of multimodal large language models (MLLMs), recent efforts [26, 50, 53, 54] have attempted to build more generalized frameworks to uniformly handle a variety of sports understanding tasks.",
+        "bbox": [
+            89,
+            700,
+            482,
+            852
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Visual-Language Models [1, 27, 28, 38, 58] have exhibited remarkable performance across extensive applications like classification, segmentation, image-text retrieval, and image",
+        "bbox": [
+            89,
+            854,
+            482,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "captioning. Recent efforts have ventured into more challenging video understanding [29, 30, 42, 59, 60] tasks, such as temporal alignment [17, 31], dense captioning [5, 56, 62], and audio description [18-20]. However, these efforts typically focus on general scenarios, limiting their adaptability to specific professional fields. Thus, this paper aims to bridge this gap by advancing visual-language models tailored for comprehensive soccer understanding.",
+        "bbox": [
+            511,
+            90,
+            903,
+            210
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Soccer Game Analysis [9] has primarily focused on tasks such as action spotting [15], replay grounding [10, 61], commentary generation [34, 36, 40], player tracking [8], state reconstruction [43], camera calibration [6, 8, 10] and foul recognition [22, 23], as facilitated by the SoccerNet [7, 10, 14, 15] series datasets, with 500 full-match videos from 2015 to 2017. Unlike existing methods that target designing specific models for distinct tasks, this paper aims to design a unified multi-modal framework that leverages spatiotemporal information within videos, serving as a specialized visual encoder for soccer video understanding.",
+        "bbox": [
+            511,
+            214,
+            906,
+            381
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3. SoccerReplay-1988 Dataset",
+        "text_level": 1,
+        "bbox": [
+            511,
+            397,
+            764,
+            414
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To establish a solid foundation for soccer understanding, we construct SoccerReplay-1988, the largest soccer dataset to date. Here, we first outline our data collection details and an overview of the dataset in Sec. 3.1; followed by elaborating on our automated data curation pipeline in Sec. 3.2; lastly, in Sec. 3.3, we present the data statistics and discussion.",
+        "bbox": [
+            511,
+            422,
+            903,
+            513
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.1. Dataset Collection",
+        "text_level": 1,
+        "bbox": [
+            511,
+            525,
+            689,
+            539
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To construct the SoccerReplay-1988 dataset, we have collected untrimmed, full-match videos from the Internet, encompassing a total of 1,988 matches from six European major soccer leagues and championships<sup>1</sup>, spanning the 2014-15 to 2023-24 seasons. For each match, we acquire textual commentaries with second-level timestamps from a sports text live website<sup>2</sup>, with part of them annotated with specific event types such as corner and goal. Additionally, we also incorporate extensive metadata, including detailed background information about the games, players, coaches, referees, and teams, providing a solid foundation for future soccer understanding research.",
+        "bbox": [
+            511,
+            547,
+            903,
+            729
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We partition the SoccerReplay-1988 dataset into train, validation, and test sets, containing 1,488, 250, and 250 full-match videos with diverse and comprehensive annotations, respectively. These sets provide rich training data for downstream tasks, such as event classification and commentary generation, while establishing comprehensive and challenging benchmarks for soccer understanding, as further discussed in subsequent sections.",
+        "bbox": [
+            511,
+            729,
+            903,
+            849
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>Premier (England), Laliga (Spain), Bundesliga (Germany), Serie-a (Italy), League-1 (France) and UEFA Champions League.",
+        "bbox": [
+            511,
+            862,
+            903,
+            887
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "2flashscore.com",
+        "bbox": [
+            514,
+            887,
+            602,
+            898
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "8385",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/ca6edb824bec182c0601875550b600151eadd1bf8ae1a66dbf292087b9c65f7b.jpg",
+        "image_caption": [
+            "Figure 2. Automated Data Curation Pipeline. The collected soccer video data are automatically processed for temporal alignment, event summarization, and anonymization by our curation pipeline."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            88,
+            477,
+            306
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Automated Data Curation",
+        "text_level": 1,
+        "bbox": [
+            89,
+            369,
+            328,
+            386
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given the potential noise in raw data, such as irrelevant video content, inaccurate timestamps, and incomplete event annotations, we design an automated data curation pipeline, comprising (i) temporal alignment, (ii) event summarization, and (iii) anonymization, as illustrated in Figure 2.",
+        "bbox": [
+            89,
+            392,
+            483,
+            469
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Temporal Alignment. Here, we divide match videos into two halves, each starting at kick-off, and adopt the temporal alignment model from MatchTime [40], to synchronize textual commentary timestamps with those of video frames.",
+        "bbox": [
+            89,
+            470,
+            483,
+            531
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Event Summarization. For samples without event annotations, we leverage LLaMA-3-70B [12] to summarize the events based on textual commentaries. Concretely, we have expanded the event categories from 17 in SoccerNet [10] to 24 types, for finer-grained soccer understanding, for example, categorizing penalties into scored and missed, and integrating modern soccer regulations like VAR. The resulting 24 event labels include: 'corner', 'goal', 'injury', 'own goal', 'penalty', 'penalty missed', 'red card', 'second yellow card', 'substitution', 'start of game (half)', 'end of game (half)', 'yellow card', 'throw in', 'free kick', 'saved by goal-keeper', 'shot off target', 'clearance', 'lead to corner', 'off-side', 'var', 'foul (no card)', 'statistics and summary', 'ball possession', and 'ball out of play'. More details on the used prompts are provided in the Appendix.",
+        "bbox": [
+            89,
+            534,
+            483,
+            762
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Anonymization. Similar to [34], we extract all person and team entity names from the metadata of SoccerReplay-1988, and replace them in textual commentaries with placeholders, such as \"[PLAYER]\", \"[TEAM]\", \"[COACH]\", and \"[REFEREE]\", ensuring consistency across tasks.",
+        "bbox": [
+            89,
+            763,
+            482,
+            840
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Moreover, our data curation pipeline can seamlessly extend to existing datasets, converting the SoccerNet series [10, 34] into our unified data format, termed SoccerNetpro. This expansion further enlarges the standardized",
+        "bbox": [
+            89,
+            840,
+            483,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/a4ff8d585d0999336f8d7af8e33491bd175228e27dd5d83ccbdf23ffa33fcb7b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td colspan=\"5\">Existing Datasets</td></tr><tr><td># Game</td><td>Duration(h)</td><td># Event</td><td># Anno.</td><td># Com.</td></tr><tr><td>SoccerNet-v1 [15]</td><td>500</td><td>764</td><td>7</td><td>6.7k</td><td>-</td></tr><tr><td>SoccerNet-v2 [10]</td><td>500</td><td>764</td><td>17</td><td>110k</td><td>-</td></tr><tr><td>MatchTime [40]</td><td>471</td><td>716</td><td>14</td><td>14k</td><td>37k</td></tr><tr><td>GOAL [36]</td><td>20</td><td>25.5</td><td>-</td><td>-</td><td>8.9k</td></tr><tr><td rowspan=\"2\"></td><td colspan=\"5\">Our Curated Datasets</td></tr><tr><td># Game</td><td>Duration(h)</td><td># Event</td><td># Anno.</td><td># Com.</td></tr><tr><td>SoccerNet-pro</td><td>500</td><td>764</td><td>24</td><td>102k</td><td>37k</td></tr><tr><td>SoccerReplay-1988</td><td>1,988</td><td>3,323</td><td>24</td><td>150k</td><td>150k</td></tr><tr><td>Integrated</td><td>2,488</td><td>4,087</td><td>24</td><td>252k</td><td>187k</td></tr></table>",
+        "bbox": [
+            517,
+            88,
+            905,
+            263
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Statistics of Soccer Datasets. Our SoccerReplay-1988 significantly surpasses existing datasets in both scale and diversity. Here, # Anno. and # Com. refer to the number of event annotations and textual commentaries, respectively.",
+        "bbox": [
+            511,
+            270,
+            906,
+            327
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "datasets available for soccer understanding tasks.",
+        "bbox": [
+            511,
+            340,
+            836,
+            357
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3. Statistics & Discussion",
+        "text_level": 1,
+        "bbox": [
+            511,
+            369,
+            723,
+            383
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Dataset Statistics. As shown in Table 1, our dataset encompasses 3,323 hours of footage from 1,988 soccer matches, with an average duration of 100.3 minutes per match. The videos range in resolution from 360p to 720p and frame rates between 25 and 30 FPS.",
+        "bbox": [
+            511,
+            393,
+            905,
+            467
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For textual annotations, this dataset features approximately 150K commentaries, averaging 76 per match, precisely temporal-aligned by the robust alignment model from MatchTime [40]. These commentaries cover 4,467 unique words, significantly surpassing the 2,873 words in existing datasets [34, 40], greatly enriching textual diversity. Automated event summarization based on these commentaries has yielded about 150K event annotations. Notably, a random sampling of  $2\\%$  of the data yields  $98\\%$  manual verification accuracy, ensuring high-quality automated labeling.",
+        "bbox": [
+            511,
+            469,
+            905,
+            621
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "SoccerReplay-test Benchmark. To facilitate a more comprehensive evaluation of soccer understanding models, we integrate 250 matches from SoccerReplay-1988 with 50 matches from the curated SoccerNet-pro, establishing SoccerReplay-test, a more challenging benchmark for event classification and commentary generation. This benchmark features nearly four times larger than existing datasets and comprises finer-grained event labels, richer textual commentaries, and up-to-date soccer regulations.",
+        "bbox": [
+            511,
+            625,
+            905,
+            761
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Discussion. To summarize, SoccerReplay-1988 exhibits advancements in three aspects: (i) it is the largest soccer video dataset to date, with nearly four times more videos than existing datasets; (ii) it features more professional and diverse annotations, more suitable for fine-grained and comprehensive soccer understanding tasks; (iii) It employs an automated curation pipeline for annotations and is thus scalable to provide a solid data foundation for future research. All data from SoccerReplay-1988, including videos",
+        "bbox": [
+            511,
+            763,
+            906,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "8386",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/e4c6e688d3460a7c195de0a3737ceca94017c7c0480ff3b70fd2a1036e793496.jpg",
+        "image_caption": [
+            "(a) Structure of MatchVision"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            89,
+            450,
+            353
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/977119086969bdcf69088f68e5878a2f6416a7dae1682c8d3d5c485bda592645.jpg",
+        "image_caption": [
+            "(b) MatchVision Pretraining",
+            "(c) Downstream Tasks",
+            "Figure 3. Overview of MatchVision. (a) The model architecture and its spatiotemporal feature extraction process; (b) Details of visual encoder pretraining, including supervised classification and video-language contrastive learning; (c) Implementation details of specific heads for various downstream tasks, including commentary generation, foul recognition, and event classification."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            464,
+            89,
+            903,
+            352
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "and annotations, are open-source for non-commercial use.",
+        "bbox": [
+            89,
+            435,
+            478,
+            450
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4. Method",
+        "text_level": 1,
+        "bbox": [
+            89,
+            462,
+            181,
+            478
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we aim to develop a soccer-specific visual encoder, MatchVision, tailored for diverse soccer video analysis tasks. We start by outlining our problem formulation in Sec. 4.1. Next, in Sec. 4.2, we detail the architecture of MatchVision. The training procedures are thoroughly discussed in Sec. 4.3. Finally, we describe the configurations for downstream tasks in Sec. 4.4, demonstrating the practical applications and effectiveness of our model.",
+        "bbox": [
+            89,
+            488,
+            483,
+            608
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1. Problem Formulation",
+        "text_level": 1,
+        "bbox": [
+            89,
+            617,
+            294,
+            631
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this work, we tackle the challenge of analyzing soccer video segments, denoted as  $\\mathcal{V} \\in \\mathbb{R}^{T \\times 3 \\times H \\times W}$ . Our goal is to utilize the visual encoder  $(\\Phi_{\\mathrm{MatchVision}})$  to extract spatiotemporal features from these segments, which are then processed by multiple task-specific heads, formulated as:",
+        "bbox": [
+            89,
+            638,
+            483,
+            715
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {E}, \\mathbf {C}, \\mathbf {F} = \\Psi (\\Phi_ {\\text {M a t c h V i s i o n}} (\\mathcal {V}))\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            179,
+            724,
+            392,
+            739
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Here,  $\\Psi = \\{\\Psi_{\\mathrm{cls}}, \\Psi_{\\mathrm{Cmt}}, \\Psi_{\\mathrm{Foul}}\\}$  represents the task-specific heads, with  $\\mathbf{E}$ ,  $\\mathbf{C}$ , and  $\\mathbf{F}$  denoting the output event types, textual commentaries, and foul types, respectively. This unified framework effectively learns relevant spatiotemporal features, and enables seamless integration across various downstream tasks for comprehensive soccer understanding.",
+        "bbox": [
+            89,
+            750,
+            482,
+            840
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2. Architecture",
+        "text_level": 1,
+        "bbox": [
+            89,
+            848,
+            225,
+            862
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "MatchVision comprises three key components: (i) Token Embedding, (ii) Spatiotemporal Attention Block, and (iii)",
+        "bbox": [
+            89,
+            869,
+            483,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Aggregation Layer, as depicted in Figure 3.",
+        "bbox": [
+            511,
+            435,
+            802,
+            450
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Token Embedding. In accordance with the convention in Vision Transformer [11], each frame  $(\\mathcal{I}_i)$  from the video segment  $(\\mathcal{V} = \\{\\mathcal{I}_1, \\mathcal{I}_2, \\dots, \\mathcal{I}_T\\})$  is divided into  $M$  non-overlapping patches of size  $P \\times P$  that span the entire frame.",
+        "bbox": [
+            511,
+            452,
+            905,
+            513
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "These patches are flattened into vectors  $(\\mathbf{x}_i^p)$ , where  $p$  and  $i$  denote the spatial and temporal positions, respectively. Each vector is transformed via an embedding layer  $(\\Phi_{\\mathrm{Emb}})$  into a token vector of size  $\\mathbb{R}^{1\\times D}$ , and then added with a spatial position embedding  $(\\mathbf{e}_{\\mathrm{s}}^{\\mathrm{pos}}\\in \\mathbb{R}^{M\\times D})$ . Subsequently, we concatenate a [c1s] token along with each frame. Finally, a temporal positional embedding  $(\\mathbf{e}_{\\mathrm{t}}^{\\mathrm{pos}}\\in \\mathbb{R}^{T\\times D})$  is added across features of all frames, as formulated below:",
+        "bbox": [
+            511,
+            513,
+            906,
+            633
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathbf {y} _ {i} = \\left[ \\mathbf {x} _ {i} ^ {\\mathrm {c l s}}, \\Phi_ {\\mathrm {E m b}} \\left(\\left[ \\mathbf {x} _ {i} ^ {1}, \\dots , \\mathbf {x} _ {i} ^ {M} \\right]\\right) + \\mathbf {e} _ {\\mathrm {s}} ^ {\\mathrm {p o s}} \\right] \\\\ \\mathbf {z} = \\left[ \\mathbf {y} _ {1}, \\dots , \\mathbf {y} _ {T} \\right] + \\mathbf {e} _ {\\mathrm {t}} ^ {\\text {p o s}} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            568,
+            640,
+            846,
+            676
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Here,  $[\\cdot ,\\cdot ]$  denotes concatenation, and  $\\mathbf{y}_i\\in \\mathbb{R}^{(M + 1)\\times D}$  represents the frame-wise features. The embedded features (z) will then serve as input for spatiotemporal attention blocks. Spatiotemporal Attention Block. Similar to TimeSformer [3], we utilize interleaved temporal and spatial attention to integrate spatiotemporal information in soccer videos. Concretely, each spatiotemporal attention block comprises a temporal self-attention layer and a spatial self-attention layer, i.e.,  $\\phi_t(\\cdot)$  and  $\\phi_s(\\cdot)$ , respectively.",
+        "bbox": [
+            511,
+            685,
+            905,
+            824
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Given a video feature  $(\\mathbf{z} \\in \\mathbb{R}^{T \\times (M + 1) \\times D})$ , we alternate temporal and spatial attention: temporal attention facilitates interactions among tokens at the same spatial positions across distinct frames, while spatial attention enables interactions among tokens within the same frame. Residual con",
+        "bbox": [
+            511,
+            824,
+            905,
+            898
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "8387",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "nections are employed in each layer. After passing through a total of  $\\mathbf{K}$  spatiotemporal attention blocks, the resulting feature  $(\\mathcal{F})$  captures both intra-frame and inter-frame relationships, i.e.,  $\\mathcal{F} = [\\phi_s(\\phi_t(\\mathbf{z}))]^{\\mathbf{K}} \\in \\mathbb{R}^{T \\times (M + 1) \\times D}$ .",
+        "bbox": [
+            89,
+            90,
+            480,
+            151
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Aggregation Layer. To obtain video-level features, we employ an aggregation layer on the frame-wise spatiotemporal features. Specifically, for the  $i$ -th frame, we utilize spatial self-attention to aggregate information into its [cls] token, denoted as  $\\hat{\\mathcal{F}}_i^{\\mathrm{cls}} = \\Phi_{\\mathrm{Agg}}(\\mathcal{F}_i)$ . Concatenating the [cls] tokens of all frames yields the final video feature  $(\\mathcal{F}_{\\mathcal{V}})$ , that effectively encapsulates spatiotemporal characteristics of soccer video segments, thus enabling it to be applicable for various downstream soccer understanding tasks. This process can be formulated as:",
+        "bbox": [
+            89,
+            154,
+            482,
+            305
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {F} _ {\\mathcal {V}} = \\Phi_ {\\mathrm {M a t c h V i s i o n}} (\\mathcal {V}) = [ \\hat {\\mathcal {F}} _ {1} ^ {\\mathrm {c l s}}, \\dots , \\hat {\\mathcal {F}} _ {T} ^ {\\mathrm {c l s}} ] \\in \\mathbb {R} ^ {T \\times D}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            112,
+            318,
+            460,
+            338
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3. MatchVision Pretraining",
+        "text_level": 1,
+        "bbox": [
+            89,
+            349,
+            318,
+            366
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this part, we aim to pretrain the visual encoder with triplet samples  $(\\{\\mathcal{V},\\mathbf{E},\\mathbf{C}\\})$ , comprising videos, event labels, and textual commentaries. Concretely, we investigate two distinct pretraining strategies: supervised classification and video-language contrastive learning.",
+        "bbox": [
+            89,
+            373,
+            482,
+            449
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Supervised Classification. One way to pretrain the visual encoder is supervised learning on event classification. To be specific, the extracted visual features  $(\\mathcal{F}_{\\mathcal{V}})$  are aggregated by a temporal self-attention layer into a learnable [cls] token, denoted as  $\\mathcal{F}_{\\mathcal{V}}^{\\mathrm{cls}}$ , similar to the spatial-wise aggregation mentioned above. This token is then fed into a linear classifier, and trained with a cross-entropy loss for event classification. The objective is denoted as  $\\mathcal{L}_{\\mathrm{sup}}$ .",
+        "bbox": [
+            89,
+            452,
+            482,
+            571
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Video-Language Contrastive Learning. As an alternative, we can also pretrain our visual encoder with video-text contrastive learning. Specifically, we adopt simple average pooling on the video feature to get the aggregated visual feature  $(\\mathcal{F}_{\\mathcal{V}}^{\\mathrm{Avg}})$ , and encode the textual commentary (C) with a text encoder  $(\\Phi_{\\mathrm{Text}})$ . We train the model with sigmoid loss ( $\\mathcal{L}_{\\mathrm{sigmoid}}$ ), as used in SigLIP [58]. Note that, some video clips may have highly similar commentaries, for example, 'start of the game', we treat the commentaries with high similarity in the same batch as positive samples when calculating loss functions. This can be expressed as follows:",
+        "bbox": [
+            89,
+            575,
+            482,
+            743
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {c o n t r a}} = \\mathcal {L} _ {\\text {s i g m o i d}} \\left(\\mathcal {F} _ {\\mathcal {V}} ^ {\\text {A v g}}, \\Phi_ {\\text {T e x t}} (\\mathbf {C})\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            160,
+            753,
+            413,
+            773
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.4. Downstream Tasks",
+        "text_level": 1,
+        "bbox": [
+            89,
+            787,
+            272,
+            801
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After the pretraining mentioned above, MatchVision can now serve as a versatile visual encoder, to map the soccer video segments into visual features  $(\\mathcal{F}_{\\mathcal{V}})$ , for training task-specific heads  $\\Psi = \\{\\Psi_{\\mathrm{cls}}, \\Psi_{\\mathrm{Cmt}}, \\Psi_{\\mathrm{Foul}}\\}$  across different downstream tasks, including: (i) event classification, (ii) commentary generation, and (iii) foul recognition.",
+        "bbox": [
+            89,
+            810,
+            482,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Event Classification. Similar to supervised classification above, we concatenate a learnable [cls] token to aggregate frame-wise visual features via temporal self-attention. This token is then fed into a linear classifier for event classification. The event classification head  $(\\Psi_{\\mathrm{cls}})$  is trained with a cross-entropy loss while freezing the visual encoder.",
+        "bbox": [
+            511,
+            90,
+            903,
+            181
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Commentary Generation. We follow the paradigm in MatchTime [40] to generate anonymized textual commentary for soccer video clips. Concretely, the commentary generation head  $(\\Psi_{\\mathrm{Cmt}})$  employs a Perceiver [25] aggregator to consolidate visual features, which are then projected by a trainable MLP, serving as prefix embeddings for a large language model (LLM). Subsequently, an off-the-shelf LLM decodes these embeddings into textual commentary. We adopt the negative log-likelihood loss, commonly used for auto-regressive next-token prediction.",
+        "bbox": [
+            511,
+            185,
+            906,
+            335
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Foul Recognition. As outlined in [22], the foul recognition task takes multi-view videos from the same scene as inputs, with each sample annotated with a foul class (8 types) and severity (4 levels). We encode these multi-view videos with MatchVision, and aggregate the extracted features into a single feature vector, via either max or average pooling, following the common practice. Subsequently, the foul recognition head  $(\\Psi_{\\mathrm{Foul}})$  employs a shared MLP and two task-specific linear classifiers, to predict foul type and severity, respectively. Similar to event classification, we use the combination of cross-entropy losses on the foul type and severity classification to jointly train  $\\Psi_{\\mathrm{Foul}}$ .",
+        "bbox": [
+            511,
+            339,
+            905,
+            520
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Discussion. Pretraining MatchVision on large-scale soccer data equips it with substantial domain-specific knowledge, enabling it to serve as a universal visual encoder adaptable to various downstream soccer understanding tasks.",
+        "bbox": [
+            511,
+            523,
+            905,
+            583
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            511,
+            599,
+            645,
+            616
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This section begins with implementation details in Sec. 5.1; followed by quantitative evaluations across downstream tasks in Sec. 5.2; then, we conduct ablation studies on our SoccerReplay-test benchmark to analyze the effectiveness of the proposed dataset and model in Sec. 5.3; finally, we provide qualitative results for comparison in Sec. 5.4.",
+        "bbox": [
+            511,
+            625,
+            905,
+            715
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1. Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            511,
+            727,
+            730,
+            742
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In our experiments, video segments are sampled at 1FPS around annotated timestamps, capturing a 30-second window for each sample. Frames are resized to  $224 \\times 224$  pixels as inputs. We initialize the embedding layer, spatial attention layers, aggregation layer, and text encoder of MatchVision with pretrained weights from SigLIP Base-16 [58] and adopt LLaMA-3 (8B) [12] as the off-the-shelf LLM decoder for commentary generation. All experiments are conducted on  $4 \\times$  Nvidia H800 GPUs with the AdamW [33] optimizer. Next, we elaborate on the training and evaluation details",
+        "bbox": [
+            511,
+            750,
+            906,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "8388",
+        "bbox": [
+            480,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/d502cee976c4dea31f3fc7e5a902e3e873bd8fdba7e42283767eafe88893f1b7.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Visual Encoder</td><td colspan=\"3\">Dataset</td><td colspan=\"3\">Classification (%)</td><td colspan=\"5\">Commentary</td></tr><tr><td>SN</td><td>MT</td><td>SR</td><td>Acc.@1</td><td>Acc.@3</td><td>Acc.@5</td><td>B@1</td><td>B@4</td><td>M</td><td>R-L</td><td>C</td></tr><tr><td colspan=\"12\">Off-the-shelf Models</td></tr><tr><td>I3D [4]</td><td>X</td><td>X</td><td>X</td><td>45.4</td><td>82.5</td><td>93.2</td><td>26.77</td><td>5.57</td><td>24.17</td><td>23.12</td><td>18.73</td></tr><tr><td>C3D [45]</td><td>X</td><td>X</td><td>X</td><td>47.8</td><td>85.1</td><td>95.0</td><td>28.13</td><td>6.64</td><td>24.52</td><td>24.23</td><td>27.88</td></tr><tr><td>ResNet [21]</td><td>X</td><td>X</td><td>X</td><td>47.2</td><td>84.6</td><td>94.4</td><td>27.34</td><td>6.57</td><td>24.72</td><td>24.43</td><td>27.29</td></tr><tr><td>CLIP [38]</td><td>X</td><td>X</td><td>X</td><td>48.5</td><td>85.5</td><td>95.2</td><td>26.25</td><td>6.51</td><td>24.27</td><td>24.75</td><td>28.17</td></tr><tr><td>InternVideo [48]</td><td>X</td><td>X</td><td>X</td><td>49.9</td><td>87.0</td><td>95.9</td><td>27.12</td><td>6.54</td><td>25.02</td><td>24.82</td><td>29.90</td></tr><tr><td>SigLIP [58]</td><td>X</td><td>X</td><td>X</td><td>50.2</td><td>86.7</td><td>95.6</td><td>27.85</td><td>6.98</td><td>25.16</td><td>25.03</td><td>31.38</td></tr><tr><td colspan=\"12\">Pretrain with Supervised Classification</td></tr><tr><td>Baidu [61]</td><td>✓</td><td>X</td><td>X</td><td>56.4</td><td>91.9</td><td>97.3</td><td>31.20</td><td>8.88</td><td>26.56</td><td>26.61</td><td>38.93</td></tr><tr><td>SigLIP</td><td>✓</td><td>X</td><td>X</td><td>55.9</td><td>89.6</td><td>94.9</td><td>28.51</td><td>7.39</td><td>25.96</td><td>25.94</td><td>35.71</td></tr><tr><td>SigLIP</td><td>✓</td><td>✓</td><td>✓</td><td>57.9</td><td>91.7</td><td>97.5</td><td>30.95</td><td>8.56</td><td>25.79</td><td>26.17</td><td>38.24</td></tr><tr><td>MatchVision</td><td>✓</td><td>X</td><td>X</td><td>82.5</td><td>96.6</td><td>98.8</td><td>29.45</td><td>7.92</td><td>26.01</td><td>26.21</td><td>36.15</td></tr><tr><td>MatchVision</td><td>✓</td><td>✓</td><td>✓</td><td>84.0</td><td>97.3</td><td>99.2</td><td>31.05</td><td>9.06</td><td>26.94</td><td>27.93</td><td>42.20</td></tr><tr><td colspan=\"12\">Pretrain with Visual-Language Contrastive Learning</td></tr><tr><td>SigLIP</td><td>X</td><td>✓</td><td>X</td><td>55.4</td><td>88.8</td><td>97.0</td><td>28.72</td><td>7.72</td><td>25.91</td><td>26.17</td><td>32.27</td></tr><tr><td>SigLIP</td><td>X</td><td>✓</td><td>✓</td><td>66.8</td><td>93.7</td><td>98.6</td><td>30.35</td><td>8.12</td><td>26.05</td><td>26.38</td><td>39.41</td></tr><tr><td>MatchVision</td><td>X</td><td>✓</td><td>X</td><td>58.9</td><td>89.0</td><td>97.1</td><td>30.33</td><td>7.97</td><td>25.48</td><td>26.33</td><td>33.87</td></tr><tr><td>MatchVision</td><td>X</td><td>✓</td><td>✓</td><td>67.9</td><td>93.9</td><td>98.6</td><td>31.94</td><td>9.12</td><td>26.24</td><td>27.56</td><td>40.76</td></tr><tr><td colspan=\"12\">Pretrain with Hybrid Supervised-Contrastive Training</td></tr><tr><td>SigLIP</td><td>✓</td><td>✓</td><td>X</td><td>71.2</td><td>94.5</td><td>98.7</td><td>28.63</td><td>7.82</td><td>25.74</td><td>25.35</td><td>34.09</td></tr><tr><td>SigLIP</td><td>✓</td><td>✓</td><td>✓</td><td>67.1</td><td>93.2</td><td>98.1</td><td>30.71</td><td>8.78</td><td>26.26</td><td>26.74</td><td>41.82</td></tr><tr><td>MatchVision</td><td>✓</td><td>✓</td><td>X</td><td>76.4</td><td>96.0</td><td>99.0</td><td>30.65</td><td>8.33</td><td>25.28</td><td>26.31</td><td>37.23</td></tr><tr><td>MatchVision</td><td>✓</td><td>✓</td><td>✓</td><td>80.1</td><td>97.1</td><td>99.1</td><td>33.58</td><td>9.14</td><td>26.82</td><td>28.21</td><td>44.18</td></tr></table>",
+        "bbox": [
+            142,
+            88,
+            852,
+            501
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Quantitative Results on Event Classification and Commentary Generation. Here, SN, MT, and SR represent finetuning with curated SoccerNet-v2 [10], MatchTime [40], and SoccerReplay-1988, respectively. B, M, R-L, and C refer to BLEU, METEOR, ROUGE-L, and CIDEr metrics, respectively. Within each unit, we denote the best performance in RED and the second-best performance in BLUE.",
+        "bbox": [
+            89,
+            508,
+            906,
+            551
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "about visual encoder pretraining and downstream tasks.",
+        "bbox": [
+            89,
+            561,
+            457,
+            577
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Visual Encoder Pretraining. For both pretraining strategies, we use a batch size of 40 for 15 epochs. The learning rate for all randomly initialized modules, including the temporal attention layers, aggregator layer, and linear classifier, is set to  $1 \\times 10^{-4}$ . Meanwhile, the learning rate for modules initialized with pretrained parameters (including the text encoder) is set to  $5 \\times 10^{-5}$ . In contrastive training, we adopt a multi-positive strategy where each textual commentary, based on its event label, considers closely related categories (e.g. \"start of game\" and \"offside\") as positive samples.",
+        "bbox": [
+            89,
+            580,
+            482,
+            731
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Downstream Tasks. In all downstream tasks, unless otherwise specified, we use the frozen visual encoder for feature extraction and only train the task-specific heads with a learning rate of  $1 \\times 10^{-4}$  for 30 epochs. The batch sizes for event classification, commentary generation, and foul recognition are set to 40, 32, and 8, respectively. We adopt specific evaluation metrics for these three tasks: (i) For event classification, we use the top-1/3/5 classification accuracy; (ii) For commentary generation, we employ several commonly-used language evaluation metrics, including BLEU [35], METEOR [2], ROUGE-L [32], and",
+        "bbox": [
+            89,
+            734,
+            483,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "CIDEr [47]; (iii) For foul recognition, we follow the common practice, and report top-1/2 and top-1 accuracy for the foul type and severity classification, respectively.",
+        "bbox": [
+            511,
+            561,
+            906,
+            607
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Benchmarks & Baselines. To ensure fair and reliable comparisons with existing work, we evaluate event classification (24 types) on 100 matches from curated SoccerNetv2 [10] test set; commentary generation on 49 matches from SN-Caption-test-align benchmark manually aligned in [40]; and foul recognition on MVFoul [22]. We consider various baselines: for the first two tasks, this includes off-the-shelf general visual encoders such as ResNet [21], C3D [45], I3D [4], CLIP [38], SigLIP [58], and InternVideo [48], along with Baidu [61] and SigLIP finetuned with soccer-specific data. For foul recognition, we follow previous work [22, 23] and adopt ResNet [21], R(2+1)D [46], and MViT [13] jointly finetuned with classifiers, as baselines.",
+        "bbox": [
+            511,
+            609,
+            908,
+            806
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2.Quantitative Evaluation",
+        "text_level": 1,
+        "bbox": [
+            511,
+            816,
+            735,
+            834
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As depicted in Table 2, we draw two observations on event classification and commentary generation: (i) visual encoders trained on soccer data substantially outperform off-the-shelf general encoders (ResNet, C3D, I3D, CLIP, and",
+        "bbox": [
+            511,
+            840,
+            906,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "8389",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/1e7396307c4fef1618267f34b5fe5bb12fa48b57efeace8acc9710a23730fc40.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"3\">Visual Encoder</td><td colspan=\"2\">Foul Class</td><td>Severity</td></tr><tr><td>Backbone</td><td>Train</td><td>Agg.</td><td>Acc.@1</td><td>Acc.@2</td><td>Acc.@1</td></tr><tr><td rowspan=\"2\">ResNet [21]</td><td rowspan=\"2\">✓</td><td>Mean</td><td>0.31</td><td>0.56</td><td>0.34</td></tr><tr><td>Max</td><td>0.32</td><td>0.60</td><td>0.32</td></tr><tr><td rowspan=\"2\">R(2+1)D [46]</td><td rowspan=\"2\">✓</td><td>Mean</td><td>0.31</td><td>0.55</td><td>0.34</td></tr><tr><td>Max</td><td>0.32</td><td>0.56</td><td>0.39</td></tr><tr><td rowspan=\"2\">MViT [13]</td><td rowspan=\"2\">✓</td><td>Mean</td><td>0.40</td><td>0.65</td><td>0.38</td></tr><tr><td>Max</td><td>0.47</td><td>0.69</td><td>0.43</td></tr><tr><td rowspan=\"2\">MatchVision</td><td rowspan=\"2\">×</td><td>Mean</td><td>0.44</td><td>0.53</td><td>0.58</td></tr><tr><td>Max</td><td>0.35</td><td>0.70</td><td>0.46</td></tr></table>",
+        "bbox": [
+            96,
+            88,
+            475,
+            267
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "InternVideo), underscoring the necessity of building specialized models for soccer understanding; (ii) almost all visual encoders, across all training settings, benefit from SoccerReplay-1988, emphasizing the value of constructing large-scale, high-quality data for soccer understanding. Next, we will delve into each task to discuss the results.",
+        "bbox": [
+            89,
+            333,
+            482,
+            422
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Event Classification. With identical training strategies and data, MatchVision considerably outperforms other methods in classification accuracy, demonstrating the superiority of its architecture, which effectively leverages spatiotemporal features within soccer videos. For example, MatchVision achieves a Top-1 accuracy of  $82.5\\%$ , significantly surpassing SigLIP's  $55.9\\%$  under the same training conditions. Moreover, models trained via supervised classification excel others, primarily because the pre-training task shares the same objectives as the downstream event classification task.",
+        "bbox": [
+            89,
+            426,
+            482,
+            578
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Commentary Generation. Visual encoders trained with visual-language contrastive learning exhibit better commentary generation performance than those trained with supervised classification, as this strategy better captures correlations between visual and textual features. Additionally, while MatchVision trained solely on SoccerNet slightly underperforms Baidu [61], incorporating SoccerReplay-1988 enables it to outperform on most metrics. This demonstrates that MatchVision can take advantage of large-scale datasets. Finally, a hybrid training approach, starting with supervised classification followed by visual-language contrastive learning, enables MatchVision to achieve optimal performance. This indicates that learning coarse-grained tasks such as classification provides a foundation for fine-grained tasks like commentary generation, and fully leveraging data unlocks the potential of soccer understanding.",
+        "bbox": [
+            89,
+            580,
+            482,
+            823
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Foul Recognition. As demonstrated in Table 3, MatchVision achieves performance comparable to jointly finetuned state-of-the-art methods in foul recognition, even with a frozen visual encoder. This highlights that MatchVision effectively learns substantial knowledge from large-scale soc",
+        "bbox": [
+            89,
+            824,
+            482,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/21608bf50b0ecb1bd5348751dc6d315743bf72ac03842ae185193d4c36abf0d4.jpg",
+        "table_caption": [
+            "Table 3. Quantitative Results on Multi-view Foul Recognition. Our frozen MatchVision encoder can achieve comparable performance with other jointly finetuned visual encoders."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"3\">Pretrain</td><td colspan=\"3\">Classification(%)</td></tr><tr><td>Sup.</td><td>Contra.</td><td>SR</td><td>Acc.@1</td><td>Acc.@3</td><td>Acc.@5</td></tr><tr><td>✓</td><td>X</td><td>X</td><td>62.67</td><td>83.00</td><td>89.81</td></tr><tr><td>✓</td><td>X</td><td>✓</td><td>68.03</td><td>86.90</td><td>92.38</td></tr><tr><td>X</td><td>✓</td><td>X</td><td>46.97</td><td>75.53</td><td>85.85</td></tr><tr><td>X</td><td>✓</td><td>✓</td><td>57.41</td><td>83.13</td><td>91.00</td></tr><tr><td>✓</td><td>✓</td><td>X</td><td>56.86</td><td>80.30</td><td>88.09</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>63.59</td><td>85.21</td><td>91.63</td></tr></table>",
+        "bbox": [
+            547,
+            88,
+            870,
+            234
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Ablations on Event Classification. We explore the impact of various training settings of our MatchVision encoder on the SoccerReplay-test benchmark. Here, Sup., Contra., and SR refer to supervised classification, visual-language contrastive learning, and the SoccerReplay-1988 dataset, respectively.",
+        "bbox": [
+            511,
+            239,
+            906,
+            311
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "cer data and adapts seamlessly to downstream tasks. Comparisons with additional baselines from the SoccerNet foul recognition challenges [9] are provided in the Appendix.",
+        "bbox": [
+            511,
+            328,
+            906,
+            375
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3. Ablation Studies",
+        "text_level": 1,
+        "bbox": [
+            511,
+            386,
+            679,
+            402
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We conduct ablation experiments on event classification and commentary generation using our SoccerReplay-test benchmark. These experiments validate the effectiveness of our proposed dataset and model, while establishing a baseline for future evaluations on this benchmark.",
+        "bbox": [
+            511,
+            409,
+            906,
+            484
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Event Classification. We evaluate event classification on 300 matches from our SoccerReplay-test benchmark using the MatchVision visual encoder pretrained with various strategies. Features are extracted by MatchVision and processed with a learnable aggregation layer and a linear classifier. The default training set is our curated SoccerNet-pro. As shown in Table 4, integrating SoccerReplay-1988 for training results in significant performance improvements across all pretraining strategies, yielding the significance of our dataset. Additionally, supervised classification outperforms visual-language contrastive learning and hybrid pretraining. This is due to its closer alignment with downstream event classification task, and the scale of event annotations is far larger than that of textual commentaries, further confirming the substantial benefits of data scaling for boosting soccer understanding.",
+        "bbox": [
+            511,
+            488,
+            908,
+            731
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Commentary Generation. With the pretrained MatchVision encoder, we train the commentary generation head on the MatchTime [40] and SoccerReplay-1988 datasets using various training strategies. By default, only the Perceiver [25] aggregation layer and projection layer within the head are trained. For joint training with the LLM decoder, considering computational costs, we incorporate LoRA [24] layers while freezing the original LLM layers. As shown in Table 5, incorporating SoccerReplay-1988 significantly improves performance on all metrics, confirming substantial advantages of our proposed dataset. This performance",
+        "bbox": [
+            511,
+            734,
+            908,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "8390",
+        "bbox": [
+            482,
+            944,
+            516,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/9d80294ea391f74578532a7a932bfd89ac9a197638371fc49f6dce60d8272ed9.jpg",
+        "image_caption": [
+            "Soccer Videos"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            90,
+            119,
+            104
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/9da4b340fa5e701f874da4eab9b7b914104a894fefb77e3bfec73dbc28c6aaee.jpg",
+        "image_caption": [
+            "Events"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            374,
+            92,
+            392,
+            104
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/dfe11a125f3ab53dc3d4ff9c58672ea87bd1c11d4c1cb608159d5405b50da324.jpg",
+        "image_caption": [
+            "Commentaries"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            90,
+            537,
+            104
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/8cdd4dbbecb8a9ec6e3fdaac06188a5bdb8d9593f88138abc2c3f127dc7a9875.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            107,
+            367,
+            150
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/68732df897e97929bec2c06a6cd4a0e00c3c827ef4c30cf5c5ca0c4f68c81307.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            372,
+            118,
+            459,
+            143
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/fc9c2bb97514f72b61de963098cb78a266efe3146a378467055940157f707b37.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            526,
+            114,
+            888,
+            147
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/52ed70c15dc4709676df0e4b456ec14925d160a95ba8aee8054e3fadab030edd.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            154,
+            367,
+            196
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/8276801c2dc956cf715cbb6e35392feb9ec6c5c1e35246581e3815e139c44262.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            372,
+            162,
+            514,
+            189
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/516894ea4acf31fed702636a2327972b70c6c0686e548ec9008c7942bb866031.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            527,
+            155,
+            890,
+            196
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/270fd650638147ed4738f17a034c436bce1a5e856c8100fe1265415bc8ef1e7e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            198,
+            367,
+            241
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b0de7aef3c8941434bfbb3eea98b327a676dacfc6f7f83d3791ac06e9b00f0f7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            372,
+            209,
+            488,
+            234
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/a380b47f26580073f501fca6a80af2d27b0dff7e9f36d75a2471d72d0c93d535.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            527,
+            200,
+            890,
+            241
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/3eefcf41d5f78e47d7d847d40c2c7378f9003d67dcc7f9abc8a12c41c3846a2e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            243,
+            367,
+            287
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/3b4d60d20742533f20aab26bf153e9159875fb7e3fc21dc2d1ab6a2e9a7e5be3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            372,
+            253,
+            477,
+            281
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6bdbf1b0e4e1e262d7f85f385ddf1be5a1f0c0e018b967e3eade70b46797fd21.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            527,
+            247,
+            888,
+            287
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/c4b3b98a4be399f96a73b58101fd17cc5b827266907bce570f39aa35c3e2d7ff.jpg",
+        "image_caption": [
+            "Figure 4. Qualitative Results for Event Classification and Commentary Generation. Here, \"w/o SR\" and \"w/ SR\" indicate models trained without and with the SoccerReplay-1988 dataset, respectively. Incorporating SoccerReplay-1988 improves event classification accuracy. Moreover, this enriched training data enables the model to demonstrate several advantages in commentary generation: (a) more detailed descriptions, (b) greater linguistic variety, (c) higher event depiction accuracy, (d) better adherence to updated rules, and (e) improved specificity in scenario response."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            290,
+            367,
+            333
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/9b8addab99c97babf3f33b9d8f1d4cb75828a7be35583514c63ea2e65e97bcee.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            372,
+            300,
+            460,
+            327
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/39aeeb1cf36f84a1d0fe1d9adf45176e87fd7b4d6118c5b90ca344994a7a1b2b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            527,
+            297,
+            879,
+            330
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/e1760a007d98c191a6ad49503d990e87dd10e26631844807f40b023b88dfe480.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Trainable</td><td colspan=\"5\">Commentary Metrics</td></tr><tr><td>V</td><td>L</td><td>B@1</td><td>B@4</td><td>M</td><td>R-L</td><td>C</td></tr><tr><td colspan=\"7\">Trained on MatchTime</td></tr><tr><td>X</td><td>X</td><td>21.65</td><td>3.27</td><td>21.02</td><td>17.79</td><td>12.90</td></tr><tr><td>✓</td><td>X</td><td>27.62</td><td>7.02</td><td>24.03</td><td>23.51</td><td>30.77</td></tr><tr><td>X</td><td>✓</td><td>27.04</td><td>6.41</td><td>24.15</td><td>23.88</td><td>31.91</td></tr><tr><td>✓</td><td>✓</td><td>27.49</td><td>6.96</td><td>24.50</td><td>23.33</td><td>30.81</td></tr><tr><td colspan=\"7\">Trained on MatchTime &amp; SoccerReplay-1988</td></tr><tr><td>X</td><td>X</td><td>24.17</td><td>4.09</td><td>20.51</td><td>20.70</td><td>15.70</td></tr><tr><td>✓</td><td>X</td><td>28.98</td><td>8.39</td><td>24.45</td><td>25.35</td><td>45.85</td></tr><tr><td>X</td><td>✓</td><td>27.54</td><td>7.76</td><td>24.50</td><td>24.70</td><td>42.79</td></tr><tr><td>✓</td><td>✓</td><td>29.21</td><td>8.22</td><td>25.25</td><td>25.54</td><td>43.18</td></tr></table>",
+        "bbox": [
+            112,
+            435,
+            457,
+            642
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5. Ablations on Commentary Generation. We investigate the impact of different training strategies and datasets on MatchVision using the SoccerReplay-test benchmark. 'V' and 'L' denote the visual encoder and the LLM decoder, respectively.",
+        "bbox": [
+            89,
+            650,
+            482,
+            705
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "gap also reflects the challenges of our established benchmark, which features diverse vocabulary, richer semantics, and updated soccer rules. Additionally, jointly finetuning the visual encoder and the LLM decoder provides a feasible approach for further improvements.",
+        "bbox": [
+            89,
+            717,
+            483,
+            792
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4. Qualitative Comparisons",
+        "text_level": 1,
+        "bbox": [
+            89,
+            801,
+            318,
+            819
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As depicted in Figure 4, we present qualitative results of MatchVision on the SoccerReplay-test benchmark, comparing models pretrained with and without SoccerReplay-1988. For event classification, incorporating our data improves accuracy, and even in misclassified cases, the results",
+        "bbox": [
+            89,
+            825,
+            483,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "remain contextually relevant. For commentary generation, hybrid training on SoccerReplay-1988 enables MatchVision to produce richer, more detailed textual commentary, reflecting a deeper understanding of soccer dynamics. More qualitative results are available in the Appendix.",
+        "bbox": [
+            511,
+            438,
+            906,
+            513
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            513,
+            541,
+            633,
+            556
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we establish a unified, scalable multi-modal framework for soccer understanding. Specifically, we introduce SoccerReplay-1988, the largest and most comprehensive soccer video dataset to date, annotated by an automated curation pipeline. This provides a solid foundation for developing soccer understanding models and serves as a more challenging benchmark. Built upon this, we develop MatchVision, an advanced soccer-specific visual encoder, which effectively leverages spatiotemporal information within soccer videos and can be applied to various tasks such as event classification and commentary generation. Extensive experiments demonstrate the superiority of our model, with MatchVision achieving state-of-the-art performance on both existing benchmarks and our newly established one. We believe this work will set a viable, universal paradigm for future research in sports understanding.",
+        "bbox": [
+            511,
+            571,
+            906,
+            814
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            513,
+            840,
+            671,
+            857
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This work is funded by National Key R&D Program of China (No.2022ZD0161400).",
+        "bbox": [
+            511,
+            869,
+            906,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "8391",
+        "bbox": [
+            482,
+            944,
+            513,
+            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] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. 2",
+            "[2] 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, pages 65-72, 2005. 6",
+            "[3] 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, page 4, 2021. 2, 4",
+            "[4] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6",
+            "[5] 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, and Sergey Tulyakov. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2",
+            "[6] Anthony Cioppa, Adrien Deliege, Floriane Magera, Silvio Giancola, Olivier Barnich, Bernard Ghanem, and Marc Van Droogenbroeck. Camera calibration and player localization in soccernet-v2 and investigation of their representations for action spotting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 4537-4546, 2021. 2",
+            "[7] Anthony Cioppa, Adrien Deliege, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. Scaling up soccer-net with multi-view spatial localization and re-identification. Scientific data, 9(1):355, 2022. 1, 2",
+            "[8] Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3491-3502, 2022. 2",
+            "[9] Anthony Cioppa, Silvio Giancola, Vladimir Somers, Victor Joos, Floriane Magera, Jan Held, Seyed Abolfazl Ghasemzadeh, Xin Zhou, Karolina Seweryn, Mateusz Kowalczyk, Zuzanna Mroz, Szymon Lukasik, Michal Halon, Hassan Mkhallati, Adrien Deliege, Carlos Hinojosa, Karen Sanchez, Amir M. Mansourian, Pierre Miralles, Olivier Barnich, Christophe De Vleeschouwer, Alexandre Alahi, Bernard Ghanem, Marc Van Droogenbroeck, Adam Gorski, Albert Clapes, Andrei Boiarov, Anton Afanasiev, Artur Xarles, Atom Scott, ByoungKwon Lim, Calvin Yeung, Cristian Gonzalez, Dominic Rufenacht, Enzo Pacilio, Fabian Deuser, Faisal Sami Altawijri, Francisco Cachon, HanKyul Kim, Haobo Wang, Hyeonmin Choe, Hyunwoo J Kim, Il-Min"
+        ],
+        "bbox": [
+            99,
+            114,
+            483,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Kim, Jae-Mo Kang, Jamshid Tursunboev, Jian Yang, Jihwan Hong, Jimin Lee, Jing Zhang, Junseok Lee, Kexin Zhang, Konrad Habel, Licheng Jiao, Linyi Li, Marc Gutierrez-Perez, Marcelo Ortega, Menglong Li, Milosz Lopatto, Nikita Kasatkin, Nikolay Nemtsev, Norbert Oswald, Oleg Udin, Pavel Kononov, Pei Geng, Saad Ghazai Alotaibi, Sehyung Kim, Sergei Ulasen, Sergio Escalera, Shanshan Zhang, Shuyuan Yang, Sunghwan Moon, Thomas B. Moeslund, Vasyl Shandyba, Vladimir Golovkin, Wei Dai, WonTaek Chung, Xinyu Liu, Yongqiang Zhu, Youngseo Kim, Yuan Li, Yuting Yang, Yuxuan Xiao, Zehua Cheng, and Zhihao Li. Soccernet 2024 challenges results. arXiv preprint arXiv:2409.10587, 2024. 2, 7",
+            "[10] Adrien Deliege, Anthony Cioppa, Silvio Giancola, Meisam J Seikavandi, Jacob V Dueholm, Kamal Nasrollahi, Bernard Ghanem, Thomas B Moeslund, and Marc Van Droogenbroeck. Soccernet-v2: A dataset and benchmarks for holistic understanding of broadcast soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 4508-4519, 2021. 1, 2, 3, 6",
+            "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations, 2021. 4",
+            "[12] 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 preprint arXiv:2407.21783, 2024. 3, 5",
+            "[13] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6824-6835, 2021. 6, 7",
+            "[14] Sushant Gautam, Mehdi Houshmand Sarkhoosh, Jan Held, Cise Midoglu, Anthony Cioppa, Silvio Giancola, Vajira Thambawita, Michael A Riegler, Pål Halvorsen, and Mubarak Shah. Soccernet-echoes: A soccer game audio commentary dataset. arXiv preprint arXiv:2405.07354, 2024. 2",
+            "[15] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. Soccernet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 1711-1721, 2018. 1, 2, 3",
+            "[16] Xiaofan Gu, Xinwei Xue, and Feng Wang. Fine-grained action recognition on a novel basketball dataset. In International Conference on Acoustics, Speech, and Signal Processing, pages 2563-2567, 2020. 2",
+            "[17] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2906-2916, 2022. 2",
+            "[18] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad: Movie description in"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "8392",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2",
+            "[19] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel-who, when, and what in movie audio description. In Proceedings of the International Conference on Computer Vision, 2023.",
+            "[20] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad iii: The prequel - back to the pixels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2",
+            "[21] 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, 7",
+            "[22] Jan Held, Anthony Cioppa, Silvio Giancola, Abdullah Hamdi, Bernard Ghanem, and Marc Van Droogenbroeck. Vars: Video assistant referee system for automated soccer decision making from multiple views. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 5085-5096, 2023. 2, 5, 6",
+            "[23] Jan Held, Hani Itani, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. X-vars: Introducing explainability in football refereeing with multimodal large language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3267-3279, 2024. 2, 6",
+            "[24] 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. In Proceedings of the International Conference on Learning Representations, 2022. 7",
+            "[25] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In Proceedings of the International Conference on Machine Learning, pages 4651-4664, 2021. 5, 7",
+            "[26] Haopeng Li, Andong Deng, Qiuhong Ke, Jun Liu, Hossein Rahmani, Yulan Guo, Bernt Schiele, and Chen Chen. Sports-qa: A large-scale video question answering benchmark for complex and professional sports. arXiv preprint arXiv:2401.01505, 2024. 2",
+            "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the International Conference on Machine Learning, pages 12888-12900, 2022. 2",
+            "[28] 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 Machine Learning, pages 19730–19742, 2023. 2",
+            "[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2"
+        ],
+        "bbox": [
+            91,
+            92,
+            482,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[30] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In Proceedings of the European Conference on Computer Vision, 2024. 2",
+            "[31] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhang, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. In Proceedings of the European Conference on Computer Vision, 2024. 2",
+            "[32] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, 2004. 6",
+            "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proceedings of the International Conference on Learning Representations, 2019. 5",
+            "[34] Hassan Mkhallati, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. Socernetc caption: Dense video captioning for soccer broadcasts commentaries. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 5074-5085, 2023. 1, 2, 3",
+            "[35] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Association for Computational Linguistics, pages 311-318, 2002. 6",
+            "[36] Ji Qi, Jifan Yu, Teng Tu, Kunyu Gao, Yifan Xu, Xinyu Guan, Xiaozhi Wang, Bin Xu, Lei Hou, Juanzi Li, et al. Goal: A challenging knowledge-grounded video captioning benchmark for real-time soccer commentary generation. In Proceedings of the ACM International Conference on Information and Knowledge Management, pages 5391-5395, 2023. 1, 2, 3",
+            "[37] Mengshi Qi, Yunhong Wang, Annan Li, and Jiebo Luo. Sports video captioning via attentive motion representation and group relationship modeling. IEEE Transactions on Circuits and Systems for Video Technology, 2019. 2",
+            "[38] 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 Proceedings of the International Conference on Machine Learning, 2021. 2, 6",
+            "[39] Prem N Ramkumar, Bryan C Luu, Heather S Haeberle, Jaret M Karnuta, Benedict U Nwachukwu, and Riley J Williams. Sports medicine and artificial intelligence: a primer. The American Journal of Sports Medicine, 50(4): 1166-1174, 2022. 2",
+            "[40] Jiayuan Rao, Haoning Wu, Chang Liu, Yanfeng Wang, and Weidi Xie. Matchtime: Towards automatic soccer game commentary generation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2024. 1, 2, 3, 5, 6, 7",
+            "[41] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2616-2625, 2020. 2",
+            "[42] Yan Shu, Zheng Liu, Peitian Zhang, Minghao Qin, Junjie Zhou, Zhengyang Liang, Tiejun Huang, and Bo Zhao."
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            901
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "8393",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Video-xl: Extra-long vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 2",
+            "[43] Vladimir Somers, Victor Joos, Anthony Cioppa, Silvio Giancola, Seyed Abolfazl Ghasemzadeh, Floriane Magera, Baptiste Standaert, Amir M Mansourian, Xin Zhou, Shohreh Kasaei, et al. Soccernet game state reconstruction: End-to-end athlete tracking and identification on a minimap. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3293-3305, 2024. 2",
+            "[44] Graham Thomas, Rikke Gade, Thomas B Moeslund, Peter Carr, and Adrian Hilton. Computer vision for sports: Current applications and research topics. Computer Vision and Image Understanding, 159:3-18, 2017. 2",
+            "[45] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the International Conference on Computer Vision, pages 4489-4497, 2015. 6",
+            "[46] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 6, 7",
+            "[47] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4566-4575, 2015. 6",
+            "[48] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 6",
+            "[49] Zhe Wang, Petar Velicković, Daniel Hennes, Nenad Tomasev, Laurel Prince, Michael Kaisers, Yoram Bachrach, Romuald Elie, Li Kevin Wenliang, Federico Piccinini, et al. Tacticai: an ai assistant for football tactics. Nature Communications, 15(1):1-13, 2024. 1, 2",
+            "[50] Dekun Wu, He Zhao, Xingce Bao, and Richard P Wildes. Sports video analysis on large-scale data. In Proceedings of the European Conference on Computer Vision, 2022. 2",
+            "[51] Zeyu Xi, Ge Shi, Xuefen Li, Junchi Yan, Zun Li, Lifang Wu, Zilin Liu, and Liang Wang. A simple yet effective knowledge guided method for entity-aware video captioning on a basketball benchmark. Neurocomputing, 2025. 2",
+            "[52] Zeyu Xi, Ge Shi, Haoying Sun, Bowen Zhang, Shuyi Li, and Lifang Wu. Eika: Explicit & implicit knowledge-augmented network for entity-aware sports video captioning. Expert Systems with Applications, 2025. 2",
+            "[53] Haotian Xia, Zhengbang Yang, Yuqing Wang, Rhys Tracy, Yun Zhao, Dongdong Huang, Zezhi Chen, Yan Zhu, Yuanfang Wang, and Weining Shen. Sportqa: A benchmark for sports understanding in large language models. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics, 2024. 2",
+            "[54] Haotian Xia, Zhengbang Yang, Junbo Zou, Rhys Tracy, Yuqing Wang, Chi Lu, Christopher Lai, Yanjun He, Xun"
+        ],
+        "bbox": [
+            91,
+            90,
+            482,
+            901
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Shao, Zhuoqing Xie, et al. Sportu: A comprehensive sports understanding benchmark for multimodal large language models. In Proceedings of the International Conference on Learning Representations, 2025. 2",
+            "[55] Jinglin Xu, Yongming Rao, Xumin Yu, Guangyi Chen, Jie Zhou, and Jiwen Lu. Finediving: A fine-grained dataset for procedure-aware action quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2949-2958, 2022. 2",
+            "[56] 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 Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2",
+            "[57] Huanyu Yu, Shuo Cheng, Bingbing Ni, Minsi Wang, Jian Zhang, and Xiaokang Yang. Fine-grained video captioning for sports narrative. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2",
+            "[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the International Conference on Computer Vision, pages 11975-11986, 2023. 2, 5, 6",
+            "[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2023. 2",
+            "[60] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2",
+            "[61] Xin Zhou, Le Kang, Zhiyu Cheng, Bo He, and Jingyu Xin. Feature combination meets attention: Baidu soccer embeddings and transformer based temporal detection. arXiv preprint arXiv:2106.14447, 2021. 2, 6, 7",
+            "[62] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            669
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "8394",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2025/Towards Universal Soccer Video Understanding/bd5f526c-bc30-45cc-bbcd-e151dfc21f1a_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e040d999cee2b6f12893a0062c59b25c6b4f72c8cdd5de5d684b7b8360bcefe
+size 7755598

2025/Towards Universal Soccer Video Understanding/full.md

@@ -0,0 +1,348 @@
+# Towards Universal Soccer Video Understanding
+
+Jiayuan Rao $^{1,2*}$ , Haoning Wu $^{1,2*}$ , Hao Jiang $^{3}$ , Ya Zhang $^{1}$ , Yanfeng Wang $^{1\dagger}$ , Weidi Xie $^{1\dagger}$ $^{1}$ School of Artificial Intelligence, Shanghai Jiao Tong University, China  
+ $^{2}$ CMIC, Shanghai Jiao Tong University, China  
+https://jyrao.github.io/UniSoccer/
+
+# Abstract
+
+As a globally celebrated sport, soccer has attracted widespread interest from fans all over the world. This paper aims to develop a comprehensive multi-modal framework for soccer video understanding. Specifically, we make the following contributions in this paper: (i) we introduce SoccerReplay-1988, the largest multi-modal soccer dataset to date, featuring videos and detailed annotations from 1,988 complete matches, with an automated annotation pipeline; (ii) we present an advanced soccer-specific visual encoder, MatchVision, which leverages spatiotemporal information across soccer videos and excels in various downstream tasks; (iii) we conduct extensive experiments and ablation studies on event classification, commentary generation, and multi-view foul recognition. MatchVision demonstrates state-of-the-art performance on all of them, substantially outperforming existing models, which highlights the superiority of our proposed data and model. We believe that this work will offer a standard paradigm for sports understanding research.
+
+"Football is one of the world's best means of communication. It is impartial, apolitical, and universal."
+
+— Franz Beckenbauer (1945 - 2024)
+
+# 1. Introduction
+
+Soccer, celebrated worldwide for its significant commercial value, has recently seen great research interest in integrating artificial intelligence (AI) for soccer video understanding. This is primarily motivated by the sport's complexity and the growing demand for enhanced analytics and improved viewing experiences. AI systems facilitate tactical analysis [49], allowing coaches to devise better strategies by uncovering patterns not apparent to the naked eye. In addition, it also supports automated content generation and enriches
+
+![](images/b768a20c91a1e5521d31cbbdecbc86cf5ce93b5b2297350c47f6ff106847ef5b.jpg)  
+Videos in SoccerReplay-1988
+
+![](images/d8fceea74ff9cd5dcca5ddf474b1fa93d9e505c697fe8845e096779cc9f8715b.jpg)
+
+![](images/ed384a51e59726eec53f918b94be181ccdfedb6766313c36bbcbed6f6b542de5.jpg)  
+Figure 1. Overview. We present SoccerReplay-1988, the largest soccer dataset to date, and a powerful soccer-specific visual encoder, MatchVision, capable of excelling in various tasks such as event classification and commentary generation.
+
+fan engagement through interactive and personalized content [34, 36, 40]. These capabilities promote a deeper understanding of soccer, simplify content creation, and foster a more engaging experience for fans and professionals.
+
+Existing research in soccer video analysis primarily revolves around the SoccerNet series datasets [7, 10, 15], which comprise 500 full-match videos for benchmarking various tasks, such as action spotting [10, 15] and commentary generation [34, 36, 40]. Despite this extensive coverage, the focus has predominantly been on designing specialized models for task-specific applications, leading to fragmented and incompatible solutions. Such fragmentation underscores the need for a unified framework capable of integrating diverse demands, enabling more holistic and scalable advancements in soccer video understanding.
+
+In this paper, we introduce SoccerReplay-1988, the largest and most comprehensive multi-modal soccer video dataset to date, featuring 1,988 complete match videos with
+
+rich annotations, such as event labels and textual commentaries. This dataset offers a solid foundation for developing advanced soccer understanding models and establishes a challenging new benchmark for the field. Additionally, we have harmonized existing datasets to be compatible with ours, further expanding the available data resources.
+
+Leveraging this dataset, we develop MatchVision, an advanced soccer-specific visual encoder tailored for diverse soccer understanding tasks. It employs the cutting-edge visual-language foundation model as the backbone, e.g., SigLIP [58]. We further extend framewise visual features into spatiotemporal representations with temporal attentions [3], by training on diverse visual-language tasks on SoccerReplay-1988, as depicted in Figure 1. As a result, MatchVision exhibits strong adaptability across various tasks, such as event classification and commentary generation, serving as a universal and unified framework for comprehensive soccer video understanding.
+
+To summarize, we make the following contributions in this paper: (i) we construct SoccerReplay-1988, the largest and most diverse soccer video dataset to date, featuring videos of 1,988 soccer matches with rich annotations, supported by an automated curation pipeline. This provides a solid foundation for developing robust and comprehensive soccer understanding models; (ii) we present a powerful soccer-specific visual encoder, termed MatchVision, which effectively leverages spatiotemporal information in soccer videos, and can adapt to various tasks such as event classification and commentary generation, serving as a unified framework for soccer understanding; (iii) we establish more comprehensive and challenging benchmarks based on our dataset, enabling more professional evaluation of soccer understanding models; (iv) extensive experiments and ablation studies demonstrate the superiority of our data and model across various downstream tasks, achieving state-of-the-art performance on both existing benchmarks and our newly established ones. We believe this work offers a viable paradigm for future sports video understanding.
+
+# 2. Related Works
+
+Sports Understanding [44] is an evolving field that encompasses multiple research topics and integrates diverse data modalities, covering various tasks such as action spotting [10, 15, 16], commentary generation [34, 37, 40, 51, 52, 57], athlete analysis [41, 55], tactical planning [49], sports health [39], and intelligent refereeing [22, 23]. Furthermore, with the rapid development of multimodal large language models (MLLMs), recent efforts [26, 50, 53, 54] have attempted to build more generalized frameworks to uniformly handle a variety of sports understanding tasks.
+
+Visual-Language Models [1, 27, 28, 38, 58] have exhibited remarkable performance across extensive applications like classification, segmentation, image-text retrieval, and image
+
+captioning. Recent efforts have ventured into more challenging video understanding [29, 30, 42, 59, 60] tasks, such as temporal alignment [17, 31], dense captioning [5, 56, 62], and audio description [18-20]. However, these efforts typically focus on general scenarios, limiting their adaptability to specific professional fields. Thus, this paper aims to bridge this gap by advancing visual-language models tailored for comprehensive soccer understanding.
+
+Soccer Game Analysis [9] has primarily focused on tasks such as action spotting [15], replay grounding [10, 61], commentary generation [34, 36, 40], player tracking [8], state reconstruction [43], camera calibration [6, 8, 10] and foul recognition [22, 23], as facilitated by the SoccerNet [7, 10, 14, 15] series datasets, with 500 full-match videos from 2015 to 2017. Unlike existing methods that target designing specific models for distinct tasks, this paper aims to design a unified multi-modal framework that leverages spatiotemporal information within videos, serving as a specialized visual encoder for soccer video understanding.
+
+# 3. SoccerReplay-1988 Dataset
+
+To establish a solid foundation for soccer understanding, we construct SoccerReplay-1988, the largest soccer dataset to date. Here, we first outline our data collection details and an overview of the dataset in Sec. 3.1; followed by elaborating on our automated data curation pipeline in Sec. 3.2; lastly, in Sec. 3.3, we present the data statistics and discussion.
+
+# 3.1. Dataset Collection
+
+To construct the SoccerReplay-1988 dataset, we have collected untrimmed, full-match videos from the Internet, encompassing a total of 1,988 matches from six European major soccer leagues and championships<sup>1</sup>, spanning the 2014-15 to 2023-24 seasons. For each match, we acquire textual commentaries with second-level timestamps from a sports text live website<sup>2</sup>, with part of them annotated with specific event types such as corner and goal. Additionally, we also incorporate extensive metadata, including detailed background information about the games, players, coaches, referees, and teams, providing a solid foundation for future soccer understanding research.
+
+We partition the SoccerReplay-1988 dataset into train, validation, and test sets, containing 1,488, 250, and 250 full-match videos with diverse and comprehensive annotations, respectively. These sets provide rich training data for downstream tasks, such as event classification and commentary generation, while establishing comprehensive and challenging benchmarks for soccer understanding, as further discussed in subsequent sections.
+
+![](images/ca6edb824bec182c0601875550b600151eadd1bf8ae1a66dbf292087b9c65f7b.jpg)  
+Figure 2. Automated Data Curation Pipeline. The collected soccer video data are automatically processed for temporal alignment, event summarization, and anonymization by our curation pipeline.
+
+# 3.2. Automated Data Curation
+
+Given the potential noise in raw data, such as irrelevant video content, inaccurate timestamps, and incomplete event annotations, we design an automated data curation pipeline, comprising (i) temporal alignment, (ii) event summarization, and (iii) anonymization, as illustrated in Figure 2.
+
+Temporal Alignment. Here, we divide match videos into two halves, each starting at kick-off, and adopt the temporal alignment model from MatchTime [40], to synchronize textual commentary timestamps with those of video frames.
+
+Event Summarization. For samples without event annotations, we leverage LLaMA-3-70B [12] to summarize the events based on textual commentaries. Concretely, we have expanded the event categories from 17 in SoccerNet [10] to 24 types, for finer-grained soccer understanding, for example, categorizing penalties into scored and missed, and integrating modern soccer regulations like VAR. The resulting 24 event labels include: 'corner', 'goal', 'injury', 'own goal', 'penalty', 'penalty missed', 'red card', 'second yellow card', 'substitution', 'start of game (half)', 'end of game (half)', 'yellow card', 'throw in', 'free kick', 'saved by goal-keeper', 'shot off target', 'clearance', 'lead to corner', 'off-side', 'var', 'foul (no card)', 'statistics and summary', 'ball possession', and 'ball out of play'. More details on the used prompts are provided in the Appendix.
+
+Anonymization. Similar to [34], we extract all person and team entity names from the metadata of SoccerReplay-1988, and replace them in textual commentaries with placeholders, such as "[PLAYER]", "[TEAM]", "[COACH]", and "[REFEREE]", ensuring consistency across tasks.
+
+Moreover, our data curation pipeline can seamlessly extend to existing datasets, converting the SoccerNet series [10, 34] into our unified data format, termed SoccerNetpro. This expansion further enlarges the standardized
+
+<table><tr><td rowspan="2"></td><td colspan="5">Existing Datasets</td></tr><tr><td># Game</td><td>Duration(h)</td><td># Event</td><td># Anno.</td><td># Com.</td></tr><tr><td>SoccerNet-v1 [15]</td><td>500</td><td>764</td><td>7</td><td>6.7k</td><td>-</td></tr><tr><td>SoccerNet-v2 [10]</td><td>500</td><td>764</td><td>17</td><td>110k</td><td>-</td></tr><tr><td>MatchTime [40]</td><td>471</td><td>716</td><td>14</td><td>14k</td><td>37k</td></tr><tr><td>GOAL [36]</td><td>20</td><td>25.5</td><td>-</td><td>-</td><td>8.9k</td></tr><tr><td rowspan="2"></td><td colspan="5">Our Curated Datasets</td></tr><tr><td># Game</td><td>Duration(h)</td><td># Event</td><td># Anno.</td><td># Com.</td></tr><tr><td>SoccerNet-pro</td><td>500</td><td>764</td><td>24</td><td>102k</td><td>37k</td></tr><tr><td>SoccerReplay-1988</td><td>1,988</td><td>3,323</td><td>24</td><td>150k</td><td>150k</td></tr><tr><td>Integrated</td><td>2,488</td><td>4,087</td><td>24</td><td>252k</td><td>187k</td></tr></table>
+
+Table 1. Statistics of Soccer Datasets. Our SoccerReplay-1988 significantly surpasses existing datasets in both scale and diversity. Here, # Anno. and # Com. refer to the number of event annotations and textual commentaries, respectively.
+
+datasets available for soccer understanding tasks.
+
+# 3.3. Statistics & Discussion
+
+Dataset Statistics. As shown in Table 1, our dataset encompasses 3,323 hours of footage from 1,988 soccer matches, with an average duration of 100.3 minutes per match. The videos range in resolution from 360p to 720p and frame rates between 25 and 30 FPS.
+
+For textual annotations, this dataset features approximately 150K commentaries, averaging 76 per match, precisely temporal-aligned by the robust alignment model from MatchTime [40]. These commentaries cover 4,467 unique words, significantly surpassing the 2,873 words in existing datasets [34, 40], greatly enriching textual diversity. Automated event summarization based on these commentaries has yielded about 150K event annotations. Notably, a random sampling of  $2\%$  of the data yields  $98\%$  manual verification accuracy, ensuring high-quality automated labeling.
+
+SoccerReplay-test Benchmark. To facilitate a more comprehensive evaluation of soccer understanding models, we integrate 250 matches from SoccerReplay-1988 with 50 matches from the curated SoccerNet-pro, establishing SoccerReplay-test, a more challenging benchmark for event classification and commentary generation. This benchmark features nearly four times larger than existing datasets and comprises finer-grained event labels, richer textual commentaries, and up-to-date soccer regulations.
+
+Discussion. To summarize, SoccerReplay-1988 exhibits advancements in three aspects: (i) it is the largest soccer video dataset to date, with nearly four times more videos than existing datasets; (ii) it features more professional and diverse annotations, more suitable for fine-grained and comprehensive soccer understanding tasks; (iii) It employs an automated curation pipeline for annotations and is thus scalable to provide a solid data foundation for future research. All data from SoccerReplay-1988, including videos
+
+![](images/e4c6e688d3460a7c195de0a3737ceca94017c7c0480ff3b70fd2a1036e793496.jpg)  
+(a) Structure of MatchVision
+
+![](images/977119086969bdcf69088f68e5878a2f6416a7dae1682c8d3d5c485bda592645.jpg)  
+(b) MatchVision Pretraining  
+(c) Downstream Tasks  
+Figure 3. Overview of MatchVision. (a) The model architecture and its spatiotemporal feature extraction process; (b) Details of visual encoder pretraining, including supervised classification and video-language contrastive learning; (c) Implementation details of specific heads for various downstream tasks, including commentary generation, foul recognition, and event classification.
+
+and annotations, are open-source for non-commercial use.
+
+# 4. Method
+
+In this paper, we aim to develop a soccer-specific visual encoder, MatchVision, tailored for diverse soccer video analysis tasks. We start by outlining our problem formulation in Sec. 4.1. Next, in Sec. 4.2, we detail the architecture of MatchVision. The training procedures are thoroughly discussed in Sec. 4.3. Finally, we describe the configurations for downstream tasks in Sec. 4.4, demonstrating the practical applications and effectiveness of our model.
+
+# 4.1. Problem Formulation
+
+In this work, we tackle the challenge of analyzing soccer video segments, denoted as  $\mathcal{V} \in \mathbb{R}^{T \times 3 \times H \times W}$ . Our goal is to utilize the visual encoder  $(\Phi_{\mathrm{MatchVision}})$  to extract spatiotemporal features from these segments, which are then processed by multiple task-specific heads, formulated as:
+
+$$
+\mathbf {E}, \mathbf {C}, \mathbf {F} = \Psi (\Phi_ {\text {M a t c h V i s i o n}} (\mathcal {V}))
+$$
+
+Here,  $\Psi = \{\Psi_{\mathrm{cls}}, \Psi_{\mathrm{Cmt}}, \Psi_{\mathrm{Foul}}\}$  represents the task-specific heads, with  $\mathbf{E}$ ,  $\mathbf{C}$ , and  $\mathbf{F}$  denoting the output event types, textual commentaries, and foul types, respectively. This unified framework effectively learns relevant spatiotemporal features, and enables seamless integration across various downstream tasks for comprehensive soccer understanding.
+
+# 4.2. Architecture
+
+MatchVision comprises three key components: (i) Token Embedding, (ii) Spatiotemporal Attention Block, and (iii)
+
+Aggregation Layer, as depicted in Figure 3.
+
+Token Embedding. In accordance with the convention in Vision Transformer [11], each frame  $(\mathcal{I}_i)$  from the video segment  $(\mathcal{V} = \{\mathcal{I}_1, \mathcal{I}_2, \dots, \mathcal{I}_T\})$  is divided into  $M$  non-overlapping patches of size  $P \times P$  that span the entire frame.
+
+These patches are flattened into vectors  $(\mathbf{x}_i^p)$ , where  $p$  and  $i$  denote the spatial and temporal positions, respectively. Each vector is transformed via an embedding layer  $(\Phi_{\mathrm{Emb}})$  into a token vector of size  $\mathbb{R}^{1\times D}$ , and then added with a spatial position embedding  $(\mathbf{e}_{\mathrm{s}}^{\mathrm{pos}}\in \mathbb{R}^{M\times D})$ . Subsequently, we concatenate a [c1s] token along with each frame. Finally, a temporal positional embedding  $(\mathbf{e}_{\mathrm{t}}^{\mathrm{pos}}\in \mathbb{R}^{T\times D})$  is added across features of all frames, as formulated below:
+
+$$
+\begin{array}{l} \mathbf {y} _ {i} = \left[ \mathbf {x} _ {i} ^ {\mathrm {c l s}}, \Phi_ {\mathrm {E m b}} \left(\left[ \mathbf {x} _ {i} ^ {1}, \dots , \mathbf {x} _ {i} ^ {M} \right]\right) + \mathbf {e} _ {\mathrm {s}} ^ {\mathrm {p o s}} \right] \\ \mathbf {z} = \left[ \mathbf {y} _ {1}, \dots , \mathbf {y} _ {T} \right] + \mathbf {e} _ {\mathrm {t}} ^ {\text {p o s}} \\ \end{array}
+$$
+
+Here,  $[\cdot ,\cdot ]$  denotes concatenation, and  $\mathbf{y}_i\in \mathbb{R}^{(M + 1)\times D}$  represents the frame-wise features. The embedded features (z) will then serve as input for spatiotemporal attention blocks. Spatiotemporal Attention Block. Similar to TimeSformer [3], we utilize interleaved temporal and spatial attention to integrate spatiotemporal information in soccer videos. Concretely, each spatiotemporal attention block comprises a temporal self-attention layer and a spatial self-attention layer, i.e.,  $\phi_t(\cdot)$  and  $\phi_s(\cdot)$ , respectively.
+
+Given a video feature  $(\mathbf{z} \in \mathbb{R}^{T \times (M + 1) \times D})$ , we alternate temporal and spatial attention: temporal attention facilitates interactions among tokens at the same spatial positions across distinct frames, while spatial attention enables interactions among tokens within the same frame. Residual con
+
+nections are employed in each layer. After passing through a total of  $\mathbf{K}$  spatiotemporal attention blocks, the resulting feature  $(\mathcal{F})$  captures both intra-frame and inter-frame relationships, i.e.,  $\mathcal{F} = [\phi_s(\phi_t(\mathbf{z}))]^{\mathbf{K}} \in \mathbb{R}^{T \times (M + 1) \times D}$ .
+
+Aggregation Layer. To obtain video-level features, we employ an aggregation layer on the frame-wise spatiotemporal features. Specifically, for the  $i$ -th frame, we utilize spatial self-attention to aggregate information into its [cls] token, denoted as  $\hat{\mathcal{F}}_i^{\mathrm{cls}} = \Phi_{\mathrm{Agg}}(\mathcal{F}_i)$ . Concatenating the [cls] tokens of all frames yields the final video feature  $(\mathcal{F}_{\mathcal{V}})$ , that effectively encapsulates spatiotemporal characteristics of soccer video segments, thus enabling it to be applicable for various downstream soccer understanding tasks. This process can be formulated as:
+
+$$
+\mathcal {F} _ {\mathcal {V}} = \Phi_ {\mathrm {M a t c h V i s i o n}} (\mathcal {V}) = [ \hat {\mathcal {F}} _ {1} ^ {\mathrm {c l s}}, \dots , \hat {\mathcal {F}} _ {T} ^ {\mathrm {c l s}} ] \in \mathbb {R} ^ {T \times D}
+$$
+
+# 4.3. MatchVision Pretraining
+
+In this part, we aim to pretrain the visual encoder with triplet samples  $(\{\mathcal{V},\mathbf{E},\mathbf{C}\})$ , comprising videos, event labels, and textual commentaries. Concretely, we investigate two distinct pretraining strategies: supervised classification and video-language contrastive learning.
+
+Supervised Classification. One way to pretrain the visual encoder is supervised learning on event classification. To be specific, the extracted visual features  $(\mathcal{F}_{\mathcal{V}})$  are aggregated by a temporal self-attention layer into a learnable [cls] token, denoted as  $\mathcal{F}_{\mathcal{V}}^{\mathrm{cls}}$ , similar to the spatial-wise aggregation mentioned above. This token is then fed into a linear classifier, and trained with a cross-entropy loss for event classification. The objective is denoted as  $\mathcal{L}_{\mathrm{sup}}$ .
+
+Video-Language Contrastive Learning. As an alternative, we can also pretrain our visual encoder with video-text contrastive learning. Specifically, we adopt simple average pooling on the video feature to get the aggregated visual feature  $(\mathcal{F}_{\mathcal{V}}^{\mathrm{Avg}})$ , and encode the textual commentary (C) with a text encoder  $(\Phi_{\mathrm{Text}})$ . We train the model with sigmoid loss ( $\mathcal{L}_{\mathrm{sigmoid}}$ ), as used in SigLIP [58]. Note that, some video clips may have highly similar commentaries, for example, 'start of the game', we treat the commentaries with high similarity in the same batch as positive samples when calculating loss functions. This can be expressed as follows:
+
+$$
+\mathcal {L} _ {\text {c o n t r a}} = \mathcal {L} _ {\text {s i g m o i d}} \left(\mathcal {F} _ {\mathcal {V}} ^ {\text {A v g}}, \Phi_ {\text {T e x t}} (\mathbf {C})\right)
+$$
+
+# 4.4. Downstream Tasks
+
+After the pretraining mentioned above, MatchVision can now serve as a versatile visual encoder, to map the soccer video segments into visual features  $(\mathcal{F}_{\mathcal{V}})$ , for training task-specific heads  $\Psi = \{\Psi_{\mathrm{cls}}, \Psi_{\mathrm{Cmt}}, \Psi_{\mathrm{Foul}}\}$  across different downstream tasks, including: (i) event classification, (ii) commentary generation, and (iii) foul recognition.
+
+Event Classification. Similar to supervised classification above, we concatenate a learnable [cls] token to aggregate frame-wise visual features via temporal self-attention. This token is then fed into a linear classifier for event classification. The event classification head  $(\Psi_{\mathrm{cls}})$  is trained with a cross-entropy loss while freezing the visual encoder.
+
+Commentary Generation. We follow the paradigm in MatchTime [40] to generate anonymized textual commentary for soccer video clips. Concretely, the commentary generation head  $(\Psi_{\mathrm{Cmt}})$  employs a Perceiver [25] aggregator to consolidate visual features, which are then projected by a trainable MLP, serving as prefix embeddings for a large language model (LLM). Subsequently, an off-the-shelf LLM decodes these embeddings into textual commentary. We adopt the negative log-likelihood loss, commonly used for auto-regressive next-token prediction.
+
+Foul Recognition. As outlined in [22], the foul recognition task takes multi-view videos from the same scene as inputs, with each sample annotated with a foul class (8 types) and severity (4 levels). We encode these multi-view videos with MatchVision, and aggregate the extracted features into a single feature vector, via either max or average pooling, following the common practice. Subsequently, the foul recognition head  $(\Psi_{\mathrm{Foul}})$  employs a shared MLP and two task-specific linear classifiers, to predict foul type and severity, respectively. Similar to event classification, we use the combination of cross-entropy losses on the foul type and severity classification to jointly train  $\Psi_{\mathrm{Foul}}$ .
+
+Discussion. Pretraining MatchVision on large-scale soccer data equips it with substantial domain-specific knowledge, enabling it to serve as a universal visual encoder adaptable to various downstream soccer understanding tasks.
+
+# 5. Experiments
+
+This section begins with implementation details in Sec. 5.1; followed by quantitative evaluations across downstream tasks in Sec. 5.2; then, we conduct ablation studies on our SoccerReplay-test benchmark to analyze the effectiveness of the proposed dataset and model in Sec. 5.3; finally, we provide qualitative results for comparison in Sec. 5.4.
+
+# 5.1. Implementation Details
+
+In our experiments, video segments are sampled at 1FPS around annotated timestamps, capturing a 30-second window for each sample. Frames are resized to  $224 \times 224$  pixels as inputs. We initialize the embedding layer, spatial attention layers, aggregation layer, and text encoder of MatchVision with pretrained weights from SigLIP Base-16 [58] and adopt LLaMA-3 (8B) [12] as the off-the-shelf LLM decoder for commentary generation. All experiments are conducted on  $4 \times$  Nvidia H800 GPUs with the AdamW [33] optimizer. Next, we elaborate on the training and evaluation details
+
+<table><tr><td rowspan="2">Visual Encoder</td><td colspan="3">Dataset</td><td colspan="3">Classification (%)</td><td colspan="5">Commentary</td></tr><tr><td>SN</td><td>MT</td><td>SR</td><td>Acc.@1</td><td>Acc.@3</td><td>Acc.@5</td><td>B@1</td><td>B@4</td><td>M</td><td>R-L</td><td>C</td></tr><tr><td colspan="12">Off-the-shelf Models</td></tr><tr><td>I3D [4]</td><td>X</td><td>X</td><td>X</td><td>45.4</td><td>82.5</td><td>93.2</td><td>26.77</td><td>5.57</td><td>24.17</td><td>23.12</td><td>18.73</td></tr><tr><td>C3D [45]</td><td>X</td><td>X</td><td>X</td><td>47.8</td><td>85.1</td><td>95.0</td><td>28.13</td><td>6.64</td><td>24.52</td><td>24.23</td><td>27.88</td></tr><tr><td>ResNet [21]</td><td>X</td><td>X</td><td>X</td><td>47.2</td><td>84.6</td><td>94.4</td><td>27.34</td><td>6.57</td><td>24.72</td><td>24.43</td><td>27.29</td></tr><tr><td>CLIP [38]</td><td>X</td><td>X</td><td>X</td><td>48.5</td><td>85.5</td><td>95.2</td><td>26.25</td><td>6.51</td><td>24.27</td><td>24.75</td><td>28.17</td></tr><tr><td>InternVideo [48]</td><td>X</td><td>X</td><td>X</td><td>49.9</td><td>87.0</td><td>95.9</td><td>27.12</td><td>6.54</td><td>25.02</td><td>24.82</td><td>29.90</td></tr><tr><td>SigLIP [58]</td><td>X</td><td>X</td><td>X</td><td>50.2</td><td>86.7</td><td>95.6</td><td>27.85</td><td>6.98</td><td>25.16</td><td>25.03</td><td>31.38</td></tr><tr><td colspan="12">Pretrain with Supervised Classification</td></tr><tr><td>Baidu [61]</td><td>✓</td><td>X</td><td>X</td><td>56.4</td><td>91.9</td><td>97.3</td><td>31.20</td><td>8.88</td><td>26.56</td><td>26.61</td><td>38.93</td></tr><tr><td>SigLIP</td><td>✓</td><td>X</td><td>X</td><td>55.9</td><td>89.6</td><td>94.9</td><td>28.51</td><td>7.39</td><td>25.96</td><td>25.94</td><td>35.71</td></tr><tr><td>SigLIP</td><td>✓</td><td>✓</td><td>✓</td><td>57.9</td><td>91.7</td><td>97.5</td><td>30.95</td><td>8.56</td><td>25.79</td><td>26.17</td><td>38.24</td></tr><tr><td>MatchVision</td><td>✓</td><td>X</td><td>X</td><td>82.5</td><td>96.6</td><td>98.8</td><td>29.45</td><td>7.92</td><td>26.01</td><td>26.21</td><td>36.15</td></tr><tr><td>MatchVision</td><td>✓</td><td>✓</td><td>✓</td><td>84.0</td><td>97.3</td><td>99.2</td><td>31.05</td><td>9.06</td><td>26.94</td><td>27.93</td><td>42.20</td></tr><tr><td colspan="12">Pretrain with Visual-Language Contrastive Learning</td></tr><tr><td>SigLIP</td><td>X</td><td>✓</td><td>X</td><td>55.4</td><td>88.8</td><td>97.0</td><td>28.72</td><td>7.72</td><td>25.91</td><td>26.17</td><td>32.27</td></tr><tr><td>SigLIP</td><td>X</td><td>✓</td><td>✓</td><td>66.8</td><td>93.7</td><td>98.6</td><td>30.35</td><td>8.12</td><td>26.05</td><td>26.38</td><td>39.41</td></tr><tr><td>MatchVision</td><td>X</td><td>✓</td><td>X</td><td>58.9</td><td>89.0</td><td>97.1</td><td>30.33</td><td>7.97</td><td>25.48</td><td>26.33</td><td>33.87</td></tr><tr><td>MatchVision</td><td>X</td><td>✓</td><td>✓</td><td>67.9</td><td>93.9</td><td>98.6</td><td>31.94</td><td>9.12</td><td>26.24</td><td>27.56</td><td>40.76</td></tr><tr><td colspan="12">Pretrain with Hybrid Supervised-Contrastive Training</td></tr><tr><td>SigLIP</td><td>✓</td><td>✓</td><td>X</td><td>71.2</td><td>94.5</td><td>98.7</td><td>28.63</td><td>7.82</td><td>25.74</td><td>25.35</td><td>34.09</td></tr><tr><td>SigLIP</td><td>✓</td><td>✓</td><td>✓</td><td>67.1</td><td>93.2</td><td>98.1</td><td>30.71</td><td>8.78</td><td>26.26</td><td>26.74</td><td>41.82</td></tr><tr><td>MatchVision</td><td>✓</td><td>✓</td><td>X</td><td>76.4</td><td>96.0</td><td>99.0</td><td>30.65</td><td>8.33</td><td>25.28</td><td>26.31</td><td>37.23</td></tr><tr><td>MatchVision</td><td>✓</td><td>✓</td><td>✓</td><td>80.1</td><td>97.1</td><td>99.1</td><td>33.58</td><td>9.14</td><td>26.82</td><td>28.21</td><td>44.18</td></tr></table>
+
+Table 2. Quantitative Results on Event Classification and Commentary Generation. Here, SN, MT, and SR represent finetuning with curated SoccerNet-v2 [10], MatchTime [40], and SoccerReplay-1988, respectively. B, M, R-L, and C refer to BLEU, METEOR, ROUGE-L, and CIDEr metrics, respectively. Within each unit, we denote the best performance in RED and the second-best performance in BLUE.
+
+about visual encoder pretraining and downstream tasks.
+
+Visual Encoder Pretraining. For both pretraining strategies, we use a batch size of 40 for 15 epochs. The learning rate for all randomly initialized modules, including the temporal attention layers, aggregator layer, and linear classifier, is set to  $1 \times 10^{-4}$ . Meanwhile, the learning rate for modules initialized with pretrained parameters (including the text encoder) is set to  $5 \times 10^{-5}$ . In contrastive training, we adopt a multi-positive strategy where each textual commentary, based on its event label, considers closely related categories (e.g. "start of game" and "offside") as positive samples.
+
+Downstream Tasks. In all downstream tasks, unless otherwise specified, we use the frozen visual encoder for feature extraction and only train the task-specific heads with a learning rate of  $1 \times 10^{-4}$  for 30 epochs. The batch sizes for event classification, commentary generation, and foul recognition are set to 40, 32, and 8, respectively. We adopt specific evaluation metrics for these three tasks: (i) For event classification, we use the top-1/3/5 classification accuracy; (ii) For commentary generation, we employ several commonly-used language evaluation metrics, including BLEU [35], METEOR [2], ROUGE-L [32], and
+
+CIDEr [47]; (iii) For foul recognition, we follow the common practice, and report top-1/2 and top-1 accuracy for the foul type and severity classification, respectively.
+
+Benchmarks & Baselines. To ensure fair and reliable comparisons with existing work, we evaluate event classification (24 types) on 100 matches from curated SoccerNetv2 [10] test set; commentary generation on 49 matches from SN-Caption-test-align benchmark manually aligned in [40]; and foul recognition on MVFoul [22]. We consider various baselines: for the first two tasks, this includes off-the-shelf general visual encoders such as ResNet [21], C3D [45], I3D [4], CLIP [38], SigLIP [58], and InternVideo [48], along with Baidu [61] and SigLIP finetuned with soccer-specific data. For foul recognition, we follow previous work [22, 23] and adopt ResNet [21], R(2+1)D [46], and MViT [13] jointly finetuned with classifiers, as baselines.
+
+# 5.2.Quantitative Evaluation
+
+As depicted in Table 2, we draw two observations on event classification and commentary generation: (i) visual encoders trained on soccer data substantially outperform off-the-shelf general encoders (ResNet, C3D, I3D, CLIP, and
+
+<table><tr><td colspan="3">Visual Encoder</td><td colspan="2">Foul Class</td><td>Severity</td></tr><tr><td>Backbone</td><td>Train</td><td>Agg.</td><td>Acc.@1</td><td>Acc.@2</td><td>Acc.@1</td></tr><tr><td rowspan="2">ResNet [21]</td><td rowspan="2">✓</td><td>Mean</td><td>0.31</td><td>0.56</td><td>0.34</td></tr><tr><td>Max</td><td>0.32</td><td>0.60</td><td>0.32</td></tr><tr><td rowspan="2">R(2+1)D [46]</td><td rowspan="2">✓</td><td>Mean</td><td>0.31</td><td>0.55</td><td>0.34</td></tr><tr><td>Max</td><td>0.32</td><td>0.56</td><td>0.39</td></tr><tr><td rowspan="2">MViT [13]</td><td rowspan="2">✓</td><td>Mean</td><td>0.40</td><td>0.65</td><td>0.38</td></tr><tr><td>Max</td><td>0.47</td><td>0.69</td><td>0.43</td></tr><tr><td rowspan="2">MatchVision</td><td rowspan="2">×</td><td>Mean</td><td>0.44</td><td>0.53</td><td>0.58</td></tr><tr><td>Max</td><td>0.35</td><td>0.70</td><td>0.46</td></tr></table>
+
+InternVideo), underscoring the necessity of building specialized models for soccer understanding; (ii) almost all visual encoders, across all training settings, benefit from SoccerReplay-1988, emphasizing the value of constructing large-scale, high-quality data for soccer understanding. Next, we will delve into each task to discuss the results.
+
+Event Classification. With identical training strategies and data, MatchVision considerably outperforms other methods in classification accuracy, demonstrating the superiority of its architecture, which effectively leverages spatiotemporal features within soccer videos. For example, MatchVision achieves a Top-1 accuracy of  $82.5\%$ , significantly surpassing SigLIP's  $55.9\%$  under the same training conditions. Moreover, models trained via supervised classification excel others, primarily because the pre-training task shares the same objectives as the downstream event classification task.
+
+Commentary Generation. Visual encoders trained with visual-language contrastive learning exhibit better commentary generation performance than those trained with supervised classification, as this strategy better captures correlations between visual and textual features. Additionally, while MatchVision trained solely on SoccerNet slightly underperforms Baidu [61], incorporating SoccerReplay-1988 enables it to outperform on most metrics. This demonstrates that MatchVision can take advantage of large-scale datasets. Finally, a hybrid training approach, starting with supervised classification followed by visual-language contrastive learning, enables MatchVision to achieve optimal performance. This indicates that learning coarse-grained tasks such as classification provides a foundation for fine-grained tasks like commentary generation, and fully leveraging data unlocks the potential of soccer understanding.
+
+Foul Recognition. As demonstrated in Table 3, MatchVision achieves performance comparable to jointly finetuned state-of-the-art methods in foul recognition, even with a frozen visual encoder. This highlights that MatchVision effectively learns substantial knowledge from large-scale soc
+
+Table 3. Quantitative Results on Multi-view Foul Recognition. Our frozen MatchVision encoder can achieve comparable performance with other jointly finetuned visual encoders.  
+
+<table><tr><td colspan="3">Pretrain</td><td colspan="3">Classification(%)</td></tr><tr><td>Sup.</td><td>Contra.</td><td>SR</td><td>Acc.@1</td><td>Acc.@3</td><td>Acc.@5</td></tr><tr><td>✓</td><td>X</td><td>X</td><td>62.67</td><td>83.00</td><td>89.81</td></tr><tr><td>✓</td><td>X</td><td>✓</td><td>68.03</td><td>86.90</td><td>92.38</td></tr><tr><td>X</td><td>✓</td><td>X</td><td>46.97</td><td>75.53</td><td>85.85</td></tr><tr><td>X</td><td>✓</td><td>✓</td><td>57.41</td><td>83.13</td><td>91.00</td></tr><tr><td>✓</td><td>✓</td><td>X</td><td>56.86</td><td>80.30</td><td>88.09</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>63.59</td><td>85.21</td><td>91.63</td></tr></table>
+
+Table 4. Ablations on Event Classification. We explore the impact of various training settings of our MatchVision encoder on the SoccerReplay-test benchmark. Here, Sup., Contra., and SR refer to supervised classification, visual-language contrastive learning, and the SoccerReplay-1988 dataset, respectively.
+
+cer data and adapts seamlessly to downstream tasks. Comparisons with additional baselines from the SoccerNet foul recognition challenges [9] are provided in the Appendix.
+
+# 5.3. Ablation Studies
+
+We conduct ablation experiments on event classification and commentary generation using our SoccerReplay-test benchmark. These experiments validate the effectiveness of our proposed dataset and model, while establishing a baseline for future evaluations on this benchmark.
+
+Event Classification. We evaluate event classification on 300 matches from our SoccerReplay-test benchmark using the MatchVision visual encoder pretrained with various strategies. Features are extracted by MatchVision and processed with a learnable aggregation layer and a linear classifier. The default training set is our curated SoccerNet-pro. As shown in Table 4, integrating SoccerReplay-1988 for training results in significant performance improvements across all pretraining strategies, yielding the significance of our dataset. Additionally, supervised classification outperforms visual-language contrastive learning and hybrid pretraining. This is due to its closer alignment with downstream event classification task, and the scale of event annotations is far larger than that of textual commentaries, further confirming the substantial benefits of data scaling for boosting soccer understanding.
+
+Commentary Generation. With the pretrained MatchVision encoder, we train the commentary generation head on the MatchTime [40] and SoccerReplay-1988 datasets using various training strategies. By default, only the Perceiver [25] aggregation layer and projection layer within the head are trained. For joint training with the LLM decoder, considering computational costs, we incorporate LoRA [24] layers while freezing the original LLM layers. As shown in Table 5, incorporating SoccerReplay-1988 significantly improves performance on all metrics, confirming substantial advantages of our proposed dataset. This performance
+
+![](images/9d80294ea391f74578532a7a932bfd89ac9a197638371fc49f6dce60d8272ed9.jpg)  
+Soccer Videos
+
+![](images/9da4b340fa5e701f874da4eab9b7b914104a894fefb77e3bfec73dbc28c6aaee.jpg)  
+Events
+
+![](images/dfe11a125f3ab53dc3d4ff9c58672ea87bd1c11d4c1cb608159d5405b50da324.jpg)  
+Commentaries
+
+![](images/8cdd4dbbecb8a9ec6e3fdaac06188a5bdb8d9593f88138abc2c3f127dc7a9875.jpg)
+
+![](images/68732df897e97929bec2c06a6cd4a0e00c3c827ef4c30cf5c5ca0c4f68c81307.jpg)
+
+![](images/fc9c2bb97514f72b61de963098cb78a266efe3146a378467055940157f707b37.jpg)
+
+![](images/52ed70c15dc4709676df0e4b456ec14925d160a95ba8aee8054e3fadab030edd.jpg)
+
+![](images/8276801c2dc956cf715cbb6e35392feb9ec6c5c1e35246581e3815e139c44262.jpg)
+
+![](images/516894ea4acf31fed702636a2327972b70c6c0686e548ec9008c7942bb866031.jpg)
+
+![](images/270fd650638147ed4738f17a034c436bce1a5e856c8100fe1265415bc8ef1e7e.jpg)
+
+![](images/b0de7aef3c8941434bfbb3eea98b327a676dacfc6f7f83d3791ac06e9b00f0f7.jpg)
+
+![](images/a380b47f26580073f501fca6a80af2d27b0dff7e9f36d75a2471d72d0c93d535.jpg)
+
+![](images/3eefcf41d5f78e47d7d847d40c2c7378f9003d67dcc7f9abc8a12c41c3846a2e.jpg)
+
+![](images/3b4d60d20742533f20aab26bf153e9159875fb7e3fc21dc2d1ab6a2e9a7e5be3.jpg)
+
+![](images/6bdbf1b0e4e1e262d7f85f385ddf1be5a1f0c0e018b967e3eade70b46797fd21.jpg)
+
+![](images/c4b3b98a4be399f96a73b58101fd17cc5b827266907bce570f39aa35c3e2d7ff.jpg)  
+Figure 4. Qualitative Results for Event Classification and Commentary Generation. Here, "w/o SR" and "w/ SR" indicate models trained without and with the SoccerReplay-1988 dataset, respectively. Incorporating SoccerReplay-1988 improves event classification accuracy. Moreover, this enriched training data enables the model to demonstrate several advantages in commentary generation: (a) more detailed descriptions, (b) greater linguistic variety, (c) higher event depiction accuracy, (d) better adherence to updated rules, and (e) improved specificity in scenario response.
+
+![](images/9b8addab99c97babf3f33b9d8f1d4cb75828a7be35583514c63ea2e65e97bcee.jpg)
+
+![](images/39aeeb1cf36f84a1d0fe1d9adf45176e87fd7b4d6118c5b90ca344994a7a1b2b.jpg)
+
+<table><tr><td colspan="2">Trainable</td><td colspan="5">Commentary Metrics</td></tr><tr><td>V</td><td>L</td><td>B@1</td><td>B@4</td><td>M</td><td>R-L</td><td>C</td></tr><tr><td colspan="7">Trained on MatchTime</td></tr><tr><td>X</td><td>X</td><td>21.65</td><td>3.27</td><td>21.02</td><td>17.79</td><td>12.90</td></tr><tr><td>✓</td><td>X</td><td>27.62</td><td>7.02</td><td>24.03</td><td>23.51</td><td>30.77</td></tr><tr><td>X</td><td>✓</td><td>27.04</td><td>6.41</td><td>24.15</td><td>23.88</td><td>31.91</td></tr><tr><td>✓</td><td>✓</td><td>27.49</td><td>6.96</td><td>24.50</td><td>23.33</td><td>30.81</td></tr><tr><td colspan="7">Trained on MatchTime &amp; SoccerReplay-1988</td></tr><tr><td>X</td><td>X</td><td>24.17</td><td>4.09</td><td>20.51</td><td>20.70</td><td>15.70</td></tr><tr><td>✓</td><td>X</td><td>28.98</td><td>8.39</td><td>24.45</td><td>25.35</td><td>45.85</td></tr><tr><td>X</td><td>✓</td><td>27.54</td><td>7.76</td><td>24.50</td><td>24.70</td><td>42.79</td></tr><tr><td>✓</td><td>✓</td><td>29.21</td><td>8.22</td><td>25.25</td><td>25.54</td><td>43.18</td></tr></table>
+
+Table 5. Ablations on Commentary Generation. We investigate the impact of different training strategies and datasets on MatchVision using the SoccerReplay-test benchmark. 'V' and 'L' denote the visual encoder and the LLM decoder, respectively.
+
+gap also reflects the challenges of our established benchmark, which features diverse vocabulary, richer semantics, and updated soccer rules. Additionally, jointly finetuning the visual encoder and the LLM decoder provides a feasible approach for further improvements.
+
+# 5.4. Qualitative Comparisons
+
+As depicted in Figure 4, we present qualitative results of MatchVision on the SoccerReplay-test benchmark, comparing models pretrained with and without SoccerReplay-1988. For event classification, incorporating our data improves accuracy, and even in misclassified cases, the results
+
+remain contextually relevant. For commentary generation, hybrid training on SoccerReplay-1988 enables MatchVision to produce richer, more detailed textual commentary, reflecting a deeper understanding of soccer dynamics. More qualitative results are available in the Appendix.
+
+# 6. Conclusion
+
+In this paper, we establish a unified, scalable multi-modal framework for soccer understanding. Specifically, we introduce SoccerReplay-1988, the largest and most comprehensive soccer video dataset to date, annotated by an automated curation pipeline. This provides a solid foundation for developing soccer understanding models and serves as a more challenging benchmark. Built upon this, we develop MatchVision, an advanced soccer-specific visual encoder, which effectively leverages spatiotemporal information within soccer videos and can be applied to various tasks such as event classification and commentary generation. Extensive experiments demonstrate the superiority of our model, with MatchVision achieving state-of-the-art performance on both existing benchmarks and our newly established one. We believe this work will set a viable, universal paradigm for future research in sports understanding.
+
+# Acknowledgments
+
+This work is funded by National Key R&D Program of China (No.2022ZD0161400).
+
+# References
+
+[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. 2  
+[2] 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, pages 65-72, 2005. 6  
+[3] 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, page 4, 2021. 2, 4  
+[4] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6  
+[5] 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, and Sergey Tulyakov. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2  
+[6] Anthony Cioppa, Adrien Deliege, Floriane Magera, Silvio Giancola, Olivier Barnich, Bernard Ghanem, and Marc Van Droogenbroeck. Camera calibration and player localization in soccernet-v2 and investigation of their representations for action spotting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 4537-4546, 2021. 2  
+[7] Anthony Cioppa, Adrien Deliege, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. Scaling up soccer-net with multi-view spatial localization and re-identification. Scientific data, 9(1):355, 2022. 1, 2  
+[8] Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3491-3502, 2022. 2  
+[9] Anthony Cioppa, Silvio Giancola, Vladimir Somers, Victor Joos, Floriane Magera, Jan Held, Seyed Abolfazl Ghasemzadeh, Xin Zhou, Karolina Seweryn, Mateusz Kowalczyk, Zuzanna Mroz, Szymon Lukasik, Michal Halon, Hassan Mkhallati, Adrien Deliege, Carlos Hinojosa, Karen Sanchez, Amir M. Mansourian, Pierre Miralles, Olivier Barnich, Christophe De Vleeschouwer, Alexandre Alahi, Bernard Ghanem, Marc Van Droogenbroeck, Adam Gorski, Albert Clapes, Andrei Boiarov, Anton Afanasiev, Artur Xarles, Atom Scott, ByoungKwon Lim, Calvin Yeung, Cristian Gonzalez, Dominic Rufenacht, Enzo Pacilio, Fabian Deuser, Faisal Sami Altawijri, Francisco Cachon, HanKyul Kim, Haobo Wang, Hyeonmin Choe, Hyunwoo J Kim, Il-Min
+
+Kim, Jae-Mo Kang, Jamshid Tursunboev, Jian Yang, Jihwan Hong, Jimin Lee, Jing Zhang, Junseok Lee, Kexin Zhang, Konrad Habel, Licheng Jiao, Linyi Li, Marc Gutierrez-Perez, Marcelo Ortega, Menglong Li, Milosz Lopatto, Nikita Kasatkin, Nikolay Nemtsev, Norbert Oswald, Oleg Udin, Pavel Kononov, Pei Geng, Saad Ghazai Alotaibi, Sehyung Kim, Sergei Ulasen, Sergio Escalera, Shanshan Zhang, Shuyuan Yang, Sunghwan Moon, Thomas B. Moeslund, Vasyl Shandyba, Vladimir Golovkin, Wei Dai, WonTaek Chung, Xinyu Liu, Yongqiang Zhu, Youngseo Kim, Yuan Li, Yuting Yang, Yuxuan Xiao, Zehua Cheng, and Zhihao Li. Soccernet 2024 challenges results. arXiv preprint arXiv:2409.10587, 2024. 2, 7  
+[10] Adrien Deliege, Anthony Cioppa, Silvio Giancola, Meisam J Seikavandi, Jacob V Dueholm, Kamal Nasrollahi, Bernard Ghanem, Thomas B Moeslund, and Marc Van Droogenbroeck. Soccernet-v2: A dataset and benchmarks for holistic understanding of broadcast soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 4508-4519, 2021. 1, 2, 3, 6  
+[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations, 2021. 4  
+[12] 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 preprint arXiv:2407.21783, 2024. 3, 5  
+[13] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6824-6835, 2021. 6, 7  
+[14] Sushant Gautam, Mehdi Houshmand Sarkhoosh, Jan Held, Cise Midoglu, Anthony Cioppa, Silvio Giancola, Vajira Thambawita, Michael A Riegler, Pål Halvorsen, and Mubarak Shah. Soccernet-echoes: A soccer game audio commentary dataset. arXiv preprint arXiv:2405.07354, 2024. 2  
+[15] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. Soccernet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 1711-1721, 2018. 1, 2, 3  
+[16] Xiaofan Gu, Xinwei Xue, and Feng Wang. Fine-grained action recognition on a novel basketball dataset. In International Conference on Acoustics, Speech, and Signal Processing, pages 2563-2567, 2020. 2  
+[17] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2906-2916, 2022. 2  
+[18] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad: Movie description in
+
+context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2  
+[19] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel-who, when, and what in movie audio description. In Proceedings of the International Conference on Computer Vision, 2023.  
+[20] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad iii: The prequel - back to the pixels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2  
+[21] 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, 7  
+[22] Jan Held, Anthony Cioppa, Silvio Giancola, Abdullah Hamdi, Bernard Ghanem, and Marc Van Droogenbroeck. Vars: Video assistant referee system for automated soccer decision making from multiple views. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 5085-5096, 2023. 2, 5, 6  
+[23] Jan Held, Hani Itani, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. X-vars: Introducing explainability in football refereeing with multimodal large language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3267-3279, 2024. 2, 6  
+[24] 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. In Proceedings of the International Conference on Learning Representations, 2022. 7  
+[25] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In Proceedings of the International Conference on Machine Learning, pages 4651-4664, 2021. 5, 7  
+[26] Haopeng Li, Andong Deng, Qiuhong Ke, Jun Liu, Hossein Rahmani, Yulan Guo, Bernt Schiele, and Chen Chen. Sports-qa: A large-scale video question answering benchmark for complex and professional sports. arXiv preprint arXiv:2401.01505, 2024. 2  
+[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the International Conference on Machine Learning, pages 12888-12900, 2022. 2  
+[28] 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 Machine Learning, pages 19730–19742, 2023. 2  
+[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2
+
+[30] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In Proceedings of the European Conference on Computer Vision, 2024. 2  
+[31] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhang, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. In Proceedings of the European Conference on Computer Vision, 2024. 2  
+[32] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, 2004. 6  
+[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proceedings of the International Conference on Learning Representations, 2019. 5  
+[34] Hassan Mkhallati, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. Socernetc caption: Dense video captioning for soccer broadcasts commentaries. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 5074-5085, 2023. 1, 2, 3  
+[35] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Association for Computational Linguistics, pages 311-318, 2002. 6  
+[36] Ji Qi, Jifan Yu, Teng Tu, Kunyu Gao, Yifan Xu, Xinyu Guan, Xiaozhi Wang, Bin Xu, Lei Hou, Juanzi Li, et al. Goal: A challenging knowledge-grounded video captioning benchmark for real-time soccer commentary generation. In Proceedings of the ACM International Conference on Information and Knowledge Management, pages 5391-5395, 2023. 1, 2, 3  
+[37] Mengshi Qi, Yunhong Wang, Annan Li, and Jiebo Luo. Sports video captioning via attentive motion representation and group relationship modeling. IEEE Transactions on Circuits and Systems for Video Technology, 2019. 2  
+[38] 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 Proceedings of the International Conference on Machine Learning, 2021. 2, 6  
+[39] Prem N Ramkumar, Bryan C Luu, Heather S Haeberle, Jaret M Karnuta, Benedict U Nwachukwu, and Riley J Williams. Sports medicine and artificial intelligence: a primer. The American Journal of Sports Medicine, 50(4): 1166-1174, 2022. 2  
+[40] Jiayuan Rao, Haoning Wu, Chang Liu, Yanfeng Wang, and Weidi Xie. Matchtime: Towards automatic soccer game commentary generation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2024. 1, 2, 3, 5, 6, 7  
+[41] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2616-2625, 2020. 2  
+[42] Yan Shu, Zheng Liu, Peitian Zhang, Minghao Qin, Junjie Zhou, Zhengyang Liang, Tiejun Huang, and Bo Zhao.
+
+Video-xl: Extra-long vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 2  
+[43] Vladimir Somers, Victor Joos, Anthony Cioppa, Silvio Giancola, Seyed Abolfazl Ghasemzadeh, Floriane Magera, Baptiste Standaert, Amir M Mansourian, Xin Zhou, Shohreh Kasaei, et al. Soccernet game state reconstruction: End-to-end athlete tracking and identification on a minimap. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3293-3305, 2024. 2  
+[44] Graham Thomas, Rikke Gade, Thomas B Moeslund, Peter Carr, and Adrian Hilton. Computer vision for sports: Current applications and research topics. Computer Vision and Image Understanding, 159:3-18, 2017. 2  
+[45] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the International Conference on Computer Vision, pages 4489-4497, 2015. 6  
+[46] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 6, 7  
+[47] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4566-4575, 2015. 6  
+[48] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 6  
+[49] Zhe Wang, Petar Velicković, Daniel Hennes, Nenad Tomasev, Laurel Prince, Michael Kaisers, Yoram Bachrach, Romuald Elie, Li Kevin Wenliang, Federico Piccinini, et al. Tacticai: an ai assistant for football tactics. Nature Communications, 15(1):1-13, 2024. 1, 2  
+[50] Dekun Wu, He Zhao, Xingce Bao, and Richard P Wildes. Sports video analysis on large-scale data. In Proceedings of the European Conference on Computer Vision, 2022. 2  
+[51] Zeyu Xi, Ge Shi, Xuefen Li, Junchi Yan, Zun Li, Lifang Wu, Zilin Liu, and Liang Wang. A simple yet effective knowledge guided method for entity-aware video captioning on a basketball benchmark. Neurocomputing, 2025. 2  
+[52] Zeyu Xi, Ge Shi, Haoying Sun, Bowen Zhang, Shuyi Li, and Lifang Wu. Eika: Explicit & implicit knowledge-augmented network for entity-aware sports video captioning. Expert Systems with Applications, 2025. 2  
+[53] Haotian Xia, Zhengbang Yang, Yuqing Wang, Rhys Tracy, Yun Zhao, Dongdong Huang, Zezhi Chen, Yan Zhu, Yuanfang Wang, and Weining Shen. Sportqa: A benchmark for sports understanding in large language models. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics, 2024. 2  
+[54] Haotian Xia, Zhengbang Yang, Junbo Zou, Rhys Tracy, Yuqing Wang, Chi Lu, Christopher Lai, Yanjun He, Xun
+
+Shao, Zhuoqing Xie, et al. Sportu: A comprehensive sports understanding benchmark for multimodal large language models. In Proceedings of the International Conference on Learning Representations, 2025. 2  
+[55] Jinglin Xu, Yongming Rao, Xumin Yu, Guangyi Chen, Jie Zhou, and Jiwen Lu. Finediving: A fine-grained dataset for procedure-aware action quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2949-2958, 2022. 2  
+[56] 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 Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2  
+[57] Huanyu Yu, Shuo Cheng, Bingbing Ni, Minsi Wang, Jian Zhang, and Xiaokang Yang. Fine-grained video captioning for sports narrative. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2  
+[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the International Conference on Computer Vision, pages 11975-11986, 2023. 2, 5, 6  
+[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2023. 2  
+[60] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2  
+[61] Xin Zhou, Le Kang, Zhiyu Cheng, Bo He, and Jingyu Xin. Feature combination meets attention: Baidu soccer embeddings and transformer based temporal detection. arXiv preprint arXiv:2106.14447, 2021. 2, 6, 7  
+[62] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2

2025/Towards Universal Soccer Video Understanding/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:662e11fa11125dd224724691dcbdc952fd4f3369d2148b293cd50500324c8547
+size 721991

2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/e496d86c-0d1a-42d1-b098-7f552a70c4ff_content_list.json

@@ -0,0 +1,1689 @@
+[
+    {
+        "type": "text",
+        "text": "Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics: Physics-Grounded Anomaly Detection",
+        "text_level": 1,
+        "bbox": [
+            94,
+            128,
+            908,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Wenqiao Li $^{1*}$  Yao Gu $^{1*}$  Xintao Chen $^{1*}$  Xiaohao Xu $^{2\\dagger}$  Ming Hu $^{3}$  Xiaonan Huang $^{2}$  Yingna Wu $^{1\\dagger}$ $^{1}$ ShanghaiTech University  $^{2}$ University of Michigan, Ann Arbor  $^{3}$ Monash University",
+        "bbox": [
+            99,
+            200,
+            944,
+            241
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/75d6ea3cc4000b98b604cf49e2eedc38b5b4f3a012591964ab21612bc71ce942.jpg",
+        "image_caption": [
+            "Figure 1. Towards visual discrimination of physical dynamics in real-world industrial object anomaly detection. We illustrate objects, interactions, and time-sequenced videos from the Physics-Grounded Anomaly Detection dataset: (a) Object; (b) Interaction: Applied actions shown with directional arrows; (c) Video with Physical Dynamics: Temporal sequences showing normal and abnormal states, highlighting anomalies like leaks, misalignments, and cracks. By focusing on the dynamic behaviors of complex objects, we enhance understanding of interactions and failure modes in real-world settings, where both structure and motion contribute to anomaly detection."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            253,
+            535,
+            542
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/406ce781989bb97be996195c92fcaa5cde7a373227f124df11e22d77c6ab559c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            540,
+            250,
+            906,
+            542
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            248,
+            640,
+            326,
+            655
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Humans detect real-world object anomalies by perceiving, interacting, and reasoning based on object-conditioned physical knowledge. The long-term goal of Industrial Anomaly Detection (IAD) is to enable machines to autonomously replicate this skill. However, current IAD algorithms are largely developed and tested on static, semantically simple datasets, which diverge from real-world scenarios where physical understanding and reasoning are essential. To bridge this gap, we introduce the Physics Anomaly Detection (Phys-AD) dataset, the first large-scale, real-world, physics-grounded video dataset for industrial anomaly detection. Collected using a real robot arm and motor, Phys-AD provides a diverse",
+        "bbox": [
+            88,
+            674,
+            485,
+            856
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "set of dynamic, semantically rich scenarios. The dataset includes more than 6400 videos across 22 real-world object categories, interacting with robot arms and motors, and exhibits 47 types of anomalies. Anomaly detection in Phys-AD requires visual reasoning, combining both physical knowledge and video content to determine object abnormality. We benchmark state-of-the-art anomaly detection methods under three settings: unsupervised AD, weakly-supervised AD, and video-understanding AD, highlighting their limitations in handling physics-grounded anomalies. Additionally, we introduce the Physics Anomaly Explanation (PAEval) metric, designed to assess the ability of visual-language foundation models to not only detect anomalies but also provide accurate explanations for their underlying physical causes. Our project is available at https://guyao2023.github.io/Phys-AD/.",
+        "bbox": [
+            511,
+            641,
+            910,
+            869
+        ],
+        "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": "*Equally contribute to this work.",
+        "bbox": [
+            107,
+            875,
+            284,
+            887
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "† Corresponding authors",
+        "bbox": [
+            109,
+            887,
+            238,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "30409",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/fc4dc8c941e0683bab05b70a96d80230028a00681e4d6d0074810515380bb537.jpg",
+        "image_caption": [
+            "Figure 2. Human-like decision-making process for physics-grounded object anomaly detection. We illustrate the sequential approach of a human-like agent for evaluating an object's normality. First, the agent perceives relevant physical attributes (e.g., plastic and elastic), then interacts by performing a physical action (e.g., squeezing), and finally reasons based on the vision feedback and attributes changes (e.g., surface shape change) to determine whether the object is normal or anomalous. This mirrors a human's natural process of reasoning over physics in objects."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            101,
+            90,
+            478,
+            311
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            91,
+            467,
+            222,
+            483
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Industrial anomaly detection (IAD) is a critical subfield in computer vision and industrial automation, aiming to identify defects or irregularities in products during manufacturing. As shown in Fig. 2, the ultimate vision is to create autonomous systems that not only perceive but also interact with and reason about objects to discriminate between anomalies and normal states, integrating complex physical principles to detect anomalies in dynamic, real-world scenarios. For example, a human inspecting a water bottle for anomalies wouldn't rely solely on visual observation; they might rotate or invert the bottle, using physical interactions and cues, such as noticing a loose cap or an irregular internal flow, to detect issues.",
+        "bbox": [
+            89,
+            492,
+            483,
+            686
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Central to advancing IAD [9] is the availability of high-quality datasets that bridge the gap between academic research and industrial needs. Datasets like MVTec-AD [4], MPDD [21], and VisA [65] have played foundational roles, enabling algorithm development for image-based anomaly detection and bringing IAD to the forefront of computer vision research. While these datasets have significantly advanced single-image anomaly detection, recent datasets, e.g., MVTec-3D [7], Real3D [29], and Anomaly-ShapeNet [27], have extended IAD to 3D, aligning research more closely with the needs of complex real-world industrial settings.",
+        "bbox": [
+            89,
+            688,
+            483,
+            854
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Yet, as factories increasingly rely on robotic arms and automated systems to perform sophisticated inspections, the limitations of current IAD datasets become apparent.",
+        "bbox": [
+            89,
+            854,
+            483,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Existing benchmarks focus on static, semantically simple environments, overlooking the physical priors and interactive reasoning required in real-world industrial contexts. This gap highlights a growing need for datasets that not only reflect real-world physical constraints but also challenge models to reason dynamically about anomalies.",
+        "bbox": [
+            511,
+            90,
+            903,
+            181
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To bridge this gap, as shown in Fig. 1, we introduce the Physics Anomaly Detection (Phys-AD) dataset, the first large-scale, physics-grounded video dataset for industrial anomaly detection. Phys-AD features over 6,400 videos of 22 categories and 49 objects interacting with robotic arms and motors, capturing 47 types of anomalies that require visual reasoning informed by physical knowledge. The short video clips in the dataset range from 60 to 240 frames in length and are filmed in real-time industrial environment, fully capturing the interaction process between robotic arm or motor and industrial objects. Additionally, to ensure our dataset meets industrial demands and matches the complexity and diversity of the real physical world, we selected industrial objects of different physical qualities, various interaction methods, and anomalies that reflect diverse physical principles and require different reasoning process. Specifically, we selected 22 object categories spanning across metals, plastics, fluid, amorphous substances and articulated objects with diverse appearances. For interaction, we use mechanical grippers, robotic arms, and motors, incorporating various interaction modes such as pressing, rotating, squeezing, and driving to handle different types of objects.",
+        "bbox": [
+            511,
+            184,
+            906,
+            516
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We benchmark state-of-the-art anomaly detection methods in three key configurations: unsupervised anomaly detection, weakly-supervised anomaly detection, and video-understanding based anomaly detection. Our findings reveal critical gaps in their ability to handle the complexities of physics-grounded scenarios, where anomalies often arise from dynamic, interdependent interactions. To advance the field, we also introduce the Physical Anomaly Explanation (PAEval) metric, designed to assess both detection performance and a model's ability to explain anomalies by identifying underlying physical causes. Furthermore, our benchmark reveals the fragility of existing methods in tackling these challenging conditions, underscoring the need for approaches that better understand object dynamics and temporal coherence in anomaly detection.",
+        "bbox": [
+            511,
+            518,
+            908,
+            744
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions can be summarized as followings:",
+        "text_level": 1,
+        "bbox": [
+            531,
+            747,
+            901,
+            762
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We introduce a novel task of detecting physical-based industrial anomalies in real-world that involves perception, physical and visual reasoning.",
+            "- We present Phys-AD, the first large-scale, physics-grounded video dataset specifically designed for industrial anomaly detection in real world, containing objects of different physical qualities, multiple interaction methods and various physical reasoning process.",
+            "- We benchmarking the anomaly detection and reasoning"
+        ],
+        "bbox": [
+            511,
+            763,
+            906,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "30410",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/2affb83bc3d5314d7dc6b2b25346b44074ea92e0f46ef8701d79e44a343e635b.jpg",
+        "table_caption": [
+            "Table 1. Comparison of Phys-AD with existing industrial anomaly detection datasets. Our Phys-AD dataset is the first to consider complex objects with physical dynamics. 'Syn', 'IR', 'D', and 'PC' denote Synthetic, Infrared, Depth, and Point Cloud, respectively. #Anomaly indicates the number of anomaly types."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Dataset</td><td rowspan=\"2\">Year</td><td rowspan=\"2\">Type</td><td rowspan=\"2\">Modality</td><td colspan=\"3\">Sample Statistics</td></tr><tr><td>Class #Anomaly</td><td colspan=\"2\">Physics</td></tr><tr><td>MVTec-AD [3]</td><td>2019</td><td>Real</td><td>RGB</td><td>15</td><td>-</td><td>X</td></tr><tr><td>BTAD [38]</td><td>2021</td><td>Real</td><td>RGB</td><td>3</td><td>3</td><td>X</td></tr><tr><td>MPDD [21]</td><td>2021</td><td>Real</td><td>RGB</td><td>6</td><td>8</td><td>X</td></tr><tr><td>VisA [66]</td><td>2021</td><td>Real</td><td>RGB</td><td>12</td><td>-</td><td>X</td></tr><tr><td>MVTec LOCO-AD [5]</td><td>2022</td><td>Real</td><td>RGB</td><td>5</td><td>-</td><td>X</td></tr><tr><td>MAD [64]</td><td>2023</td><td>Syn+Real</td><td>RGB</td><td>20</td><td>3</td><td>X</td></tr><tr><td>LOCO-Annotations [62]</td><td>2024</td><td>Real</td><td>RGB</td><td>5</td><td>5</td><td>X</td></tr><tr><td>Real-IAD [46]</td><td>2024</td><td>Real</td><td>RGB</td><td>30</td><td>8</td><td>X</td></tr><tr><td>GDXray [37]</td><td>2015</td><td>Real</td><td>X-ray</td><td>5</td><td>15</td><td>X</td></tr><tr><td>PVEL-AD [43]</td><td>2023</td><td>Real</td><td>IR</td><td>1</td><td>10</td><td>X</td></tr><tr><td>MVTec3D-AD [6]</td><td>2021</td><td>Real</td><td>RGB-D</td><td>10</td><td>3-5</td><td>X</td></tr><tr><td>Eyecandies [8]</td><td>2022</td><td>Syn</td><td>RGB-D</td><td>10</td><td>3</td><td>X</td></tr><tr><td>Real3D-AD [30]</td><td>2023</td><td>Real</td><td>PC</td><td>12</td><td>2</td><td>X</td></tr><tr><td>Anomaly-ShapeNet [27]</td><td>2024</td><td>Syn</td><td>PC</td><td>40</td><td>6</td><td>X</td></tr><tr><td>Phys-AD (Ours)</td><td>2024</td><td>Real</td><td>RGB</td><td>49</td><td>47</td><td>✓</td></tr></table>",
+        "bbox": [
+            91,
+            161,
+            483,
+            392
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "performance of popular video AD methods and Visual Language Foundation Models on the Phys-AD dataset in several settings, establishing a practical and challenging benchmark to promote the development of the physics-related anomaly detection field.",
+        "bbox": [
+            102,
+            419,
+            483,
+            494
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2. Related work",
+        "text_level": 1,
+        "bbox": [
+            89,
+            511,
+            227,
+            527
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Industrial anomaly detection datasets. Existing datasets primarily focus on static, semantically simple scenarios, deviate significantly from real world where physical understanding and reasoning are essential. Datasets like MVTec-AD [3], BTAD [38], MPDD [21], and VisA [66] focus on surface level image anomaly detection with one single-view RGB image, limiting their effectiveness in capturing holistic object structures. MVTec-LOCO-AD [5], focusing on structure and local information in industries, is limited by its relatively simple and constrained data content. While MVTec3D-AD [6] and Eyecandies [8] incorporate depth data, they remain static and single-view image anomaly detection, neglecting object level information. To explore object level anomaly detection, multi-view IAD datasets like MAD [64] and Real-IAD [46], point cloud IAD datasets like Real3D-AD [30] and Anomaly-ShapeNet [27], offering richer visual and geometric cues, but are still limited to static objects without dynamic interaction or reasoning. In summary, current industrial anomaly detection datasets focus on relatively simple and static anomaly detection scenarios, lacking complex physical rules, dynamic interactions, and visual reasoning requirements. Therefore, existing IAD datasets generally applied to limited industrial scenarios and there is no IAD dataset could meet the demands of detecting",
+        "bbox": [
+            89,
+            537,
+            485,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "complex anomalies in real world which need physical priors and reasoning.",
+        "bbox": [
+            511,
+            90,
+            906,
+            121
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Video anomaly detection. Deep learning methods [18, 35, 54, 61] now dominate Video Anomaly Detection (VAD), categorized into unsupervised, weakly-supervised, and fully-supervised approaches. Unsupervised methods learn normal patterns via reconstruction [16, 18, 52], prediction [32], or hybrids [33], while some methods [45, 55] train with both unlabeled normal and abnormal data. Weakly-supervised methods[14, 49, 60, 63] use video-level or glance-based annotations, and fully-supervised methods[25, 31] remain rare due to costly frame-level labeling. Visual-language models like CLIP [42] have recently been applied to enhance anomaly detection [24, 41, 51], focusing on semantic anomalies. Open-vocabulary VAD [50] and prompt-based anomaly scoring [56] leverage LLMs [53, 61], but performance relies heavily on the base models, often lacking domain-specific tuning. However, existing video anomaly detection algorithms lack the capability to handle complex industrial anomaly detection scenarios and understand physical rules. This gap highlights the need for models that can capture dynamic behaviors and physical laws in industrial environments, as addressed by Phys-AD, which targets industrial anomalies in objects of various physical properties.",
+        "bbox": [
+            511,
+            125,
+            908,
+            458
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Visual reasoning. Visual reasoning is a critical task in computer vision, aiming to enable machines to interpret perceptual information like humans. Several visual reasoning tasks have been proposed to evaluate reasoning capabilities, including Visual Question Answering (VQA), 2D puzzles, and physical dynamics prediction. In VQA, agents are required to combine natural language and visual cues to answer questions [1, 17, 19, 23]. For 2D puzzles, tasks involve discovering relationships among visual elements and making inferences [22, 26, 39, 57, 58]. Physical dynamics prediction tasks require machines to perceive and reason about physical interactions [2, 12, 13, 20]. In contrast to these work, we introduce the first benchmark featuring real-world industrial objects with dynamic physical properties, focusing on distinguishing diverse dynamics through vision.",
+        "bbox": [
+            511,
+            460,
+            910,
+            690
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Dataset: Phys-AD",
+        "text_level": 1,
+        "bbox": [
+            511,
+            702,
+            689,
+            718
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Object Preparation and Interaction Selection",
+        "text_level": 1,
+        "bbox": [
+            511,
+            727,
+            897,
+            743
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We selected 22 categories spanning across various materials, including metals, plastics, amorphous substances and articulated objects, with diverse shapes, sizes, and physical properties. For the physical properties of different objects, we correspondingly select different interaction methods like push, rotate, pull with robotics arms and motors. For instance, we use robotics arms to grab and extruding the deformable objects like basketball to determine whether there are any elastic anomalies or surface defects by the morphological change of the objects. To make our dataset more practical",
+        "bbox": [
+            511,
+            750,
+            908,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "30411",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/d1fb0c57315c70053b2eee90ea39523b6bc62eca99f6e996b4571a08712f0d89.jpg",
+        "image_caption": [
+            "U Disk"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            88,
+            225,
+            145
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/04c3a0c007d031f41abd5c7c98604aad9ab1aaf4129e18d28427e92deee1eb37.jpg",
+        "image_caption": [
+            "Zipper"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            227,
+            88,
+            359,
+            145
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/45bcc2f974a31f7bc5fa856043b21b280954cb11438e7f549aef3542277029ad.jpg",
+        "image_caption": [
+            "Rolling Bearing"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            364,
+            89,
+            496,
+            146
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/b7d51d7075f154f89366e928da49136a71a85cee7477c389e2eb7f9afe4c02c5.jpg",
+        "image_caption": [
+            "Sticky Roller"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            90,
+            632,
+            147
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/138a01ee9acd66f4fd6acc7aa80c8d225c1ae73de2587940ad0378ef6c70897b.jpg",
+        "image_caption": [
+            "Slider"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            637,
+            90,
+            767,
+            146
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/6b60ccaaf439eb96ea39c5a3b9fa413b510be46701455468c04e51538753bdc3.jpg",
+        "image_caption": [
+            "Toothpaste"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            90,
+            903,
+            146
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/0b3e191f36b3ed06901894a1493699ae31532b6433e29d9994adcf9e77502b7d.jpg",
+        "image_caption": [
+            "Door Lock"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            93,
+            164,
+            225,
+            222
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/de166171b1fd2c1d1d14cfe9f71d84aeee36e3408be644fb0740baed93afadf8.jpg",
+        "image_caption": [
+            "Toy Car"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            228,
+            164,
+            359,
+            222
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/90f777c39de74347453e303735bb358ab7dd2c4abc98c699069be3d12fd72216.jpg",
+        "image_caption": [
+            "Clip"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            364,
+            164,
+            496,
+            222
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/1324e3603fe14c0ee8596a894f5b4b71efea4e5df255eef133e2c529b00767b4.jpg",
+        "image_caption": [
+            "Fan"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            165,
+            632,
+            222
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/79a5bcee9d78b800f3c96612ba744acd04d2d65ab5444af43ed720b70edc3bcb.jpg",
+        "image_caption": [
+            "Rubber Band"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            637,
+            165,
+            767,
+            222
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/a93b776c01182baf642c844fd4003427bef1d20f803d1921c7b718ef3b4de6e3.jpg",
+        "image_caption": [
+            "Magnet"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            165,
+            903,
+            223
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/ef6d8ff60375a009ba23164350c9f70fa0b56f37d63ffd49a93da1a6e504579a.jpg",
+        "image_caption": [
+            "Servo"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            244,
+            223,
+            301
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/4d6dc56067eb39e38a2c6d664695e2645359b85f86b75c0ba77d4771ba20a87f.jpg",
+        "image_caption": [
+            "Gear Assembly"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            228,
+            244,
+            359,
+            301
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/143b400b63356474d7a583349339b8c3fd2ca6429bf0a91836a0f9968ad6acbb.jpg",
+        "image_caption": [
+            "Bottled Water"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            364,
+            244,
+            496,
+            301
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/db4e4ea50cb523a34663c3de3d8555feba8fc3e602d608de9919f4a00022692f.jpg",
+        "image_caption": [
+            "Screw"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            244,
+            632,
+            301
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/5d07bbf049c4ab67cb7cefb294f93b1adecd62b7b746b16a2b25c80a1f2688ff.jpg",
+        "image_caption": [
+            "Spherical Bearing"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            637,
+            244,
+            767,
+            301
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/ae3b8ab1d44dca0c371de14f0f6e04fbffda0755c5a46cb2291463b29b4ca39e.jpg",
+        "image_caption": [
+            "Motion"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            792,
+            253,
+            888,
+            276
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/30b520f00ef352421bb17433f5d67fdc49435e0a6a68430c5acb0b5e28c6e3da.jpg",
+        "image_caption": [
+            "Button"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            93,
+            325,
+            223,
+            388
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/d477cbe0731ea4b43372c2b74c2807e6242f0ecfdb1ee561599aa52988b7e88b.jpg",
+        "image_caption": [
+            "Clock"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            228,
+            325,
+            359,
+            388
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/32405e8d76a58bd525059e3dbd12ba4e95ee2af92a48352e8327c8bf0cb5a0a5.jpg",
+        "image_caption": [
+            "Caster Wheel"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            364,
+            325,
+            493,
+            388
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/d020a461c07f5b611e5475dca37bd6451a9ee9fd00ee94a5502173746bfcce9d.jpg",
+        "image_caption": [
+            "Hinge"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            325,
+            632,
+            388
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/5d39dd0d1961330a708b001bbd0af0330e7e2274175ee736aa65e240d83b341e.jpg",
+        "image_caption": [
+            "Ball"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            637,
+            325,
+            769,
+            388
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/ee9640bd724f2776c510eb025a226563c9518fd3ee2fc98b17e081ca1a1b534d.jpg",
+        "image_caption": [
+            "Direction For Interaction",
+            "Figure 3. Interactions for understanding implicit physical laws in the Phys-AD dataset. We showcase various object interactions from the Phys-AD dataset, where different actions (indicated by motion directions) are used to explore and reason about the underlying physical properties and behaviors of each object. The colored arrows indicate the interaction directions and axes, highlighting how physical interactions reveal the implicit physics governing each object."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            325,
+            903,
+            398
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/b3f99f89cbed8591fccb3110e6a970871e8521d23d9bbcbaaea6ed729034642a.jpg",
+        "image_caption": [
+            "Normal: The U Disk rotates freely over time."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            491,
+            493,
+            558
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/6768b2f107bdf1fc48650b414447219c6c19cc384203b8c07f0ee7c7436e2cb8.jpg",
+        "image_caption": [
+            "Abnormal: The U Disk is stuck over time.",
+            "(a) Persistent Anomalies"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            597,
+            493,
+            662
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/fc43f6a365f2a63ec36586f1a0cf725753bce008123e8aed6dd463dcf945a622.jpg",
+        "image_caption": [
+            "Normal: The sticky roller rotates smoothly on the surface."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            491,
+            901,
+            558
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/0fd51d148225da32c20aed5e679cb20d87bc532ca66ea2e17693bcb16094c8c2.jpg",
+        "image_caption": [
+            "Normal: The sticky roller functions properly at the beginning.",
+            "Abnormal: The sticky roller breaks at a certain point.",
+            "(b) Intermittent Anomalies",
+            "Figure 4. Categorization of anomalies based on persistence in the Phys-AD dataset. We show examples of normal and abnormal functioning in common objects, divided into two anomaly types: persistent and intermittent. (a) Persistent anomalies, such as continuous obstruction in the U Disk or permanent malfunction of the Sticky Roller, are visible throughout the operation. (b) In contrast, intermittent anomalies, like occasional jamming of the U Disk or breakage in the Sticky Roller after initial operation, only appear at specific points in time. This classification provides insight into both constant and sporadic failures in object interactions."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            599,
+            901,
+            664
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "and challenging, we introduce 47 distinct defect types, some of which just rely on single frame content and physical rules for anomaly reasoning, while the other part needs to combine the content of the whole video and physical priors together to judge whether the object is abnormal or not. Different objects information with their corresponding interaction methods and anomalies are listed in Table 2. Fig 3 shows the interac",
+        "bbox": [
+            89,
+            782,
+            485,
+            888
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "tions for understanding implicit physical laws in the Phys-AD dataset. Fig 4 provides two representative examples: one where a U Disk requires analyzing the entire video content to determine if there is an anomaly, and another where a sticky roller only requires a few frames from the video to judge whether the object is abnormal.",
+        "bbox": [
+            511,
+            782,
+            906,
+            873
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "30412",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/c3ffd1eb93313cb723ec32e3219be946461890aafac571d6445aa6c70c56801c.jpg",
+        "image_caption": [
+            "Figure 5. Data collection pipeline for the Phys-AD dataset. (a) Manipulation of a toothpaste tube using a UR5 robotic arm. (b) Manipulation of a U Disk and fan via servo and motor."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            89,
+            287,
+            208
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/e4add3a5da0de42ca6b80027e5c1fbdd5fd5651b741bdb4da4c922c84fbf789b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            289,
+            90,
+            480,
+            207
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/cbcf4e193c05768860b9b372ece0ae1bbe405dad03bd5e2829372390fe2dbb53.jpg",
+        "table_caption": [
+            "Table 2. Interaction methods and anomalies of the object categories in our Phys-AD dataset. Motor. means one object driven by motor. Multi. means one object with multiple anomalies."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Category</td><td>Interaction</td><td>Anomaly Types</td></tr><tr><td>Car</td><td>Drag, Slide</td><td>Different wheels stuck</td></tr><tr><td>Fan</td><td>Motor., Rotate</td><td>Stuck, Uneven Rotation, Vibration</td></tr><tr><td>Rolling Bearing</td><td>Motor., Rotate</td><td>Lack of friction</td></tr><tr><td>Spherical Bearing</td><td>Grab, Rotate</td><td>Internal block</td></tr><tr><td>Servo</td><td>Grab, Motor., Rotate</td><td>Angle restricted, Vibration, No calibration</td></tr><tr><td>Clip</td><td>Press</td><td>Unable to press down, Unable to rebound</td></tr><tr><td>U Disk</td><td>Grab, Motor., Rotate</td><td>Cover jam</td></tr><tr><td>Hinge</td><td>Grab, Rotate</td><td>Angle restricted, Shaft off</td></tr><tr><td>Sticky Roller</td><td>Grab, Pull</td><td>Detach, Unable to rotate</td></tr><tr><td>Caster Wheel</td><td>Slide</td><td>Axle axis stuck, Swivel axis stuck</td></tr><tr><td>Screw</td><td>Press, Rotate</td><td>Loosen, Unable to insert</td></tr><tr><td>Lock</td><td>Motor., Rotate</td><td>Latch jam, Loose</td></tr><tr><td>Gear</td><td>Motor., Rotate</td><td>Stuck, Not meshed, Multi.</td></tr><tr><td>Clock</td><td>Motor., Rotate</td><td>Pointer stops, Vibration</td></tr><tr><td>Slide</td><td>Grab, Slide</td><td>Detach, Shedding, Jam</td></tr><tr><td>Zipper</td><td>Grab, Close</td><td>Stuck, Unable to close</td></tr><tr><td>Button</td><td>Press</td><td>No light, Unable to press down, Unable to rebound, Multi.</td></tr><tr><td>Liquid</td><td>Grab, Shake</td><td>Water out, Foreign body</td></tr><tr><td>Rubber Band</td><td>Stretch</td><td>Crack</td></tr><tr><td>Ball</td><td>Pinch, Press</td><td>Insufficient gas, Leakage</td></tr><tr><td>Magnet</td><td>Grab, Press, Move</td><td>Degaussing, Shell detached</td></tr><tr><td>Toothpaste</td><td>Pinch, Press</td><td>Leakage</td></tr></table>",
+        "bbox": [
+            94,
+            319,
+            480,
+            670
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2. Data Collection and Processing",
+        "text_level": 1,
+        "bbox": [
+            89,
+            694,
+            366,
+            710
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Data collection pipeline. Most of the data collection for the Phys-AD dataset is driven by manipulation-guided video sequences, captured using a UR5 robotic arm equipped with an RGB camera (see Fig. 5a). In order to reproduce real industrial scenes, we also include adjustable light sources in our data capture process. The RGB Camera, with a 1080p resolution and a frame rate of 60 FPS, providing high-quality video sequences. Some kinds of the objects like U Disk that are not suitable for the manipulation of robotics arms, are driven by the motor or servo (Fig. 5b). After the data collection, we used video editing software to crop out irrelevant frames and retain the complete interaction process.",
+        "bbox": [
+            89,
+            719,
+            483,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/5c88e1548cab93edd3fe5a31ee082684208e011607747650969dc1fda726e079.jpg",
+        "table_caption": [
+            "Table 3. Phys-AD dataset statistics. We denote the total number of frames and videos for each category as #Images and #Videos. Note that the Train split does not contain anomaly data."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Category</td><td>Train #Frames</td><td>Train #Videos</td><td>Test #Frames</td><td>Test #Videos</td><td>Test #Videos Nor.</td><td>Test #Videos Ab.</td><td>Obj. Types</td><td>Def. Types</td></tr><tr><td>Car</td><td>18,000</td><td>300</td><td>36,000</td><td>600</td><td>150</td><td>450</td><td>5</td><td>3</td></tr><tr><td>Fan</td><td>32,400</td><td>180</td><td>64,800</td><td>360</td><td>90</td><td>270</td><td>3</td><td>3</td></tr><tr><td>Rolling Bearing</td><td>3,600</td><td>60</td><td>3,600</td><td>60</td><td>30</td><td>30</td><td>1</td><td>1</td></tr><tr><td>Spherical Bearing</td><td>3,600</td><td>60</td><td>3,600</td><td>60</td><td>30</td><td>30</td><td>1</td><td>1</td></tr><tr><td>Servo</td><td>14,400</td><td>120</td><td>28,800</td><td>240</td><td>60</td><td>180</td><td>1</td><td>3</td></tr><tr><td>Clip</td><td>28,800</td><td>240</td><td>43,200</td><td>360</td><td>120</td><td>240</td><td>4</td><td>2</td></tr><tr><td>U Disk</td><td>14,400</td><td>240</td><td>14,400</td><td>240</td><td>120</td><td>120</td><td>4</td><td>1</td></tr><tr><td>Hinge</td><td>3,600</td><td>30</td><td>7,200</td><td>60</td><td>15</td><td>45</td><td>1</td><td>2</td></tr><tr><td>Sticky Roller</td><td>5,400</td><td>30</td><td>8,100</td><td>45</td><td>15</td><td>30</td><td>1</td><td>2</td></tr><tr><td>Caster Wheel</td><td>5,400</td><td>30</td><td>10,800</td><td>60</td><td>15</td><td>45</td><td>1</td><td>2</td></tr><tr><td>Screw</td><td>5,400</td><td>30</td><td>8,100</td><td>45</td><td>15</td><td>30</td><td>1</td><td>2</td></tr><tr><td>Lock</td><td>7,200</td><td>120</td><td>10,800</td><td>180</td><td>60</td><td>120</td><td>1</td><td>2</td></tr><tr><td>Gear</td><td>21,600</td><td>180</td><td>54,000</td><td>450</td><td>90</td><td>360</td><td>3</td><td>4</td></tr><tr><td>Clock</td><td>27,000</td><td>150</td><td>40,320</td><td>224</td><td>73</td><td>151</td><td>5</td><td>2</td></tr><tr><td>Slide</td><td>7,200</td><td>60</td><td>18,000</td><td>150</td><td>30</td><td>120</td><td>1</td><td>3</td></tr><tr><td>Zipper</td><td>14,400</td><td>120</td><td>21,600</td><td>180</td><td>60</td><td>120</td><td>2</td><td>2</td></tr><tr><td>Button</td><td>21,600</td><td>120</td><td>54,000</td><td>300</td><td>60</td><td>240</td><td>4</td><td>4</td></tr><tr><td>Liquid</td><td>5,400</td><td>30</td><td>8,100</td><td>45</td><td>15</td><td>30</td><td>1</td><td>2</td></tr><tr><td>Rubber Band</td><td>10,800</td><td>60</td><td>10,800</td><td>60</td><td>30</td><td>30</td><td>1</td><td>1</td></tr><tr><td>Ball</td><td>21,600</td><td>90</td><td>32,400</td><td>135</td><td>45</td><td>30</td><td>3</td><td>2</td></tr><tr><td>Magnet</td><td>10,800</td><td>60</td><td>16,200</td><td>90</td><td>30</td><td>60</td><td>2</td><td>2</td></tr><tr><td>Toothpaste</td><td>16,200</td><td>90</td><td>16,200</td><td>90</td><td>45</td><td>45</td><td>3</td><td>1</td></tr><tr><td>Total</td><td>298,800</td><td>2400</td><td>511,020</td><td>4,034</td><td>1,198</td><td>2,836</td><td>49</td><td>47</td></tr></table>",
+        "bbox": [
+            516,
+            133,
+            903,
+            357
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3. Data Statistics",
+        "text_level": 1,
+        "bbox": [
+            513,
+            376,
+            661,
+            390
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Dataset sample distribution. Table 3 provides a detailed breakdown of the Phys-AD dataset, which includes information on category, the number of training videos and frames, the number of testing videos and frames, the distribution of normal and abnormal samples in the test set, and the number of object and defect types. The length of the video clips in the dataset ranges from 60 to 240 frames, ensuring to fully capture the interaction process in a short time. The dataset contains 2400 training videos and 4034 test videos spreading across 22 categories, 49 object types, and 47 defect types. The frames are extracted from the videos at a rate of 60 FPS. The test set includes 1,198 normal samples and 2,836 abnormal samples. Overall, the Phys-AD dataset covers a wide range of categories and contains a large scale of data, which helps to train more robust anomaly detection models and provides a reasonable evaluation setting.",
+        "bbox": [
+            509,
+            398,
+            906,
+            641
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Interaction and defect types across categories. Table 2 details the defect types and corresponding interaction methods for each category in the Phys-AD dataset. Covering 22 categories and including 47 defect types, the dataset features various interactions (e.g., Rotate, Grab). Importantly, interaction methods or defect types with identical names can differ across categories—for example, \"rotation\" for a fan describes the motor-driven movement of its blades, while for a hinge it refers to the turning of its pages around the shaft under robotic control. Furthermore, to detect anomalies with complex physical properties in a human-like manner, we sometimes combine multiple interactions. For instance, when inspecting a screw, we first press it into the hole and then use a robotic arm to rotate it further. Overall, Phys-AD provides a diverse, physics-grounded anomaly detection framework that fosters the development of advanced methods for tackling real-world challenges.",
+        "bbox": [
+            511,
+            643,
+            908,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "30413",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/37b08d3a354aa271391e904fcbd765b6bb939dde01b0e2dacc73c68dd212e63c.jpg",
+        "image_caption": [
+            "Figure 6. PhysAD-Agent: A Large Language Model (LLM)-powered system for physics anomaly detection label augmentation and automatic evaluation. This agent framework consists of two main components: (a) Rich Annotation Generation, where an LLM-based data agent generates detailed anomaly annotations based on video, anomaly state, prompt, and human-annotated physical priors, and (b) Automatic Evaluation, where an LLM-based evaluation agent assesses model predictions to calculate the Physics Anomaly Explanation (PAEval) metric."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            90,
+            480,
+            351
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Labels. For the test set videos, we provide two types of labels. First, there are the common video-level labels: all anomalous videos in the test set are labeled as 1, while normal videos are labeled as 0. For evaluating visual-language models, we also provide text labels. Specifically, for each type of anomaly in the test set, we manually design a textual description label. This label includes both a description of the video content and a physical explanation of the reason for the anomaly. To ensure diversity in our text labels, we also use ChatGPT-4o for text augmentation. For further details on the labeling process, please refer to Fig 6a.",
+        "bbox": [
+            89,
+            508,
+            483,
+            675
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Phys-AD Benchmark",
+        "text_level": 1,
+        "bbox": [
+            89,
+            686,
+            290,
+            704
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Problem Definition and Challenges",
+        "text_level": 1,
+        "bbox": [
+            89,
+            712,
+            395,
+            729
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We establish unsupervised and weakly supervised AD settings for Phys-AD, using unsupervised as the default in our problem definition.",
+        "bbox": [
+            89,
+            734,
+            482,
+            777
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We formulate our Phys-AD setting as two steps.",
+        "bbox": [
+            89,
+            780,
+            406,
+            795
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Step 1: Rules deduction Given a set of training objects  $\\mathcal{T} = \\{t_i\\}_{i=1}^N$  from category  $c_i$ , we first use robotics arms and motors to interact with  $\\mathcal{T}$  in corresponding methods  $I_i$ , and we get the interaction process in video format as  $V_i$ . Then, we feed the video sequence  $V_i$  and the corresponding category's physical prior  $P_i$  together to the deduction function  $f$  to obtain the normal rules  $r_i$ . After rules deduction for all the",
+        "bbox": [
+            89,
+            795,
+            483,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "categories, we get the rules bank  $\\mathcal{M} = \\{r_i\\}_{i=1}^N$ . Step 1 can be formulated as below:",
+        "bbox": [
+            511,
+            89,
+            906,
+            119
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {T} \\Theta I _ {i} = V _ {i} \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            669,
+            133,
+            906,
+            148
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf \\left(V _ {i} \\mid P _ {i}\\right) = r _ {i} \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            663,
+            159,
+            906,
+            175
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {M} = \\left\\{r _ {i} \\right\\} _ {i = 1} ^ {N} \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            665,
+            180,
+            906,
+            199
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "$\\Theta$  represents interaction between objects and robotics arms.  $N$  is the total number of categories.",
+        "bbox": [
+            511,
+            205,
+            906,
+            234
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Step 2: Anomaly reasoning During test time, we first use robotics arms and motors to interact with  $\\mathcal{T}$  from the test set in corresponding methods  $I_{i}$ , and we get the interaction process in video format as  $V_{i}$ . Then, we feed the video sequence  $V_{i}$  and rules  $r_i$  from step 1 to the reasoner  $R$  to predict the object's anomaly score  $S$ . Step 2 can be formulated as:",
+        "bbox": [
+            511,
+            234,
+            906,
+            325
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {T} \\Theta I _ {i} = V _ {i} \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            671,
+            337,
+            906,
+            353
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR \\left(V _ {i} \\mid r _ {i}\\right) = S, r _ {i} \\in \\mathcal {M} \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            638,
+            364,
+            906,
+            380
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Challenges. Key challenges include combining video information with objects' physical priors to deduct normal rules, capturing long-term temporal dependencies and fine-grained frame-level information, ensuring robust reasoning for anomaly prediction, and generalizing across diverse anomaly patterns. Addressing these challenges is essential for advancing anomaly detection in real-world environment.",
+        "bbox": [
+            511,
+            386,
+            908,
+            492
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Evaluation Settings",
+        "text_level": 1,
+        "bbox": [
+            511,
+            501,
+            699,
+            517
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Unsupervised AD. Unsupervised AD is the most common IAD setting for existing IAD datasets and algorithms. Under unsupervised AD setting, the training set contains only normal video data, and the algorithm needs to capture the normal distribution of the data from the training set. In the test set, both normal and abnormal data are included, and the algorithm must distinguish between normal and abnormal data based on the distribution learned during training.",
+        "bbox": [
+            511,
+            523,
+            906,
+            643
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Weakly-supervised AD. Our dataset is in the form of videos for complex anomaly detection in industrial scenarios. In the video context, it is inevitable that we need to discuss weakly supervised anomaly detection. Under weakly-supervised Phys-AD setting, 2~4 video-level labeled anomaly samples are sampled from all possible anomaly classes in the test set in our Phys-AD dataset. These sampled anomalies are then removed from the test set. It's worth noting that we only provide video level label in our test set. This is because most anomalies in our test set require reasoning based on the entire video content and physical knowledge of the objects.",
+        "bbox": [
+            511,
+            643,
+            908,
+            810
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Video-Understanding AD. Video-Understanding models are another potential solution to our Phys-AD setting. Similar to unsupervised setting, we only provide normal videos during training. The visual language models need to provide explicit normal rules in language format during training. In the test phase, these video-understanding AD",
+        "bbox": [
+            511,
+            810,
+            908,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "30414",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/ce97f53bb8944f82e5793cc25a917db16081023a8abd500324d9f0518cf27c17.jpg",
+        "table_caption": [
+            "Table 4. Video-level AUROC (↑) result of 22 categories on Phys-AD dataset. We include Unsupervised, Weakly-supervised and Video-understanding methods. The best per-category result for each class of methods is highlighted in **bold**. 'ZS ImgB', 'V-ChatGPT', 'V-LLaMA', 'V-LLaVA' denote ZS ImageBind, Video-ChatGPT, Video-LLaMA and Video-LLaVA."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Category.</td><td colspan=\"5\">Unsupervised</td><td colspan=\"3\">Weakly-supervised</td><td colspan=\"6\">Video-understanding</td></tr><tr><td>MPN [34]</td><td>MemAE [16]</td><td>MNAD.p [40]</td><td>MNAD.r [40]</td><td>SVM [44]</td><td>VADClip [51]</td><td>S3R [48]</td><td>MGFN [10]</td><td>LAVAD [56]</td><td>ZS Clip [42]</td><td>ZS ImgB [15]</td><td>V-ChatGPT [36]</td><td>V-LLaMA [59]</td><td>V-LLaVA [28]</td></tr><tr><td>Car</td><td>0.229</td><td>0.523</td><td>0.492</td><td>0.944</td><td>0.587</td><td>0.581</td><td>0.606</td><td>0.571</td><td>0.557</td><td>0.500</td><td>0.500</td><td>0.500</td><td>0.678</td><td>0.522</td></tr><tr><td>Fan</td><td>0.811</td><td>0.371</td><td>0.810</td><td>0.542</td><td>0.500</td><td>0.624</td><td>0.640</td><td>0.542</td><td>0.510</td><td>0.500</td><td>0.500</td><td>0.549</td><td>0.592</td><td>0.611</td></tr><tr><td>Rolling Bearing</td><td>0.353</td><td>0.044</td><td>0.352</td><td>0.800</td><td>0.933</td><td>0.589</td><td>0.601</td><td>0.680</td><td>0.532</td><td>0.500</td><td>0.500</td><td>0.300</td><td>0.500</td><td>0.500</td></tr><tr><td>Spherical Bearing</td><td>0.113</td><td>0.092</td><td>0.962</td><td>0.813</td><td>0.650</td><td>0.500</td><td>0.682</td><td>0.583</td><td>0.435</td><td>0.500</td><td>0.500</td><td>0.450</td><td>0.550</td><td>0.500</td></tr><tr><td>Servo</td><td>0.364</td><td>0.445</td><td>0.975</td><td>0.878</td><td>0.500</td><td>0.518</td><td>0.592</td><td>0.556</td><td>0.502</td><td>0.500</td><td>0.500</td><td>0.506</td><td>0.683</td><td>0.464</td></tr><tr><td>Clip</td><td>0.535</td><td>0.443</td><td>0.630</td><td>0.333</td><td>0.500</td><td>0.412</td><td>0.561</td><td>0.563</td><td>0.516</td><td>0.500</td><td>0.500</td><td>0.669</td><td>0.556</td><td>0.458</td></tr><tr><td>U Disk</td><td>0.240</td><td>0.617</td><td>0.609</td><td>0.940</td><td>0.500</td><td>0.530</td><td>0.549</td><td>0.570</td><td>0.513</td><td>0.500</td><td>0.500</td><td>0.565</td><td>0.575</td><td>0.500</td></tr><tr><td>Hinge</td><td>0.769</td><td>0.870</td><td>0.705</td><td>0.895</td><td>0.500</td><td>0.737</td><td>0.561</td><td>0.550</td><td>0.564</td><td>0.500</td><td>0.500</td><td>0.500</td><td>0.692</td><td>0.500</td></tr><tr><td>Sticky Roller</td><td>0.967</td><td>0.967</td><td>0.451</td><td>0.936</td><td>0.500</td><td>0.542</td><td>0.835</td><td>0.669</td><td>0.266</td><td>0.500</td><td>0.500</td><td>0.450</td><td>0.544</td><td>0.467</td></tr><tr><td>Caster Wheel</td><td>0.364</td><td>0.523</td><td>0.271</td><td>0.508</td><td>0.500</td><td>0.587</td><td>0.676</td><td>0.676</td><td>0.615</td><td>0.500</td><td>0.500</td><td>0.444</td><td>0.642</td><td>0.500</td></tr><tr><td>Screw</td><td>0.522</td><td>0.567</td><td>0.680</td><td>0.547</td><td>0.500</td><td>0.500</td><td>0.657</td><td>0.541</td><td>0.688</td><td>0.500</td><td>0.500</td><td>0.472</td><td>0.256</td><td>0.550</td></tr><tr><td>Lock</td><td>0.563</td><td>0.597</td><td>0.430</td><td>0.641</td><td>0.733</td><td>0.452</td><td>0.523</td><td>0.626</td><td>0.341</td><td>0.500</td><td>0.500</td><td>0.279</td><td>0.494</td><td>0.500</td></tr><tr><td>Gear</td><td>0.529</td><td>0.510</td><td>0.652</td><td>0.694</td><td>0.500</td><td>0.519</td><td>0.580</td><td>0.544</td><td>0.603</td><td>0.500</td><td>0.500</td><td>0.544</td><td>0.603</td><td>0.517</td></tr><tr><td>Clock</td><td>0.340</td><td>0.542</td><td>0.395</td><td>0.587</td><td>0.500</td><td>0.500</td><td>0.549</td><td>0.509</td><td>0.500</td><td>0.500</td><td>0.500</td><td>0.501</td><td>0.360</td><td>0.500</td></tr><tr><td>Slide</td><td>0.517</td><td>0.962</td><td>0.917</td><td>0.784</td><td>0.500</td><td>0.531</td><td>0.611</td><td>0.568</td><td>0.425</td><td>0.500</td><td>0.500</td><td>0.562</td><td>0.567</td><td>0.179</td></tr><tr><td>Zipper</td><td>0.815</td><td>0.592</td><td>0.829</td><td>0.421</td><td>0.500</td><td>0.504</td><td>0.636</td><td>0.633</td><td>0.535</td><td>0.500</td><td>0.500</td><td>0.547</td><td>0.489</td><td>0.500</td></tr><tr><td>Button</td><td>0.853</td><td>0.365</td><td>0.660</td><td>0.568</td><td>0.500</td><td>0.627</td><td>0.515</td><td>0.566</td><td>0.439</td><td>0.500</td><td>0.500</td><td>0.515</td><td>0.517</td><td>0.360</td></tr><tr><td>Liquid</td><td>0.184</td><td>0.700</td><td>0.671</td><td>0.831</td><td>0.500</td><td>0.542</td><td>0.595</td><td>0.793</td><td>0.504</td><td>0.500</td><td>0.500</td><td>0.410</td><td>0.278</td><td>0.217</td></tr><tr><td>Rubber Band</td><td>0.374</td><td>0.368</td><td>0.366</td><td>0.307</td><td>0.567</td><td>0.482</td><td>0.623</td><td>0.604</td><td>0.511</td><td>0.500</td><td>0.500</td><td>0.517</td><td>0.450</td><td>0.450</td></tr><tr><td>Ball</td><td>0.543</td><td>0.383</td><td>0.728</td><td>0.687</td><td>0.500</td><td>0.500</td><td>0.667</td><td>0.567</td><td>0.603</td><td>0.500</td><td>0.500</td><td>0.562</td><td>0.636</td><td>0.533</td></tr><tr><td>Magnet</td><td>0.671</td><td>0.464</td><td>0.691</td><td>0.438</td><td>0.500</td><td>0.500</td><td>0.548</td><td>0.719</td><td>0.502</td><td>0.500</td><td>0.500</td><td>0.683</td><td>0.300</td><td>0.400</td></tr><tr><td>Toothpaste</td><td>0.587</td><td>0.889</td><td>0.520</td><td>0.631</td><td>0.500</td><td>0.500</td><td>0.686</td><td>0.711</td><td>0.562</td><td>0.500</td><td>0.500</td><td>0.376</td><td>0.550</td><td>0.467</td></tr><tr><td>Average</td><td>0.511</td><td>0.538</td><td>0.627</td><td>0.669</td><td>0.544</td><td>0.535</td><td>0.613</td><td>0.606</td><td>0.510</td><td>0.500</td><td>0.500</td><td>0.496</td><td>0.523</td><td>0.463</td></tr></table>",
+        "bbox": [
+            93,
+            136,
+            903,
+            465
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "methods have to truly understand the videos and predict the anomalies based on the language rules withdrawn from the training phase.",
+        "bbox": [
+            89,
+            484,
+            480,
+            530
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3. Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            89,
+            556,
+            274,
+            571
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Physics anomaly accuracy metrics. We use the Area Under the Receiver Operating Characteristic Curve (AUROC) to evaluate video-level anomaly detection performance. We also report the average precision (AP), i.e. the area under the video-level precision-recall curve and the acc (ACCURACY), following previous works [47, 49].",
+        "bbox": [
+            89,
+            579,
+            482,
+            670
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Physics Anomaly Explanation (PAEval) metric. Directly utilizing existing VLMs [59][36] to understand the videos in our dataset and detect the anomalies is a potential solution to our challenge. The key point is: Can existing VLMs truly understand physical rules and reason in a right way? Specifically designed for video-understanding methods, we introduce a new metric named Physics Anomaly Explanation (PAEval) metric. To be more specific, PAEval evaluate the anomaly detection performance of algorithms based on VLMs from three different perspectives: classification, description, and explanation. Classification refers to the traditional anomaly detection metrics like AUROC, etc. Inspired by the works [11, 61], PAEval also introduces two additional evaluation metrics: description and physical explanation. Description refers to the model's ability to",
+        "bbox": [
+            89,
+            674,
+            482,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "correctly describe the content of the video, used to assess whether the model has the capability to describe physical phenomena. Explanation refers to the model's ability to correctly explain the physical causes of anomalies in the video, used to evaluate the VLM's reasoning ability. To provide labels for description and explanation metrics, we manually labeled each type of defects from all categories and performed data augmentation by ChatGPT to ensure the robustness of the detection. The whole pipeline of PAEval metric is depicted in Fig 6.",
+        "bbox": [
+            511,
+            484,
+            906,
+            636
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5. Benchmarking Results",
+        "text_level": 1,
+        "bbox": [
+            511,
+            654,
+            725,
+            671
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1. Benchmark Methods Selection",
+        "text_level": 1,
+        "bbox": [
+            511,
+            680,
+            784,
+            695
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "For the Phys-AD setting, we select popular and reproducible video anomaly detection methods across unsupervised, weakly-supervised, and video-understanding setting. In the unsupervised setting, we focus on reconstruction, prediction, and embedding-based models like MemAE [16] and MNAD [40], which use memory modules to enhance anomaly discrimination. For weakly-supervised anomaly detection, we adopt methods such as VadCLIP [51] and MGFN [10], which use feature magnitudes and vision-language associations. In video-understanding, we evaluate video-language models like Video-ChatGPT [36] and image-language models like CLIP [42] to predict detailed anomaly descriptions and scores through cross-modal embeddings.",
+        "bbox": [
+            511,
+            703,
+            908,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "30415",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/d8144d5b51de9e11ae4893f04065b9163b95c6453985312f9855c3e67382fa9d.jpg",
+        "table_caption": [
+            "Table 5. Physics Anomaly Explanation (PAEval) metric results on our Phys-AD dataset."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>Classification (↑)</td><td>Description (↑)</td><td>Explanation (↑)</td></tr><tr><td>LAVAD[56]</td><td>0.510</td><td>0.157</td><td>0.000</td></tr><tr><td>Video-ChatGPT[36]</td><td>0.496</td><td>0.131</td><td>0.160</td></tr><tr><td>Video-LLaMA[59]</td><td>0.523</td><td>0.137</td><td>0.303</td></tr><tr><td>Video-LLaVA[28]</td><td>0.463</td><td>0.219</td><td>0.282</td></tr></table>",
+        "bbox": [
+            93,
+            119,
+            480,
+            189
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Code and Experiment details. We provide code and toolkit for our dataset and benchmark. More experiment details are listed in Appendix.",
+        "bbox": [
+            89,
+            209,
+            482,
+            255
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2. Main Findings",
+        "text_level": 1,
+        "bbox": [
+            89,
+            263,
+            238,
+            280
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Overall anomaly detection benchmarking results. Table 4 shows that existing video anomaly detection and videowhildunderstanding methods achieve limited performance on the Phys-AD dataset, with the highest AUROC only reaching  $66.9\\%$  for MNAD.r [40]. This result underscores the heightened complexity of Phys-AD compared to existing datasets, highlighting a domain shift that challenges current industrial anomaly detection algorithms, which are often tuned to visually distinct anomaly patterns in single frames rather than complex temporal or physical cues.",
+        "bbox": [
+            89,
+            285,
+            483,
+            436
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Anomaly detection via unsupervised and weakly supervised methods. Our experiment includes unsupervised methods (e.g., MemAE [16], MPN [34], MNAD [40]) and weakly supervised methods (e.g., S3R [48], MGFN [10], VADClip [51]). Unsupervised methods like MNAD.p [40] performed better on temporal anomalies (81.0% for fan, 68.0% for screw) by leveraging prediction-based approaches, which excel in scenarios requiring temporal understanding. Weakly supervised methods improve baseline scores across complex classes, preventing extreme low scores in challenging categories such as spherical bearing. However, weakly supervised methods show marginally lower performance in simpler anomaly classes, indicating trade-offs introduced by anomaly samples in training.",
+        "bbox": [
+            89,
+            440,
+            483,
+            651
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Unsupervised anomaly detection method performance by category. Among unsupervised methods, MemAE [16] achieves high AUROC on objects with spatial anomalies (e.g., sticky roller,  $96.7\\%$ ) but underperforms in temporal anomaly classes like rolling and spherical bearings, where it achieves only  $4.4\\%$  and  $9.2\\%$ , respectively. MNAD [40] improves temporal sensitivity by incorporating temporal prototypes, achieving better scores for bearings but still struggling with complex anomaly types, revealing limitations in purely prototype-based approaches.",
+        "bbox": [
+            89,
+            655,
+            482,
+            806
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Anomaly detection via Video-Language Models. VLM-based methods (e.g., Video-ChatGPT [36], VideoLLaMA [59]) struggle, with top-performing VideoLLaMA [59] reaching only  $52.3\\%$  average AUROC. Performance is impacted by reliance on pre-trained weights that are not optimized for physics-grounded video content, evident",
+        "bbox": [
+            89,
+            810,
+            482,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "from low AUROC scores in categories with nuanced physical dynamics, such as hinges and screws. Further, PAEval results suggest that these models lack effective reasoning about object physical dynamics and behaviors influenced by physical forces, underscoring a gap between VLM capabilities and the demands of IAD tasks.",
+        "bbox": [
+            511,
+            90,
+            906,
+            181
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Anomaly explanation via Video-Language Models. Table 5 reports PAEval metric results, with the best-performing VLM achieving only  $21.9\\%$  in description and  $30.3\\%$  in explanation. These findings emphasize that current VLMs lack the depth in physical reasoning and temporal coherence required for understanding real-world physics-based scenarios, which our Phys-AD dataset demands.",
+        "bbox": [
+            511,
+            185,
+            908,
+            291
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Summary. Overall, the results reveal Phys-AD's unique challenge in requiring both high spatial detail and temporal comprehension, areas where existing methods and models underperform. This analysis points to the need for future research in models that integrate temporal reasoning with physics-based anomaly detection.",
+        "bbox": [
+            511,
+            294,
+            908,
+            386
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            513,
+            398,
+            633,
+            414
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we introduce the first industrial anomaly detection task focusing on real-world scenarios where physical understanding and reasoning are essential for anomaly detection. We present the Physics Anomaly Detection (Phys-AD) dataset, a large-scale, physics-grounded video dataset with over 6400 videos across 22 categories and 49 object types interacting with robotic systems, capturing 47 anomaly types that necessitate visual and physical understanding. We assess Phys-AD, highlighting the lack of baseline approaches for high-level reasoning in anomaly detection. Additionally, we propose the Physics Anomaly Explanation (PAEval) metric to evaluate visual language models (VLMs) on physics-based reasoning. Experiments show that current VLMs fall short of human-level understanding in physics-based anomaly scenarios. This work marks a milestone for industrial anomaly detection, promoting physics-grounded reasoning in complex industrial settings.",
+        "bbox": [
+            511,
+            422,
+            908,
+            680
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitation and future work. Although our Phys-AD dataset provides a large variety of objects with diverse physical properties and various types of interaction methods, we plan to add even more diverse interaction methods and objects in the future to better meet the demands of complex real-world industrial scenarios. Due to the significant differences between our Phys-AD dataset and current industrial anomaly detection and video anomaly detection datasets, most existing anomaly detection algorithms cannot be directly applied to our dataset. In the future, we will test more algorithms on our Phys-AD dataset and provide experimental results across more settings like zero-shot, few-shot, semi-supervised settings, etc.",
+        "bbox": [
+            511,
+            684,
+            908,
+            866
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "30416",
+        "bbox": [
+            478,
+            944,
+            519,
+            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] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015. 3",
+            "[2] Fabien Baradel, Natalia Neverova, Julien Mille, Greg Mori, and Christian Wolf. Cophy: Counterfactual learning of physical dynamics. arXiv preprint arXiv:1909.12000, 2019.3",
+            "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9584-9592, 2019. 3",
+            "[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. MVtec AD – A comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, 2019. 2",
+            "[5] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130 (4):947-969, 2022. 3",
+            "[6] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. SCITEPRESS - Science and Technology Publications, 2022. 3",
+            "[7] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The MVTec 3d-AD dataset for unsupervised 3d anomaly detection and localization. In VISAPP, 2022. 2",
+            "[8] Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. In Proceedings of the Asian Conference on Computer Vision, pages 3586-3602, 2022. 3",
+            "[9] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 2",
+            "[10] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In Proceedings of the AAAI conference on artificial intelligence, pages 387-395, 2023. 7, 8",
+            "[11] Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, et al. Uncovering what why and how: A comprehensive benchmark for causation understanding of video anomaly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18793-18803, 2024. 7",
+            "[12] Jiafei Duan, Samson Yu, and Cheston Tan. Space: A simulator for physical interactions and causal learning in 3d environments. In ICCV, pages 2058-2063, 2021. 3",
+            "[13] Jiafei Duan, Samson Yu, Soujanya Poria, Bihan Wen, and Cheston Tan. Pip: Physical interaction prediction via mental"
+        ],
+        "bbox": [
+            93,
+            114,
+            483,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "simulation with span selection. In ECCV, pages 405-421. Springer, 2022. 3",
+            "[14] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14009-14018, 2021. 3",
+            "[15] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 7",
+            "[16] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1705-1714, 2019. 3, 7, 8",
+            "[17] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 3",
+            "[18] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K Roy-Chowdhury, and Larry S Davis. Learning temporal regularity in video sequences. In CVPR, 2016. 3",
+            "[19] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 3",
+            "[20] Steeven Janny, Fabien Baradel, Natalia Neverova, Madiha Nadri, Greg Mori, and Christian Wolf. Filtered-cophy: Unsupervised learning of counterfactual physics in pixel space. In ICLR, 2022. 3",
+            "[21] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT), pages 66-71, 2021. 2, 3",
+            "[22] Huaizu Jiang, Xiaojian Ma, Weili Nie, Zhiding Yu, Yuke Zhu, and Anima Anandkumar. Bongard-hoi: Benchmarking few-shot visual reasoning for human-object interactions. In CVPR, pages 19056-19065, 2022. 3",
+            "[23] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 3",
+            "[24] Hyekang Kevin Joo, Khoa Vo, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 3230-3234. IEEE, 2023. 3",
+            "[25] Federico Landi, Cees GM Snok, and Rita Cucchiara. Anomaly locality in video surveillance. arXiv preprint arXiv:1901.10364, 2019. 3",
+            "[26] Ting Li, Charles Dubout, Emma K Wampler, Steven Yantis, Donald Geman, et al. Comparing machines and humans on a visual categorization test. 2011. 3"
+        ],
+        "bbox": [
+            516,
+            92,
+            906,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "30417",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[27] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. arXiv preprint, 2023. 2, 3",
+            "[28] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 7, 8",
+            "[29] Jiaqi Liu, Guoyang Xie, Rui Chen, et al. Real3d-ad: A dataset of point cloud anomaly detection. NeurIPS, 2023. 2",
+            "[30] Jiaqi Liu, Guoyang Xie, Ruitao Chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. Real3dad: A dataset of point cloud anomaly detection. Advances in Neural Information Processing Systems, 36, 2024. 3",
+            "[31] Kun Liu and Huadong Ma. Exploring background-bias for anomaly detection in surveillance videos. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1490-1499, 2019. 3",
+            "[32] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection-a new baseline. In CVPR, 2018. 3",
+            "[33] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588-13597, 2021. 3",
+            "[34] Hui Lv, Chen Chen, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Learning normal dynamics in videos with meta prototype network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15425-15434, 2021. 7, 8",
+            "[35] Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8022-8031, 2023. 3",
+            "[36] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 7, 8",
+            "[37] Domingo Mery, Vladimir Riffo, Uwe Zscherpel, German Mondragon, Ivan Lillo, Irene Zuccar, Hans Lobel, and Miguel Carrasco. Gdxray: The database of x-ray images for nondestructive testing. Journal of Nondestructive Evaluation, 34 (4):42, 2015. 3",
+            "[38] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adj: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE). IEEE, 2021. 3",
+            "[39] Weili Nie, Zhiding Yu, Lei Mao, Ankit B Patel, Yuke Zhu, and Animashree Anandkumar. Bongard-logo: A new benchmark for human-level concept learning and reasoning. In NeurIPS, 2020. 3",
+            "[40] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceed-"
+        ],
+        "bbox": [
+            91,
+            92,
+            482,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14372-14381, 2020. 7, 8",
+            "[41] Yujiang Pu, Xiaoyu Wu, and Shengjin Wang. Learning prompt-enhanced context features for weakly-supervised video anomaly detection. arXiv preprint arXiv:2306.14451, 2023. 3",
+            "[42] 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. 3, 7",
+            "[43] Binyi Su, Zhong Zhou, and Haiyong Chen. Pvel-ad: A large-scale open-world dataset for photovoltaic cell anomaly detection. IEEE Transactions on Industrial Informatics, 19 (1):404-413, 2023. 3",
+            "[44] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, 2018. 7",
+            "[45] Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. In Proceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 5541-5550, 2023. 3",
+            "[46] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jianning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection, 2024. 3",
+            "[47] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032-8041, 2023. 7",
+            "[48] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 7, 8",
+            "[49] Peng Wu, Jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 322-339. Springer, 2020. 3, 7",
+            "[50] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection. arXiv preprint arXiv:2311.07042, 2023. 3",
+            "[51] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. arXiv preprint arXiv:2308.11681, 2023.3,7,8",
+            "[52] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. Detecting anomalous events in videos by learning deep representations of appearance and motion. Computer Vision and Image Understanding, 156:117-127, 2017. 3",
+            "[53] Xiaohao Xu, Yunkang Cao, Yongqi Chen, Weiming Shen, and Xiaonan Huang. Customizing visual-language foundation models for multi-modal anomaly detection and reasoning. arXiv preprint arXiv:2403.11083, 2024. 3"
+        ],
+        "bbox": [
+            516,
+            92,
+            906,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "30418",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[54] Zhiwei Yang, Jing Liu, Zhaoyang Wu, Peng Wu, and Xiaotao Liu. Video event restoration based on keyframes for video anomaly detection. In CVPR, 2023. 3",
+            "[55] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14744-14754, 2022. 3",
+            "[56] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. arXiv preprint arXiv:2404.01014, 2024. 3, 7, 8",
+            "[57] Aimen Zerroug, Mohit Vaishnav, Julien Colin, Sebastian Musslick, and Thomas Serre. A benchmark for compositional visual reasoning. arXiv preprint arXiv:2206.05379, 2022. 3",
+            "[58] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In CVPR, 2019. 3",
+            "[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 7, 8",
+            "[60] Huaxin Zhang, Xiang Wang, Xiaohao Xu, Xiaonan Huang, Chuchu Han, Yuehuan Wang, Changxin Gao, Shanjun Zhang, and Nong Sang. Glancevad: Exploring glance supervision for label-efficient video anomaly detection. arXiv preprint arXiv:2403.06154, 2024. 3",
+            "[61] Huaxin Zhang, Xiaohao Xu, Xiang Wang, Jialong Zuo, Chuchu Han, Xiaonan Huang, Changxin Gao, Yuehuan Wang, and Nong Sang. Holmes-vad: Towards unbiased and explainable video anomaly detection via multi-modal llm. arXiv preprint arXiv:2406.12235, 2024. 3, 7",
+            "[62] Yiheng Zhang, Yunkang Cao, Xiaohao Xu, and Weiming Shen. Logic: An llm-driven framework for logical anomaly detection. IEEE Transactions on Automation Science and Engineering, pages 1-0, 2024. 3",
+            "[63] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In CVPR, 2019. 3",
+            "[64] Qiang Zhou, Weize Li, Lihan Jiang, Guoliang Wang, Guyue Zhou, Shanghang Zhang, and Hao Zhao. Pad: A dataset and benchmark for pose-agnostic anomaly detection, 2023. 3",
+            "[65] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In ECCV, 2022. 2",
+            "[66] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation, 2022. 3"
+        ],
+        "bbox": [
+            91,
+            92,
+            482,
+            785
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "30419",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/e496d86c-0d1a-42d1-b098-7f552a70c4ff_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8bfb8a7a032072416846dba924bf24ba41acba814630d98a722422664d82b18
+size 1307578

2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/full.md

@@ -0,0 +1,349 @@
+# Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics: Physics-Grounded Anomaly Detection
+
+Wenqiao Li $^{1*}$  Yao Gu $^{1*}$  Xintao Chen $^{1*}$  Xiaohao Xu $^{2\dagger}$  Ming Hu $^{3}$  Xiaonan Huang $^{2}$  Yingna Wu $^{1\dagger}$ $^{1}$ ShanghaiTech University  $^{2}$ University of Michigan, Ann Arbor  $^{3}$ Monash University
+
+![](images/75d6ea3cc4000b98b604cf49e2eedc38b5b4f3a012591964ab21612bc71ce942.jpg)  
+Figure 1. Towards visual discrimination of physical dynamics in real-world industrial object anomaly detection. We illustrate objects, interactions, and time-sequenced videos from the Physics-Grounded Anomaly Detection dataset: (a) Object; (b) Interaction: Applied actions shown with directional arrows; (c) Video with Physical Dynamics: Temporal sequences showing normal and abnormal states, highlighting anomalies like leaks, misalignments, and cracks. By focusing on the dynamic behaviors of complex objects, we enhance understanding of interactions and failure modes in real-world settings, where both structure and motion contribute to anomaly detection.
+
+![](images/406ce781989bb97be996195c92fcaa5cde7a373227f124df11e22d77c6ab559c.jpg)
+
+# Abstract
+
+Humans detect real-world object anomalies by perceiving, interacting, and reasoning based on object-conditioned physical knowledge. The long-term goal of Industrial Anomaly Detection (IAD) is to enable machines to autonomously replicate this skill. However, current IAD algorithms are largely developed and tested on static, semantically simple datasets, which diverge from real-world scenarios where physical understanding and reasoning are essential. To bridge this gap, we introduce the Physics Anomaly Detection (Phys-AD) dataset, the first large-scale, real-world, physics-grounded video dataset for industrial anomaly detection. Collected using a real robot arm and motor, Phys-AD provides a diverse
+
+set of dynamic, semantically rich scenarios. The dataset includes more than 6400 videos across 22 real-world object categories, interacting with robot arms and motors, and exhibits 47 types of anomalies. Anomaly detection in Phys-AD requires visual reasoning, combining both physical knowledge and video content to determine object abnormality. We benchmark state-of-the-art anomaly detection methods under three settings: unsupervised AD, weakly-supervised AD, and video-understanding AD, highlighting their limitations in handling physics-grounded anomalies. Additionally, we introduce the Physics Anomaly Explanation (PAEval) metric, designed to assess the ability of visual-language foundation models to not only detect anomalies but also provide accurate explanations for their underlying physical causes. Our project is available at https://guyao2023.github.io/Phys-AD/.
+
+![](images/fc4dc8c941e0683bab05b70a96d80230028a00681e4d6d0074810515380bb537.jpg)  
+Figure 2. Human-like decision-making process for physics-grounded object anomaly detection. We illustrate the sequential approach of a human-like agent for evaluating an object's normality. First, the agent perceives relevant physical attributes (e.g., plastic and elastic), then interacts by performing a physical action (e.g., squeezing), and finally reasons based on the vision feedback and attributes changes (e.g., surface shape change) to determine whether the object is normal or anomalous. This mirrors a human's natural process of reasoning over physics in objects.
+
+# 1. Introduction
+
+Industrial anomaly detection (IAD) is a critical subfield in computer vision and industrial automation, aiming to identify defects or irregularities in products during manufacturing. As shown in Fig. 2, the ultimate vision is to create autonomous systems that not only perceive but also interact with and reason about objects to discriminate between anomalies and normal states, integrating complex physical principles to detect anomalies in dynamic, real-world scenarios. For example, a human inspecting a water bottle for anomalies wouldn't rely solely on visual observation; they might rotate or invert the bottle, using physical interactions and cues, such as noticing a loose cap or an irregular internal flow, to detect issues.
+
+Central to advancing IAD [9] is the availability of high-quality datasets that bridge the gap between academic research and industrial needs. Datasets like MVTec-AD [4], MPDD [21], and VisA [65] have played foundational roles, enabling algorithm development for image-based anomaly detection and bringing IAD to the forefront of computer vision research. While these datasets have significantly advanced single-image anomaly detection, recent datasets, e.g., MVTec-3D [7], Real3D [29], and Anomaly-ShapeNet [27], have extended IAD to 3D, aligning research more closely with the needs of complex real-world industrial settings.
+
+Yet, as factories increasingly rely on robotic arms and automated systems to perform sophisticated inspections, the limitations of current IAD datasets become apparent.
+
+Existing benchmarks focus on static, semantically simple environments, overlooking the physical priors and interactive reasoning required in real-world industrial contexts. This gap highlights a growing need for datasets that not only reflect real-world physical constraints but also challenge models to reason dynamically about anomalies.
+
+To bridge this gap, as shown in Fig. 1, we introduce the Physics Anomaly Detection (Phys-AD) dataset, the first large-scale, physics-grounded video dataset for industrial anomaly detection. Phys-AD features over 6,400 videos of 22 categories and 49 objects interacting with robotic arms and motors, capturing 47 types of anomalies that require visual reasoning informed by physical knowledge. The short video clips in the dataset range from 60 to 240 frames in length and are filmed in real-time industrial environment, fully capturing the interaction process between robotic arm or motor and industrial objects. Additionally, to ensure our dataset meets industrial demands and matches the complexity and diversity of the real physical world, we selected industrial objects of different physical qualities, various interaction methods, and anomalies that reflect diverse physical principles and require different reasoning process. Specifically, we selected 22 object categories spanning across metals, plastics, fluid, amorphous substances and articulated objects with diverse appearances. For interaction, we use mechanical grippers, robotic arms, and motors, incorporating various interaction modes such as pressing, rotating, squeezing, and driving to handle different types of objects.
+
+We benchmark state-of-the-art anomaly detection methods in three key configurations: unsupervised anomaly detection, weakly-supervised anomaly detection, and video-understanding based anomaly detection. Our findings reveal critical gaps in their ability to handle the complexities of physics-grounded scenarios, where anomalies often arise from dynamic, interdependent interactions. To advance the field, we also introduce the Physical Anomaly Explanation (PAEval) metric, designed to assess both detection performance and a model's ability to explain anomalies by identifying underlying physical causes. Furthermore, our benchmark reveals the fragility of existing methods in tackling these challenging conditions, underscoring the need for approaches that better understand object dynamics and temporal coherence in anomaly detection.
+
+# Our contributions can be summarized as followings:
+
+- We introduce a novel task of detecting physical-based industrial anomalies in real-world that involves perception, physical and visual reasoning.  
+- We present Phys-AD, the first large-scale, physics-grounded video dataset specifically designed for industrial anomaly detection in real world, containing objects of different physical qualities, multiple interaction methods and various physical reasoning process.  
+- We benchmarking the anomaly detection and reasoning
+
+Table 1. Comparison of Phys-AD with existing industrial anomaly detection datasets. Our Phys-AD dataset is the first to consider complex objects with physical dynamics. 'Syn', 'IR', 'D', and 'PC' denote Synthetic, Infrared, Depth, and Point Cloud, respectively. #Anomaly indicates the number of anomaly types.  
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">Year</td><td rowspan="2">Type</td><td rowspan="2">Modality</td><td colspan="3">Sample Statistics</td></tr><tr><td>Class #Anomaly</td><td colspan="2">Physics</td></tr><tr><td>MVTec-AD [3]</td><td>2019</td><td>Real</td><td>RGB</td><td>15</td><td>-</td><td>X</td></tr><tr><td>BTAD [38]</td><td>2021</td><td>Real</td><td>RGB</td><td>3</td><td>3</td><td>X</td></tr><tr><td>MPDD [21]</td><td>2021</td><td>Real</td><td>RGB</td><td>6</td><td>8</td><td>X</td></tr><tr><td>VisA [66]</td><td>2021</td><td>Real</td><td>RGB</td><td>12</td><td>-</td><td>X</td></tr><tr><td>MVTec LOCO-AD [5]</td><td>2022</td><td>Real</td><td>RGB</td><td>5</td><td>-</td><td>X</td></tr><tr><td>MAD [64]</td><td>2023</td><td>Syn+Real</td><td>RGB</td><td>20</td><td>3</td><td>X</td></tr><tr><td>LOCO-Annotations [62]</td><td>2024</td><td>Real</td><td>RGB</td><td>5</td><td>5</td><td>X</td></tr><tr><td>Real-IAD [46]</td><td>2024</td><td>Real</td><td>RGB</td><td>30</td><td>8</td><td>X</td></tr><tr><td>GDXray [37]</td><td>2015</td><td>Real</td><td>X-ray</td><td>5</td><td>15</td><td>X</td></tr><tr><td>PVEL-AD [43]</td><td>2023</td><td>Real</td><td>IR</td><td>1</td><td>10</td><td>X</td></tr><tr><td>MVTec3D-AD [6]</td><td>2021</td><td>Real</td><td>RGB-D</td><td>10</td><td>3-5</td><td>X</td></tr><tr><td>Eyecandies [8]</td><td>2022</td><td>Syn</td><td>RGB-D</td><td>10</td><td>3</td><td>X</td></tr><tr><td>Real3D-AD [30]</td><td>2023</td><td>Real</td><td>PC</td><td>12</td><td>2</td><td>X</td></tr><tr><td>Anomaly-ShapeNet [27]</td><td>2024</td><td>Syn</td><td>PC</td><td>40</td><td>6</td><td>X</td></tr><tr><td>Phys-AD (Ours)</td><td>2024</td><td>Real</td><td>RGB</td><td>49</td><td>47</td><td>✓</td></tr></table>
+
+performance of popular video AD methods and Visual Language Foundation Models on the Phys-AD dataset in several settings, establishing a practical and challenging benchmark to promote the development of the physics-related anomaly detection field.
+
+# 2. Related work
+
+Industrial anomaly detection datasets. Existing datasets primarily focus on static, semantically simple scenarios, deviate significantly from real world where physical understanding and reasoning are essential. Datasets like MVTec-AD [3], BTAD [38], MPDD [21], and VisA [66] focus on surface level image anomaly detection with one single-view RGB image, limiting their effectiveness in capturing holistic object structures. MVTec-LOCO-AD [5], focusing on structure and local information in industries, is limited by its relatively simple and constrained data content. While MVTec3D-AD [6] and Eyecandies [8] incorporate depth data, they remain static and single-view image anomaly detection, neglecting object level information. To explore object level anomaly detection, multi-view IAD datasets like MAD [64] and Real-IAD [46], point cloud IAD datasets like Real3D-AD [30] and Anomaly-ShapeNet [27], offering richer visual and geometric cues, but are still limited to static objects without dynamic interaction or reasoning. In summary, current industrial anomaly detection datasets focus on relatively simple and static anomaly detection scenarios, lacking complex physical rules, dynamic interactions, and visual reasoning requirements. Therefore, existing IAD datasets generally applied to limited industrial scenarios and there is no IAD dataset could meet the demands of detecting
+
+complex anomalies in real world which need physical priors and reasoning.
+
+Video anomaly detection. Deep learning methods [18, 35, 54, 61] now dominate Video Anomaly Detection (VAD), categorized into unsupervised, weakly-supervised, and fully-supervised approaches. Unsupervised methods learn normal patterns via reconstruction [16, 18, 52], prediction [32], or hybrids [33], while some methods [45, 55] train with both unlabeled normal and abnormal data. Weakly-supervised methods[14, 49, 60, 63] use video-level or glance-based annotations, and fully-supervised methods[25, 31] remain rare due to costly frame-level labeling. Visual-language models like CLIP [42] have recently been applied to enhance anomaly detection [24, 41, 51], focusing on semantic anomalies. Open-vocabulary VAD [50] and prompt-based anomaly scoring [56] leverage LLMs [53, 61], but performance relies heavily on the base models, often lacking domain-specific tuning. However, existing video anomaly detection algorithms lack the capability to handle complex industrial anomaly detection scenarios and understand physical rules. This gap highlights the need for models that can capture dynamic behaviors and physical laws in industrial environments, as addressed by Phys-AD, which targets industrial anomalies in objects of various physical properties.
+
+Visual reasoning. Visual reasoning is a critical task in computer vision, aiming to enable machines to interpret perceptual information like humans. Several visual reasoning tasks have been proposed to evaluate reasoning capabilities, including Visual Question Answering (VQA), 2D puzzles, and physical dynamics prediction. In VQA, agents are required to combine natural language and visual cues to answer questions [1, 17, 19, 23]. For 2D puzzles, tasks involve discovering relationships among visual elements and making inferences [22, 26, 39, 57, 58]. Physical dynamics prediction tasks require machines to perceive and reason about physical interactions [2, 12, 13, 20]. In contrast to these work, we introduce the first benchmark featuring real-world industrial objects with dynamic physical properties, focusing on distinguishing diverse dynamics through vision.
+
+# 3. Dataset: Phys-AD
+
+# 3.1. Object Preparation and Interaction Selection
+
+We selected 22 categories spanning across various materials, including metals, plastics, amorphous substances and articulated objects, with diverse shapes, sizes, and physical properties. For the physical properties of different objects, we correspondingly select different interaction methods like push, rotate, pull with robotics arms and motors. For instance, we use robotics arms to grab and extruding the deformable objects like basketball to determine whether there are any elastic anomalies or surface defects by the morphological change of the objects. To make our dataset more practical
+
+![](images/d1fb0c57315c70053b2eee90ea39523b6bc62eca99f6e996b4571a08712f0d89.jpg)  
+U Disk
+
+![](images/04c3a0c007d031f41abd5c7c98604aad9ab1aaf4129e18d28427e92deee1eb37.jpg)  
+Zipper
+
+![](images/45bcc2f974a31f7bc5fa856043b21b280954cb11438e7f549aef3542277029ad.jpg)  
+Rolling Bearing
+
+![](images/b7d51d7075f154f89366e928da49136a71a85cee7477c389e2eb7f9afe4c02c5.jpg)  
+Sticky Roller
+
+![](images/138a01ee9acd66f4fd6acc7aa80c8d225c1ae73de2587940ad0378ef6c70897b.jpg)  
+Slider
+
+![](images/6b60ccaaf439eb96ea39c5a3b9fa413b510be46701455468c04e51538753bdc3.jpg)  
+Toothpaste
+
+![](images/0b3e191f36b3ed06901894a1493699ae31532b6433e29d9994adcf9e77502b7d.jpg)  
+Door Lock
+
+![](images/de166171b1fd2c1d1d14cfe9f71d84aeee36e3408be644fb0740baed93afadf8.jpg)  
+Toy Car
+
+![](images/90f777c39de74347453e303735bb358ab7dd2c4abc98c699069be3d12fd72216.jpg)  
+Clip
+
+![](images/1324e3603fe14c0ee8596a894f5b4b71efea4e5df255eef133e2c529b00767b4.jpg)  
+Fan
+
+![](images/79a5bcee9d78b800f3c96612ba744acd04d2d65ab5444af43ed720b70edc3bcb.jpg)  
+Rubber Band
+
+![](images/a93b776c01182baf642c844fd4003427bef1d20f803d1921c7b718ef3b4de6e3.jpg)  
+Magnet
+
+![](images/ef6d8ff60375a009ba23164350c9f70fa0b56f37d63ffd49a93da1a6e504579a.jpg)  
+Servo
+
+![](images/4d6dc56067eb39e38a2c6d664695e2645359b85f86b75c0ba77d4771ba20a87f.jpg)  
+Gear Assembly
+
+![](images/143b400b63356474d7a583349339b8c3fd2ca6429bf0a91836a0f9968ad6acbb.jpg)  
+Bottled Water
+
+![](images/db4e4ea50cb523a34663c3de3d8555feba8fc3e602d608de9919f4a00022692f.jpg)  
+Screw
+
+![](images/5d07bbf049c4ab67cb7cefb294f93b1adecd62b7b746b16a2b25c80a1f2688ff.jpg)  
+Spherical Bearing
+
+![](images/ae3b8ab1d44dca0c371de14f0f6e04fbffda0755c5a46cb2291463b29b4ca39e.jpg)  
+Motion
+
+![](images/30b520f00ef352421bb17433f5d67fdc49435e0a6a68430c5acb0b5e28c6e3da.jpg)  
+Button
+
+![](images/d477cbe0731ea4b43372c2b74c2807e6242f0ecfdb1ee561599aa52988b7e88b.jpg)  
+Clock
+
+![](images/32405e8d76a58bd525059e3dbd12ba4e95ee2af92a48352e8327c8bf0cb5a0a5.jpg)  
+Caster Wheel
+
+![](images/d020a461c07f5b611e5475dca37bd6451a9ee9fd00ee94a5502173746bfcce9d.jpg)  
+Hinge
+
+![](images/5d39dd0d1961330a708b001bbd0af0330e7e2274175ee736aa65e240d83b341e.jpg)  
+Ball
+
+![](images/ee9640bd724f2776c510eb025a226563c9518fd3ee2fc98b17e081ca1a1b534d.jpg)  
+Direction For Interaction  
+Figure 3. Interactions for understanding implicit physical laws in the Phys-AD dataset. We showcase various object interactions from the Phys-AD dataset, where different actions (indicated by motion directions) are used to explore and reason about the underlying physical properties and behaviors of each object. The colored arrows indicate the interaction directions and axes, highlighting how physical interactions reveal the implicit physics governing each object.
+
+![](images/b3f99f89cbed8591fccb3110e6a970871e8521d23d9bbcbaaea6ed729034642a.jpg)  
+Normal: The U Disk rotates freely over time.
+
+![](images/6768b2f107bdf1fc48650b414447219c6c19cc384203b8c07f0ee7c7436e2cb8.jpg)  
+Abnormal: The U Disk is stuck over time.  
+(a) Persistent Anomalies
+
+![](images/fc43f6a365f2a63ec36586f1a0cf725753bce008123e8aed6dd463dcf945a622.jpg)  
+Normal: The sticky roller rotates smoothly on the surface.
+
+![](images/0fd51d148225da32c20aed5e679cb20d87bc532ca66ea2e17693bcb16094c8c2.jpg)  
+Normal: The sticky roller functions properly at the beginning.  
+Abnormal: The sticky roller breaks at a certain point.  
+(b) Intermittent Anomalies  
+Figure 4. Categorization of anomalies based on persistence in the Phys-AD dataset. We show examples of normal and abnormal functioning in common objects, divided into two anomaly types: persistent and intermittent. (a) Persistent anomalies, such as continuous obstruction in the U Disk or permanent malfunction of the Sticky Roller, are visible throughout the operation. (b) In contrast, intermittent anomalies, like occasional jamming of the U Disk or breakage in the Sticky Roller after initial operation, only appear at specific points in time. This classification provides insight into both constant and sporadic failures in object interactions.
+
+and challenging, we introduce 47 distinct defect types, some of which just rely on single frame content and physical rules for anomaly reasoning, while the other part needs to combine the content of the whole video and physical priors together to judge whether the object is abnormal or not. Different objects information with their corresponding interaction methods and anomalies are listed in Table 2. Fig 3 shows the interac
+
+tions for understanding implicit physical laws in the Phys-AD dataset. Fig 4 provides two representative examples: one where a U Disk requires analyzing the entire video content to determine if there is an anomaly, and another where a sticky roller only requires a few frames from the video to judge whether the object is abnormal.
+
+![](images/c3ffd1eb93313cb723ec32e3219be946461890aafac571d6445aa6c70c56801c.jpg)  
+Figure 5. Data collection pipeline for the Phys-AD dataset. (a) Manipulation of a toothpaste tube using a UR5 robotic arm. (b) Manipulation of a U Disk and fan via servo and motor.
+
+![](images/e4add3a5da0de42ca6b80027e5c1fbdd5fd5651b741bdb4da4c922c84fbf789b.jpg)
+
+Table 2. Interaction methods and anomalies of the object categories in our Phys-AD dataset. Motor. means one object driven by motor. Multi. means one object with multiple anomalies.  
+
+<table><tr><td>Category</td><td>Interaction</td><td>Anomaly Types</td></tr><tr><td>Car</td><td>Drag, Slide</td><td>Different wheels stuck</td></tr><tr><td>Fan</td><td>Motor., Rotate</td><td>Stuck, Uneven Rotation, Vibration</td></tr><tr><td>Rolling Bearing</td><td>Motor., Rotate</td><td>Lack of friction</td></tr><tr><td>Spherical Bearing</td><td>Grab, Rotate</td><td>Internal block</td></tr><tr><td>Servo</td><td>Grab, Motor., Rotate</td><td>Angle restricted, Vibration, No calibration</td></tr><tr><td>Clip</td><td>Press</td><td>Unable to press down, Unable to rebound</td></tr><tr><td>U Disk</td><td>Grab, Motor., Rotate</td><td>Cover jam</td></tr><tr><td>Hinge</td><td>Grab, Rotate</td><td>Angle restricted, Shaft off</td></tr><tr><td>Sticky Roller</td><td>Grab, Pull</td><td>Detach, Unable to rotate</td></tr><tr><td>Caster Wheel</td><td>Slide</td><td>Axle axis stuck, Swivel axis stuck</td></tr><tr><td>Screw</td><td>Press, Rotate</td><td>Loosen, Unable to insert</td></tr><tr><td>Lock</td><td>Motor., Rotate</td><td>Latch jam, Loose</td></tr><tr><td>Gear</td><td>Motor., Rotate</td><td>Stuck, Not meshed, Multi.</td></tr><tr><td>Clock</td><td>Motor., Rotate</td><td>Pointer stops, Vibration</td></tr><tr><td>Slide</td><td>Grab, Slide</td><td>Detach, Shedding, Jam</td></tr><tr><td>Zipper</td><td>Grab, Close</td><td>Stuck, Unable to close</td></tr><tr><td>Button</td><td>Press</td><td>No light, Unable to press down, Unable to rebound, Multi.</td></tr><tr><td>Liquid</td><td>Grab, Shake</td><td>Water out, Foreign body</td></tr><tr><td>Rubber Band</td><td>Stretch</td><td>Crack</td></tr><tr><td>Ball</td><td>Pinch, Press</td><td>Insufficient gas, Leakage</td></tr><tr><td>Magnet</td><td>Grab, Press, Move</td><td>Degaussing, Shell detached</td></tr><tr><td>Toothpaste</td><td>Pinch, Press</td><td>Leakage</td></tr></table>
+
+# 3.2. Data Collection and Processing
+
+Data collection pipeline. Most of the data collection for the Phys-AD dataset is driven by manipulation-guided video sequences, captured using a UR5 robotic arm equipped with an RGB camera (see Fig. 5a). In order to reproduce real industrial scenes, we also include adjustable light sources in our data capture process. The RGB Camera, with a 1080p resolution and a frame rate of 60 FPS, providing high-quality video sequences. Some kinds of the objects like U Disk that are not suitable for the manipulation of robotics arms, are driven by the motor or servo (Fig. 5b). After the data collection, we used video editing software to crop out irrelevant frames and retain the complete interaction process.
+
+Table 3. Phys-AD dataset statistics. We denote the total number of frames and videos for each category as #Images and #Videos. Note that the Train split does not contain anomaly data.  
+
+<table><tr><td>Category</td><td>Train #Frames</td><td>Train #Videos</td><td>Test #Frames</td><td>Test #Videos</td><td>Test #Videos Nor.</td><td>Test #Videos Ab.</td><td>Obj. Types</td><td>Def. Types</td></tr><tr><td>Car</td><td>18,000</td><td>300</td><td>36,000</td><td>600</td><td>150</td><td>450</td><td>5</td><td>3</td></tr><tr><td>Fan</td><td>32,400</td><td>180</td><td>64,800</td><td>360</td><td>90</td><td>270</td><td>3</td><td>3</td></tr><tr><td>Rolling Bearing</td><td>3,600</td><td>60</td><td>3,600</td><td>60</td><td>30</td><td>30</td><td>1</td><td>1</td></tr><tr><td>Spherical Bearing</td><td>3,600</td><td>60</td><td>3,600</td><td>60</td><td>30</td><td>30</td><td>1</td><td>1</td></tr><tr><td>Servo</td><td>14,400</td><td>120</td><td>28,800</td><td>240</td><td>60</td><td>180</td><td>1</td><td>3</td></tr><tr><td>Clip</td><td>28,800</td><td>240</td><td>43,200</td><td>360</td><td>120</td><td>240</td><td>4</td><td>2</td></tr><tr><td>U Disk</td><td>14,400</td><td>240</td><td>14,400</td><td>240</td><td>120</td><td>120</td><td>4</td><td>1</td></tr><tr><td>Hinge</td><td>3,600</td><td>30</td><td>7,200</td><td>60</td><td>15</td><td>45</td><td>1</td><td>2</td></tr><tr><td>Sticky Roller</td><td>5,400</td><td>30</td><td>8,100</td><td>45</td><td>15</td><td>30</td><td>1</td><td>2</td></tr><tr><td>Caster Wheel</td><td>5,400</td><td>30</td><td>10,800</td><td>60</td><td>15</td><td>45</td><td>1</td><td>2</td></tr><tr><td>Screw</td><td>5,400</td><td>30</td><td>8,100</td><td>45</td><td>15</td><td>30</td><td>1</td><td>2</td></tr><tr><td>Lock</td><td>7,200</td><td>120</td><td>10,800</td><td>180</td><td>60</td><td>120</td><td>1</td><td>2</td></tr><tr><td>Gear</td><td>21,600</td><td>180</td><td>54,000</td><td>450</td><td>90</td><td>360</td><td>3</td><td>4</td></tr><tr><td>Clock</td><td>27,000</td><td>150</td><td>40,320</td><td>224</td><td>73</td><td>151</td><td>5</td><td>2</td></tr><tr><td>Slide</td><td>7,200</td><td>60</td><td>18,000</td><td>150</td><td>30</td><td>120</td><td>1</td><td>3</td></tr><tr><td>Zipper</td><td>14,400</td><td>120</td><td>21,600</td><td>180</td><td>60</td><td>120</td><td>2</td><td>2</td></tr><tr><td>Button</td><td>21,600</td><td>120</td><td>54,000</td><td>300</td><td>60</td><td>240</td><td>4</td><td>4</td></tr><tr><td>Liquid</td><td>5,400</td><td>30</td><td>8,100</td><td>45</td><td>15</td><td>30</td><td>1</td><td>2</td></tr><tr><td>Rubber Band</td><td>10,800</td><td>60</td><td>10,800</td><td>60</td><td>30</td><td>30</td><td>1</td><td>1</td></tr><tr><td>Ball</td><td>21,600</td><td>90</td><td>32,400</td><td>135</td><td>45</td><td>30</td><td>3</td><td>2</td></tr><tr><td>Magnet</td><td>10,800</td><td>60</td><td>16,200</td><td>90</td><td>30</td><td>60</td><td>2</td><td>2</td></tr><tr><td>Toothpaste</td><td>16,200</td><td>90</td><td>16,200</td><td>90</td><td>45</td><td>45</td><td>3</td><td>1</td></tr><tr><td>Total</td><td>298,800</td><td>2400</td><td>511,020</td><td>4,034</td><td>1,198</td><td>2,836</td><td>49</td><td>47</td></tr></table>
+
+# 3.3. Data Statistics
+
+Dataset sample distribution. Table 3 provides a detailed breakdown of the Phys-AD dataset, which includes information on category, the number of training videos and frames, the number of testing videos and frames, the distribution of normal and abnormal samples in the test set, and the number of object and defect types. The length of the video clips in the dataset ranges from 60 to 240 frames, ensuring to fully capture the interaction process in a short time. The dataset contains 2400 training videos and 4034 test videos spreading across 22 categories, 49 object types, and 47 defect types. The frames are extracted from the videos at a rate of 60 FPS. The test set includes 1,198 normal samples and 2,836 abnormal samples. Overall, the Phys-AD dataset covers a wide range of categories and contains a large scale of data, which helps to train more robust anomaly detection models and provides a reasonable evaluation setting.
+
+Interaction and defect types across categories. Table 2 details the defect types and corresponding interaction methods for each category in the Phys-AD dataset. Covering 22 categories and including 47 defect types, the dataset features various interactions (e.g., Rotate, Grab). Importantly, interaction methods or defect types with identical names can differ across categories—for example, "rotation" for a fan describes the motor-driven movement of its blades, while for a hinge it refers to the turning of its pages around the shaft under robotic control. Furthermore, to detect anomalies with complex physical properties in a human-like manner, we sometimes combine multiple interactions. For instance, when inspecting a screw, we first press it into the hole and then use a robotic arm to rotate it further. Overall, Phys-AD provides a diverse, physics-grounded anomaly detection framework that fosters the development of advanced methods for tackling real-world challenges.
+
+![](images/37b08d3a354aa271391e904fcbd765b6bb939dde01b0e2dacc73c68dd212e63c.jpg)  
+Figure 6. PhysAD-Agent: A Large Language Model (LLM)-powered system for physics anomaly detection label augmentation and automatic evaluation. This agent framework consists of two main components: (a) Rich Annotation Generation, where an LLM-based data agent generates detailed anomaly annotations based on video, anomaly state, prompt, and human-annotated physical priors, and (b) Automatic Evaluation, where an LLM-based evaluation agent assesses model predictions to calculate the Physics Anomaly Explanation (PAEval) metric.
+
+Labels. For the test set videos, we provide two types of labels. First, there are the common video-level labels: all anomalous videos in the test set are labeled as 1, while normal videos are labeled as 0. For evaluating visual-language models, we also provide text labels. Specifically, for each type of anomaly in the test set, we manually design a textual description label. This label includes both a description of the video content and a physical explanation of the reason for the anomaly. To ensure diversity in our text labels, we also use ChatGPT-4o for text augmentation. For further details on the labeling process, please refer to Fig 6a.
+
+# 4. Phys-AD Benchmark
+
+# 4.1. Problem Definition and Challenges
+
+We establish unsupervised and weakly supervised AD settings for Phys-AD, using unsupervised as the default in our problem definition.
+
+We formulate our Phys-AD setting as two steps.
+
+Step 1: Rules deduction Given a set of training objects  $\mathcal{T} = \{t_i\}_{i=1}^N$  from category  $c_i$ , we first use robotics arms and motors to interact with  $\mathcal{T}$  in corresponding methods  $I_i$ , and we get the interaction process in video format as  $V_i$ . Then, we feed the video sequence  $V_i$  and the corresponding category's physical prior  $P_i$  together to the deduction function  $f$  to obtain the normal rules  $r_i$ . After rules deduction for all the
+
+categories, we get the rules bank  $\mathcal{M} = \{r_i\}_{i=1}^N$ . Step 1 can be formulated as below:
+
+$$
+\mathcal {T} \Theta I _ {i} = V _ {i} \tag {1}
+$$
+
+$$
+f \left(V _ {i} \mid P _ {i}\right) = r _ {i} \tag {2}
+$$
+
+$$
+\mathcal {M} = \left\{r _ {i} \right\} _ {i = 1} ^ {N} \tag {3}
+$$
+
+$\Theta$  represents interaction between objects and robotics arms.  $N$  is the total number of categories.
+
+Step 2: Anomaly reasoning During test time, we first use robotics arms and motors to interact with  $\mathcal{T}$  from the test set in corresponding methods  $I_{i}$ , and we get the interaction process in video format as  $V_{i}$ . Then, we feed the video sequence  $V_{i}$  and rules  $r_i$  from step 1 to the reasoner  $R$  to predict the object's anomaly score  $S$ . Step 2 can be formulated as:
+
+$$
+\mathcal {T} \Theta I _ {i} = V _ {i} \tag {4}
+$$
+
+$$
+R \left(V _ {i} \mid r _ {i}\right) = S, r _ {i} \in \mathcal {M} \tag {5}
+$$
+
+Challenges. Key challenges include combining video information with objects' physical priors to deduct normal rules, capturing long-term temporal dependencies and fine-grained frame-level information, ensuring robust reasoning for anomaly prediction, and generalizing across diverse anomaly patterns. Addressing these challenges is essential for advancing anomaly detection in real-world environment.
+
+# 4.2. Evaluation Settings
+
+Unsupervised AD. Unsupervised AD is the most common IAD setting for existing IAD datasets and algorithms. Under unsupervised AD setting, the training set contains only normal video data, and the algorithm needs to capture the normal distribution of the data from the training set. In the test set, both normal and abnormal data are included, and the algorithm must distinguish between normal and abnormal data based on the distribution learned during training.
+
+Weakly-supervised AD. Our dataset is in the form of videos for complex anomaly detection in industrial scenarios. In the video context, it is inevitable that we need to discuss weakly supervised anomaly detection. Under weakly-supervised Phys-AD setting, 2~4 video-level labeled anomaly samples are sampled from all possible anomaly classes in the test set in our Phys-AD dataset. These sampled anomalies are then removed from the test set. It's worth noting that we only provide video level label in our test set. This is because most anomalies in our test set require reasoning based on the entire video content and physical knowledge of the objects.
+
+Video-Understanding AD. Video-Understanding models are another potential solution to our Phys-AD setting. Similar to unsupervised setting, we only provide normal videos during training. The visual language models need to provide explicit normal rules in language format during training. In the test phase, these video-understanding AD
+
+Table 4. Video-level AUROC (↑) result of 22 categories on Phys-AD dataset. We include Unsupervised, Weakly-supervised and Video-understanding methods. The best per-category result for each class of methods is highlighted in **bold**. 'ZS ImgB', 'V-ChatGPT', 'V-LLaMA', 'V-LLaVA' denote ZS ImageBind, Video-ChatGPT, Video-LLaMA and Video-LLaVA.  
+
+<table><tr><td rowspan="2">Category.</td><td colspan="5">Unsupervised</td><td colspan="3">Weakly-supervised</td><td colspan="6">Video-understanding</td></tr><tr><td>MPN [34]</td><td>MemAE [16]</td><td>MNAD.p [40]</td><td>MNAD.r [40]</td><td>SVM [44]</td><td>VADClip [51]</td><td>S3R [48]</td><td>MGFN [10]</td><td>LAVAD [56]</td><td>ZS Clip [42]</td><td>ZS ImgB [15]</td><td>V-ChatGPT [36]</td><td>V-LLaMA [59]</td><td>V-LLaVA [28]</td></tr><tr><td>Car</td><td>0.229</td><td>0.523</td><td>0.492</td><td>0.944</td><td>0.587</td><td>0.581</td><td>0.606</td><td>0.571</td><td>0.557</td><td>0.500</td><td>0.500</td><td>0.500</td><td>0.678</td><td>0.522</td></tr><tr><td>Fan</td><td>0.811</td><td>0.371</td><td>0.810</td><td>0.542</td><td>0.500</td><td>0.624</td><td>0.640</td><td>0.542</td><td>0.510</td><td>0.500</td><td>0.500</td><td>0.549</td><td>0.592</td><td>0.611</td></tr><tr><td>Rolling Bearing</td><td>0.353</td><td>0.044</td><td>0.352</td><td>0.800</td><td>0.933</td><td>0.589</td><td>0.601</td><td>0.680</td><td>0.532</td><td>0.500</td><td>0.500</td><td>0.300</td><td>0.500</td><td>0.500</td></tr><tr><td>Spherical Bearing</td><td>0.113</td><td>0.092</td><td>0.962</td><td>0.813</td><td>0.650</td><td>0.500</td><td>0.682</td><td>0.583</td><td>0.435</td><td>0.500</td><td>0.500</td><td>0.450</td><td>0.550</td><td>0.500</td></tr><tr><td>Servo</td><td>0.364</td><td>0.445</td><td>0.975</td><td>0.878</td><td>0.500</td><td>0.518</td><td>0.592</td><td>0.556</td><td>0.502</td><td>0.500</td><td>0.500</td><td>0.506</td><td>0.683</td><td>0.464</td></tr><tr><td>Clip</td><td>0.535</td><td>0.443</td><td>0.630</td><td>0.333</td><td>0.500</td><td>0.412</td><td>0.561</td><td>0.563</td><td>0.516</td><td>0.500</td><td>0.500</td><td>0.669</td><td>0.556</td><td>0.458</td></tr><tr><td>U Disk</td><td>0.240</td><td>0.617</td><td>0.609</td><td>0.940</td><td>0.500</td><td>0.530</td><td>0.549</td><td>0.570</td><td>0.513</td><td>0.500</td><td>0.500</td><td>0.565</td><td>0.575</td><td>0.500</td></tr><tr><td>Hinge</td><td>0.769</td><td>0.870</td><td>0.705</td><td>0.895</td><td>0.500</td><td>0.737</td><td>0.561</td><td>0.550</td><td>0.564</td><td>0.500</td><td>0.500</td><td>0.500</td><td>0.692</td><td>0.500</td></tr><tr><td>Sticky Roller</td><td>0.967</td><td>0.967</td><td>0.451</td><td>0.936</td><td>0.500</td><td>0.542</td><td>0.835</td><td>0.669</td><td>0.266</td><td>0.500</td><td>0.500</td><td>0.450</td><td>0.544</td><td>0.467</td></tr><tr><td>Caster Wheel</td><td>0.364</td><td>0.523</td><td>0.271</td><td>0.508</td><td>0.500</td><td>0.587</td><td>0.676</td><td>0.676</td><td>0.615</td><td>0.500</td><td>0.500</td><td>0.444</td><td>0.642</td><td>0.500</td></tr><tr><td>Screw</td><td>0.522</td><td>0.567</td><td>0.680</td><td>0.547</td><td>0.500</td><td>0.500</td><td>0.657</td><td>0.541</td><td>0.688</td><td>0.500</td><td>0.500</td><td>0.472</td><td>0.256</td><td>0.550</td></tr><tr><td>Lock</td><td>0.563</td><td>0.597</td><td>0.430</td><td>0.641</td><td>0.733</td><td>0.452</td><td>0.523</td><td>0.626</td><td>0.341</td><td>0.500</td><td>0.500</td><td>0.279</td><td>0.494</td><td>0.500</td></tr><tr><td>Gear</td><td>0.529</td><td>0.510</td><td>0.652</td><td>0.694</td><td>0.500</td><td>0.519</td><td>0.580</td><td>0.544</td><td>0.603</td><td>0.500</td><td>0.500</td><td>0.544</td><td>0.603</td><td>0.517</td></tr><tr><td>Clock</td><td>0.340</td><td>0.542</td><td>0.395</td><td>0.587</td><td>0.500</td><td>0.500</td><td>0.549</td><td>0.509</td><td>0.500</td><td>0.500</td><td>0.500</td><td>0.501</td><td>0.360</td><td>0.500</td></tr><tr><td>Slide</td><td>0.517</td><td>0.962</td><td>0.917</td><td>0.784</td><td>0.500</td><td>0.531</td><td>0.611</td><td>0.568</td><td>0.425</td><td>0.500</td><td>0.500</td><td>0.562</td><td>0.567</td><td>0.179</td></tr><tr><td>Zipper</td><td>0.815</td><td>0.592</td><td>0.829</td><td>0.421</td><td>0.500</td><td>0.504</td><td>0.636</td><td>0.633</td><td>0.535</td><td>0.500</td><td>0.500</td><td>0.547</td><td>0.489</td><td>0.500</td></tr><tr><td>Button</td><td>0.853</td><td>0.365</td><td>0.660</td><td>0.568</td><td>0.500</td><td>0.627</td><td>0.515</td><td>0.566</td><td>0.439</td><td>0.500</td><td>0.500</td><td>0.515</td><td>0.517</td><td>0.360</td></tr><tr><td>Liquid</td><td>0.184</td><td>0.700</td><td>0.671</td><td>0.831</td><td>0.500</td><td>0.542</td><td>0.595</td><td>0.793</td><td>0.504</td><td>0.500</td><td>0.500</td><td>0.410</td><td>0.278</td><td>0.217</td></tr><tr><td>Rubber Band</td><td>0.374</td><td>0.368</td><td>0.366</td><td>0.307</td><td>0.567</td><td>0.482</td><td>0.623</td><td>0.604</td><td>0.511</td><td>0.500</td><td>0.500</td><td>0.517</td><td>0.450</td><td>0.450</td></tr><tr><td>Ball</td><td>0.543</td><td>0.383</td><td>0.728</td><td>0.687</td><td>0.500</td><td>0.500</td><td>0.667</td><td>0.567</td><td>0.603</td><td>0.500</td><td>0.500</td><td>0.562</td><td>0.636</td><td>0.533</td></tr><tr><td>Magnet</td><td>0.671</td><td>0.464</td><td>0.691</td><td>0.438</td><td>0.500</td><td>0.500</td><td>0.548</td><td>0.719</td><td>0.502</td><td>0.500</td><td>0.500</td><td>0.683</td><td>0.300</td><td>0.400</td></tr><tr><td>Toothpaste</td><td>0.587</td><td>0.889</td><td>0.520</td><td>0.631</td><td>0.500</td><td>0.500</td><td>0.686</td><td>0.711</td><td>0.562</td><td>0.500</td><td>0.500</td><td>0.376</td><td>0.550</td><td>0.467</td></tr><tr><td>Average</td><td>0.511</td><td>0.538</td><td>0.627</td><td>0.669</td><td>0.544</td><td>0.535</td><td>0.613</td><td>0.606</td><td>0.510</td><td>0.500</td><td>0.500</td><td>0.496</td><td>0.523</td><td>0.463</td></tr></table>
+
+methods have to truly understand the videos and predict the anomalies based on the language rules withdrawn from the training phase.
+
+# 4.3. Evaluation Metrics
+
+Physics anomaly accuracy metrics. We use the Area Under the Receiver Operating Characteristic Curve (AUROC) to evaluate video-level anomaly detection performance. We also report the average precision (AP), i.e. the area under the video-level precision-recall curve and the acc (ACCURACY), following previous works [47, 49].
+
+Physics Anomaly Explanation (PAEval) metric. Directly utilizing existing VLMs [59][36] to understand the videos in our dataset and detect the anomalies is a potential solution to our challenge. The key point is: Can existing VLMs truly understand physical rules and reason in a right way? Specifically designed for video-understanding methods, we introduce a new metric named Physics Anomaly Explanation (PAEval) metric. To be more specific, PAEval evaluate the anomaly detection performance of algorithms based on VLMs from three different perspectives: classification, description, and explanation. Classification refers to the traditional anomaly detection metrics like AUROC, etc. Inspired by the works [11, 61], PAEval also introduces two additional evaluation metrics: description and physical explanation. Description refers to the model's ability to
+
+correctly describe the content of the video, used to assess whether the model has the capability to describe physical phenomena. Explanation refers to the model's ability to correctly explain the physical causes of anomalies in the video, used to evaluate the VLM's reasoning ability. To provide labels for description and explanation metrics, we manually labeled each type of defects from all categories and performed data augmentation by ChatGPT to ensure the robustness of the detection. The whole pipeline of PAEval metric is depicted in Fig 6.
+
+# 5. Benchmarking Results
+
+# 5.1. Benchmark Methods Selection
+
+For the Phys-AD setting, we select popular and reproducible video anomaly detection methods across unsupervised, weakly-supervised, and video-understanding setting. In the unsupervised setting, we focus on reconstruction, prediction, and embedding-based models like MemAE [16] and MNAD [40], which use memory modules to enhance anomaly discrimination. For weakly-supervised anomaly detection, we adopt methods such as VadCLIP [51] and MGFN [10], which use feature magnitudes and vision-language associations. In video-understanding, we evaluate video-language models like Video-ChatGPT [36] and image-language models like CLIP [42] to predict detailed anomaly descriptions and scores through cross-modal embeddings.
+
+Table 5. Physics Anomaly Explanation (PAEval) metric results on our Phys-AD dataset.  
+
+<table><tr><td>Methods</td><td>Classification (↑)</td><td>Description (↑)</td><td>Explanation (↑)</td></tr><tr><td>LAVAD[56]</td><td>0.510</td><td>0.157</td><td>0.000</td></tr><tr><td>Video-ChatGPT[36]</td><td>0.496</td><td>0.131</td><td>0.160</td></tr><tr><td>Video-LLaMA[59]</td><td>0.523</td><td>0.137</td><td>0.303</td></tr><tr><td>Video-LLaVA[28]</td><td>0.463</td><td>0.219</td><td>0.282</td></tr></table>
+
+Code and Experiment details. We provide code and toolkit for our dataset and benchmark. More experiment details are listed in Appendix.
+
+# 5.2. Main Findings
+
+Overall anomaly detection benchmarking results. Table 4 shows that existing video anomaly detection and videowhildunderstanding methods achieve limited performance on the Phys-AD dataset, with the highest AUROC only reaching  $66.9\%$  for MNAD.r [40]. This result underscores the heightened complexity of Phys-AD compared to existing datasets, highlighting a domain shift that challenges current industrial anomaly detection algorithms, which are often tuned to visually distinct anomaly patterns in single frames rather than complex temporal or physical cues.
+
+Anomaly detection via unsupervised and weakly supervised methods. Our experiment includes unsupervised methods (e.g., MemAE [16], MPN [34], MNAD [40]) and weakly supervised methods (e.g., S3R [48], MGFN [10], VADClip [51]). Unsupervised methods like MNAD.p [40] performed better on temporal anomalies (81.0% for fan, 68.0% for screw) by leveraging prediction-based approaches, which excel in scenarios requiring temporal understanding. Weakly supervised methods improve baseline scores across complex classes, preventing extreme low scores in challenging categories such as spherical bearing. However, weakly supervised methods show marginally lower performance in simpler anomaly classes, indicating trade-offs introduced by anomaly samples in training.
+
+Unsupervised anomaly detection method performance by category. Among unsupervised methods, MemAE [16] achieves high AUROC on objects with spatial anomalies (e.g., sticky roller,  $96.7\%$ ) but underperforms in temporal anomaly classes like rolling and spherical bearings, where it achieves only  $4.4\%$  and  $9.2\%$ , respectively. MNAD [40] improves temporal sensitivity by incorporating temporal prototypes, achieving better scores for bearings but still struggling with complex anomaly types, revealing limitations in purely prototype-based approaches.
+
+Anomaly detection via Video-Language Models. VLM-based methods (e.g., Video-ChatGPT [36], VideoLLaMA [59]) struggle, with top-performing VideoLLaMA [59] reaching only  $52.3\%$  average AUROC. Performance is impacted by reliance on pre-trained weights that are not optimized for physics-grounded video content, evident
+
+from low AUROC scores in categories with nuanced physical dynamics, such as hinges and screws. Further, PAEval results suggest that these models lack effective reasoning about object physical dynamics and behaviors influenced by physical forces, underscoring a gap between VLM capabilities and the demands of IAD tasks.
+
+Anomaly explanation via Video-Language Models. Table 5 reports PAEval metric results, with the best-performing VLM achieving only  $21.9\%$  in description and  $30.3\%$  in explanation. These findings emphasize that current VLMs lack the depth in physical reasoning and temporal coherence required for understanding real-world physics-based scenarios, which our Phys-AD dataset demands.
+
+Summary. Overall, the results reveal Phys-AD's unique challenge in requiring both high spatial detail and temporal comprehension, areas where existing methods and models underperform. This analysis points to the need for future research in models that integrate temporal reasoning with physics-based anomaly detection.
+
+# 6. Conclusion
+
+In this paper, we introduce the first industrial anomaly detection task focusing on real-world scenarios where physical understanding and reasoning are essential for anomaly detection. We present the Physics Anomaly Detection (Phys-AD) dataset, a large-scale, physics-grounded video dataset with over 6400 videos across 22 categories and 49 object types interacting with robotic systems, capturing 47 anomaly types that necessitate visual and physical understanding. We assess Phys-AD, highlighting the lack of baseline approaches for high-level reasoning in anomaly detection. Additionally, we propose the Physics Anomaly Explanation (PAEval) metric to evaluate visual language models (VLMs) on physics-based reasoning. Experiments show that current VLMs fall short of human-level understanding in physics-based anomaly scenarios. This work marks a milestone for industrial anomaly detection, promoting physics-grounded reasoning in complex industrial settings.
+
+Limitation and future work. Although our Phys-AD dataset provides a large variety of objects with diverse physical properties and various types of interaction methods, we plan to add even more diverse interaction methods and objects in the future to better meet the demands of complex real-world industrial scenarios. Due to the significant differences between our Phys-AD dataset and current industrial anomaly detection and video anomaly detection datasets, most existing anomaly detection algorithms cannot be directly applied to our dataset. In the future, we will test more algorithms on our Phys-AD dataset and provide experimental results across more settings like zero-shot, few-shot, semi-supervised settings, etc.
+
+# References
+
+[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015. 3  
+[2] Fabien Baradel, Natalia Neverova, Julien Mille, Greg Mori, and Christian Wolf. Cophy: Counterfactual learning of physical dynamics. arXiv preprint arXiv:1909.12000, 2019.3  
+[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9584-9592, 2019. 3  
+[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. MVtec AD – A comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, 2019. 2  
+[5] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130 (4):947-969, 2022. 3  
+[6] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. SCITEPRESS - Science and Technology Publications, 2022. 3  
+[7] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The MVTec 3d-AD dataset for unsupervised 3d anomaly detection and localization. In VISAPP, 2022. 2  
+[8] Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. In Proceedings of the Asian Conference on Computer Vision, pages 3586-3602, 2022. 3  
+[9] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 2  
+[10] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In Proceedings of the AAAI conference on artificial intelligence, pages 387-395, 2023. 7, 8  
+[11] Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, et al. Uncovering what why and how: A comprehensive benchmark for causation understanding of video anomaly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18793-18803, 2024. 7  
+[12] Jiafei Duan, Samson Yu, and Cheston Tan. Space: A simulator for physical interactions and causal learning in 3d environments. In ICCV, pages 2058-2063, 2021. 3  
+[13] Jiafei Duan, Samson Yu, Soujanya Poria, Bihan Wen, and Cheston Tan. Pip: Physical interaction prediction via mental
+
+simulation with span selection. In ECCV, pages 405-421. Springer, 2022. 3  
+[14] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14009-14018, 2021. 3  
+[15] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 7  
+[16] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1705-1714, 2019. 3, 7, 8  
+[17] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 3  
+[18] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K Roy-Chowdhury, and Larry S Davis. Learning temporal regularity in video sequences. In CVPR, 2016. 3  
+[19] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 3  
+[20] Steeven Janny, Fabien Baradel, Natalia Neverova, Madiha Nadri, Greg Mori, and Christian Wolf. Filtered-cophy: Unsupervised learning of counterfactual physics in pixel space. In ICLR, 2022. 3  
+[21] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT), pages 66-71, 2021. 2, 3  
+[22] Huaizu Jiang, Xiaojian Ma, Weili Nie, Zhiding Yu, Yuke Zhu, and Anima Anandkumar. Bongard-hoi: Benchmarking few-shot visual reasoning for human-object interactions. In CVPR, pages 19056-19065, 2022. 3  
+[23] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 3  
+[24] Hyekang Kevin Joo, Khoa Vo, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 3230-3234. IEEE, 2023. 3  
+[25] Federico Landi, Cees GM Snok, and Rita Cucchiara. Anomaly locality in video surveillance. arXiv preprint arXiv:1901.10364, 2019. 3  
+[26] Ting Li, Charles Dubout, Emma K Wampler, Steven Yantis, Donald Geman, et al. Comparing machines and humans on a visual categorization test. 2011. 3
+
+[27] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. arXiv preprint, 2023. 2, 3  
+[28] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 7, 8  
+[29] Jiaqi Liu, Guoyang Xie, Rui Chen, et al. Real3d-ad: A dataset of point cloud anomaly detection. NeurIPS, 2023. 2  
+[30] Jiaqi Liu, Guoyang Xie, Ruitao Chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. Real3dad: A dataset of point cloud anomaly detection. Advances in Neural Information Processing Systems, 36, 2024. 3  
+[31] Kun Liu and Huadong Ma. Exploring background-bias for anomaly detection in surveillance videos. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1490-1499, 2019. 3  
+[32] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection-a new baseline. In CVPR, 2018. 3  
+[33] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588-13597, 2021. 3  
+[34] Hui Lv, Chen Chen, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Learning normal dynamics in videos with meta prototype network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15425-15434, 2021. 7, 8  
+[35] Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8022-8031, 2023. 3  
+[36] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 7, 8  
+[37] Domingo Mery, Vladimir Riffo, Uwe Zscherpel, German Mondragon, Ivan Lillo, Irene Zuccar, Hans Lobel, and Miguel Carrasco. Gdxray: The database of x-ray images for nondestructive testing. Journal of Nondestructive Evaluation, 34 (4):42, 2015. 3  
+[38] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adj: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE). IEEE, 2021. 3  
+[39] Weili Nie, Zhiding Yu, Lei Mao, Ankit B Patel, Yuke Zhu, and Animashree Anandkumar. Bongard-logo: A new benchmark for human-level concept learning and reasoning. In NeurIPS, 2020. 3  
+[40] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceed-
+
+ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14372-14381, 2020. 7, 8  
+[41] Yujiang Pu, Xiaoyu Wu, and Shengjin Wang. Learning prompt-enhanced context features for weakly-supervised video anomaly detection. arXiv preprint arXiv:2306.14451, 2023. 3  
+[42] 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. 3, 7  
+[43] Binyi Su, Zhong Zhou, and Haiyong Chen. Pvel-ad: A large-scale open-world dataset for photovoltaic cell anomaly detection. IEEE Transactions on Industrial Informatics, 19 (1):404-413, 2023. 3  
+[44] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, 2018. 7  
+[45] Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. In Proceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 5541-5550, 2023. 3  
+[46] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jianning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection, 2024. 3  
+[47] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032-8041, 2023. 7  
+[48] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 7, 8  
+[49] Peng Wu, Jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 322-339. Springer, 2020. 3, 7  
+[50] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection. arXiv preprint arXiv:2311.07042, 2023. 3  
+[51] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. arXiv preprint arXiv:2308.11681, 2023.3,7,8  
+[52] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. Detecting anomalous events in videos by learning deep representations of appearance and motion. Computer Vision and Image Understanding, 156:117-127, 2017. 3  
+[53] Xiaohao Xu, Yunkang Cao, Yongqi Chen, Weiming Shen, and Xiaonan Huang. Customizing visual-language foundation models for multi-modal anomaly detection and reasoning. arXiv preprint arXiv:2403.11083, 2024. 3
+
+[54] Zhiwei Yang, Jing Liu, Zhaoyang Wu, Peng Wu, and Xiaotao Liu. Video event restoration based on keyframes for video anomaly detection. In CVPR, 2023. 3  
+[55] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14744-14754, 2022. 3  
+[56] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. arXiv preprint arXiv:2404.01014, 2024. 3, 7, 8  
+[57] Aimen Zerroug, Mohit Vaishnav, Julien Colin, Sebastian Musslick, and Thomas Serre. A benchmark for compositional visual reasoning. arXiv preprint arXiv:2206.05379, 2022. 3  
+[58] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In CVPR, 2019. 3  
+[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 7, 8  
+[60] Huaxin Zhang, Xiang Wang, Xiaohao Xu, Xiaonan Huang, Chuchu Han, Yuehuan Wang, Changxin Gao, Shanjun Zhang, and Nong Sang. Glancevad: Exploring glance supervision for label-efficient video anomaly detection. arXiv preprint arXiv:2403.06154, 2024. 3  
+[61] Huaxin Zhang, Xiaohao Xu, Xiang Wang, Jialong Zuo, Chuchu Han, Xiaonan Huang, Changxin Gao, Yuehuan Wang, and Nong Sang. Holmes-vad: Towards unbiased and explainable video anomaly detection via multi-modal llm. arXiv preprint arXiv:2406.12235, 2024. 3, 7  
+[62] Yiheng Zhang, Yunkang Cao, Xiaohao Xu, and Weiming Shen. Logic: An llm-driven framework for logical anomaly detection. IEEE Transactions on Automation Science and Engineering, pages 1-0, 2024. 3  
+[63] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In CVPR, 2019. 3  
+[64] Qiang Zhou, Weize Li, Lihan Jiang, Guoliang Wang, Guyue Zhou, Shanghang Zhang, and Hao Zhao. Pad: A dataset and benchmark for pose-agnostic anomaly detection, 2023. 3  
+[65] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In ECCV, 2022. 2  
+[66] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation, 2022. 3

2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18674aaffabf880257995cbcd8ef3216381832b284b7184dc4fa30f9572bfb4e
+size 800142

2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/9698d3f1-26f6-4c2e-aef0-60a71f074ba9_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6450f755868e2711934435f2e2b2e3d1d32933181dd0fe57f294b21fd65944f3
+size 5366530

2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/full.md

@@ -0,0 +1,471 @@
+# Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models
+
+Jiacong Xu $^{1*}$  Shao-Yuan Lo $^{2}$  Bardia Safaei $^{1}$  Vishal M. Patel $^{1}$  Isht Dwivedi $^{2}$ $^{1}$ Johns Hopkins University  $^{2}$ Honda Research Institute USA
+
+{jxu155, bsafaei1, vpatel36}@jhu.edu {shao-yuan.lo, idwivedi}@honda-ri.com
+
+# Abstract
+
+Zero-Shot Anomaly Detection (ZSAD) is an emerging AD paradigm. Unlike the traditional unsupervised AD setting that requires a large number of normal samples to train a model, ZSAD is more practical for handling data-restricted real-world scenarios. Recently, Multimodal Large Language Models (MLLMs) have shown revolutionary reasoning capabilities in various vision tasks. However, the reasoning of image abnormalities remains underexplored due to the lack of corresponding datasets and benchmarks. To facilitate research in AD & reasoning, we establish the first visual instruction tuning dataset, Anomaly-Instruct-125k, and the evaluation benchmark, VisA-D&R. Through investigation with our benchmark, we reveal that current MLLMs like GPT-4o cannot accurately detect and describe fine-grained anomalous details in images. To address this, we propose Anomaly-OneVision (Anomaly-OV), the first specialist visual assistant for ZSAD and reasoning. Inspired by human behavior in visual inspection, Anomaly-OV leverages a Look-Twice Feature Matching (LTFM) mechanism to adaptively select and emphasize abnormal visual tokens. Extensive experiments demonstrate that Anomaly-OV achieves significant improvements over advanced generalist models in both detection and reasoning. Extensions to medical and 3D AD are provided for future study. The link to our project page: https://xujiacong.github.io/Anomaly-OV/
+
+# 1. Introduction
+
+Visual Anomaly Detection (AD) is a well-established task in computer vision, extensively applied in scenarios such as industrial defect inspection [2, 3, 5, 12, 35, 69, 76, 77, 92, 98] and medical image diagnosis [1, 24, 31, 36, 89, 90, 103, 106]. In the traditional unsupervised AD (a.k.a. one-class AD) setting, models learn the distribution of normal visual features from normal samples and are required to identify anomaly samples during inference. While recent advance
+
+![](images/87ca03e000aaeacd11d7caf00c0f3e4662be350fb651c38b989f04967522b4d5.jpg)  
+Figure 1. Visualization of the image-level AUROC comparison between our Anomaly-OV and current state-of-the-art ZSAD methods (WinCLIP [38], AnoVL [19], AnomalyCLIP [110], AdaCLIP [6]). Notably, our zero-shot performance on VisA even surpasses most recent advances in the few-shot setting [28, 51, 112].
+
+ments [9, 25, 32–34, 37, 42, 82, 84, 95, 97, 104] have significantly improved the detection performance, these approaches assume the availability of a substantial number of normal samples. However, this assumption becomes impractical in certain scenarios due to strict data privacy policies and the significant human effort required for data classification, sometimes involving experts or specialists. Therefore, Zero-Shot Anomaly Detection (ZSAD) is emerging as a popular research direction, leading to the development of many innovative methods [6, 17, 27, 38, 43, 52, 78, 79, 110, 113].
+
+Recent advances in Multimodal Large Language Models (MLLMs) [7, 15, 44, 45, 47, 48, 57, 58, 111] have shown revolutionary reasoning capabilities in various vision tasks [14, 29, 67, 70, 80, 91, 94, 107, 109]. However, the rea
+
+soning of image abnormalities has not been explored due to the challenges of collecting large-scale datasets and establishing benchmarks. Existing methods simply predict the likelihood of an anomaly without providing rationales [6, 11, 19, 38, 110]. In contrast, for better interpretability, robustness, and trustworthiness, people would expect models to explain why an image is considered anomalous and provide visual evidence. Interestingly, we find that recent advanced MLLMs, such as GPT-4o [72], fall short in AD & reasoning. As shown in Figure 2, while the detection is correct, the explanation from GPT-4o lacks accuracy, indicating a gap in a comprehensive understanding of the anomaly.
+
+To expedite research in AD & reasoning, we establish the first visual instruction tuning dataset, Anomaly-Instruct-125k, and the evaluation benchmark, VisA-D&R, through intensive human efforts. After evaluating current generalist MLLMs, we observe that these models fail to accurately detect and describe fine-grained anomalous details in images. To address this, we propose Anomaly-OneVision (Anomaly-OV), the first specialist visual assistant for ZSAD and reasoning. Unlike existing ZSAD methods [6, 11, 19, 38, 110], Anomaly-OV directly learns object-awareness abnormality embeddings in feature space using only the visual encoder. Inspired by human behavior in visual inspection, Anomaly-OV employs a Look-Twice Feature Matching (LTFM) mechanism to assist its LLM in adaptively selecting and emphasizing the most suspicious abnormal visual tokens.
+
+Extensive experiments demonstrate that Anomaly-OV achieves significant improvements over advanced generalist models in both detection and reasoning. Extended results of Anomaly-OV, from applications in industrial defect detection to 3D inspection and medical image diagnosis, are provided for future study. With precise descriptions and rationales of visual anomalies, our model can infer potential causes (see Figure 2), assess current impacts, and offer improvement suggestions, positioning itself as a reliable assistant for visual inspection. Our contributions are in two folds:
+
+- We establish the first visual instruction tuning dataset and benchmark for anomaly detection and reasoning.  
+- We propose the first specialist visual assistant with state-of-the-art performance for this new impactful domain.
+
+# 2. Related Work
+
+Multimodal Large Language Models. Vision-Language Models (VLMs), such as CLIP [73], exhibit robust zero-shot classification capabilities and have been applied to a range of downstream vision tasks [10, 53, 56, 61, 86]. Combining a VLM's vision encoder and an LLM [20, 62, 74], MLLMs [15, 48, 49, 57, 111] enable text-based interactions related to visual content. MLLMs have shown remarkable reasoning capability, particularly when incorporated with prompting strategies such as Chain-of-Thought [66, 88, 105]. Recent studies have harnessed MLLMs to provide reasoning for
+
+![](images/13bec8769ce795b7c1bf630673506d45bbe937ad7c185058bb030de265045c11.jpg)
+
+![](images/25d9ca1d82563a8f93cbf3b8f88b477c4c9f7175c783c82e46e24750d5b1726d.jpg)
+
+GT: The top-left candle shows a missing chunk of wax and an irregular indentation near the wick.
+
+Do you see any anomalies in the image?
+
+![](images/fa1ec3fecb2f2f7a21fcf15c025d1cc5e0202f2e27fadac7851d6ee400a6f176.jpg)
+
+![](images/1028b0d63a08272fee646e1912e0f6c44eeb5805de96a4139539c94247c752f7.jpg)
+
+The candle in the top left appears to have a damaged or missing wick, which is unusual compared to the other candles.
+
+![](images/01126e20ac9b600e067b4454d12d690d857d9b915c2df3b957d02dc35024e36f.jpg)  
+Figure 2. Industrial image anomaly reasoning results from GPT-4o [72] and our Anomaly-OV. The responses for fine-grained anomaly reasoning are highlighted, with the ground truth given for reference.
+
+Yes, there is an anomaly in the top left candle. It has a small hole or crack near the wick, which is not present in the other candles.
+
+What is the cause for the anomaly?
+
+![](images/0853c89fa1b7fdb473555b96d7d47ac50e03bb39bccc3193e3b271babe46203d.jpg)
+
+![](images/ec51fc591b957718275465597f05c2ab9414421124c387006f6117f901285a47.jpg)
+
+The anomaly could be due to manufacturing defects, such as improper molding or cooling, or it might result from external damage like a puncture.
+
+downstream tasks, e.g., video anomaly detection [67, 94], affective computing [14, 29], and visual commonsense reasoning [109], revealing more interpretability.
+
+Unsupervised Anomaly Detection. Due to the scarcity and difficulty of collecting anomalous data, researchers often focus on the unsupervised AD setting, which exclusively uses normal data to train an AD model. Earlier studies, such as reconstruction-based [63, 69, 99], student-teacher [18, 85, 102], and augmentation-based [46] approaches, assume a large amount of normal data is available. These traditional approaches are less practical when data are restricted or expensive, such as in the medical domain.
+
+Zero-Shot Anomaly Detection. Unlike unsupervised AD [35, 77] and few-shot AD [22, 28, 36, 51, 112], ZSAD models directly access the likelihood of abnormality for a given image without requiring data specific to the target object. Existing works [6, 19, 38, 110] accomplish ZSAD by comparing visual and textual features encoded by visual and text encoders of CLIP and constructing their positive (anomaly) and negative (normal) prompts in the format of:
+
+$$
+\begin{array}{l} \mathcal {P} ^ {+} = \left[ V _ {1} \right] \left[ V _ {2} \right] \dots \left[ V _ {n} \right] [ o b j e c t ] \\ \mathcal {P} ^ {-} = \left[ W _ {1} \right] \left[ W _ {2} \right] \dots \left[ W _ {n} \right] [ o b j e c t ] \\ \end{array}
+$$
+
+where  $V_{i}$  and  $W_{i}$  are handcrafted or learnable tokens, and object refers to the word object or the class name of the object. However, simply utilizing object to represent all kinds of objects cannot capture the class-awareness abnormality types. Also, for an intelligent visual assistant, the images should be totally blind to the user (object-agnostic).
+
+![](images/c9b0b96175b01f00b0c7a0ecc318ecfbc5b060ab17aea4571e60fc25539fed13.jpg)  
+Figure 3. Overview of the Anomaly-0V architecture. It consists of two training stages: (1) professional training for the anomaly expert, and (2) visual instruction tuning for anomaly detection and reasoning. Text and visual tokens are distinguished by different colors.
+
+# 3. Method
+
+# 3.1. Preliminary
+
+Training an MLLM from scratch demands extensive data and computational resources to align the visual and textual embedding spaces and develop robust instruction-following capabilities. Recent studies [23, 83, 93] reveal that pretrained MLLMs function as generalists, possessing a broad knowledge base but underperforming in specialized domains. Therefore, our goal is to introduce an auxiliary specialist or expert model designed to guide the generalist in selecting and utilizing critical visual tokens. This approach circumvents the need for large-scale pre-training while preserving the generalization capacity of the original model.
+
+We choose  $LLaVA-OneVision$  [44] as our base MLLM because it is open-sourced and performs similarly to other commercial models.  $LLaVA-OneVision$  follows the model architectures for LLaVA family [47, 57-59] and other generic MLLMs, which typically consist of three major components: Visual Encoder, Projector, and LLM. The visual encoder [73, 100] extracts the visual information from the raw images, the projector aligns the spaces of visual features with the word embedding, and the LLM is responsible for textual instruction processing and complex reasoning. Since the image resolution for CLIP pre-training is fixed,  $LLaVA-OneVision$  leverages AnyRes with pooling strategy to scale
+
+up the input raw image resolution. Specifically, the high-resolution images are divided into a prototyped number of crops, and the visual encoder independently processes the image crops before final spatial pooling.
+
+# 3.2. Architecture Overview
+
+With the same image-splitting strategy AnyRes as LLaVA-OneVision, the input high-resolution image is split into several crops, and the new image set can be written as:
+
+$$
+\mathcal {I} = \left\{I _ {0}, I _ {1}, I _ {2}, \dots , I _ {n - 1} \right\} \tag {1}
+$$
+
+where  $I_0$  is the resized original image and  $I_{j\neq 0}$  refers to the image crops. As shown in Figure 3, the image set  $\mathcal{I}$  will be processed by the visual encoder  $\mathcal{F}_{\theta}$  to generate the final visual features  $\{\mathbf{v}_j^o\}$ . Similar to AnomalyCLIP [110], we store the outputs for four selected layers in the ViT [21] to capture the image representations from different levels and apply four adapters to compress the feature dimension. Then, the extracted visual features can be written as:
+
+$$
+\mathbf {v} _ {j} ^ {i} = \mathcal {F} _ {\theta} ^ {i} (I _ {j}) \tag {2}
+$$
+
+where  $i$  denotes the  $i$ -th level and  $j$  refers to the index of corresponding image in  $\mathcal{I}$ . These multi-level features have been demonstrated to be effective in capturing fine-grained local semantics by recent works [6, 28, 110].
+
+![](images/f22e345a2ef3107a4d55a4c39a16623ba30cea442cceb34a735f0e07cee87875.jpg)  
+Figure 4. Simulation of visual anomaly inspection by humans.
+
+The large-scale pre-trained CLIP models align the projection spaces of the textual and visual encoder. Therefore, the encoded image features already contain the class information required by ZSAD. To avoid human involvement in object classification and reduce the model complexity, we remove the heavy text encoder commonly utilized in existing works and let the visual model itself parse the information for suspicious classes or objects. Specifically, the output visual features for the original image  $\mathbf{v}_0^o$  are leveraged to provide the global description of the target object or regions in the look-back path. With the multi-level features and the global embeddings, the LTFM module is responsible for the recognition and localization of suspicious tokens.
+
+Drawing inspiration from human visual inspection, where suspicious objects or regions are identified and then inspected closely (see Figure 4), we design the VT selector module for aggregating (zooming in) the crucial visual tokens and explicitly assisting the LLM in distinguishing these tokens from many irrelevant ones when dealing with instructions regarding anomaly detection and reasoning. Additionally, the original visual features are preserved to maintain the generalization capability of the base model on regular instructions, such as Can you describe the content of the image?
+
+# 3.3. Look-Twice Feature Matching
+
+Given the global object information  $\mathbf{v}_0^o$  provided by the lookback path, we generate the class-awareness abnormality description by merging  $\mathbf{v}_0^o$  with two learnable embeddings:  $\mathbf{e}^{+}\in \mathbb{R}^{D}$  and  $\mathbf{e}^{-}\in \mathbb{R}^{D}$ , where  $+$  and  $-$  indicate positive (anomalous) and negative (normal) patterns and  $D$  is the embedding dimension. Specifically, a linear layer  $\mathcal{T}_i^o$  is applied along the token dimension to select and fuse useful tokens from  $\mathbf{v}_0^o$ , and then the fused vector will be concatenated with  $\mathbf{e}^{+}$  and  $\mathbf{e}^{-}$  independently and pass through two MLPs  $\{\mathcal{G}_i^+, \mathcal{G}_i^-\}$  to generate the abnormality and normality descriptions  $\{\mathbf{d}_i^+, \mathbf{d}_i^-\}$ , which can be represented by:
+
+$$
+\left\{\mathbf {d} _ {i} ^ {+}, \mathbf {d} _ {i} ^ {-} \right\} = \left\{ \begin{array}{l} \mathcal {G} _ {i} ^ {+} \left(\mathbf {e} ^ {+} \circ \mathcal {T} _ {i} ^ {o} \left(\mathbf {v} _ {0} ^ {o}\right)\right) \\ \mathcal {G} _ {i} ^ {-} \left(\mathbf {e} ^ {-} \circ \mathcal {T} _ {i} ^ {o} \left(\mathbf {v} _ {0} ^ {o}\right)\right) \end{array} \right. \tag {3}
+$$
+
+The visual features extracted from different levels of the ViT focus on different scales of semantics. Thus, the parameters of  $\mathcal{T}_i^o$  and  $\{\mathcal{G}_i^+, \mathcal{G}_i^-\}$  should be independent for different levels, where  $i$  indicate the level number.
+
+Similar to the zero-shot classification mechanism of CLIP
+
+models, we calculate the possibilities of each patch token in  $\mathbf{v}_j^i$  belonging to the anomalous patterns by combining cosine similarity and softmax operations:
+
+$$
+\mathbf {m} _ {j} ^ {i} = \frac {\exp \left(<   \mathbf {d} _ {i} ^ {+} , \mathbf {v} _ {j} ^ {i} > / \tau\right)}{\exp \left(<   \mathbf {d} _ {i} ^ {+} , \mathbf {v} _ {j} ^ {i} > / \tau\right) + \exp \left(<   \mathbf {d} _ {i} ^ {-} , \mathbf {v} _ {j} ^ {i} > / \tau\right)} \tag {4}
+$$
+
+where  $\mathbf{m}_j^i$  represents the significance map for visual tokens,  $\tau$  is the temperature hyperparameter, and  $<,>$  refers to the cosine similarity operator. The patch weight in  $\mathbf{m}_j^i$  indicates the closeness of the corresponding visual token to the anomalous pattern. Then, all the maps are averaged to capture the token significances from low to high levels:
+
+$$
+\mathbf {m} _ {j} = \sum_ {i = 1} ^ {4} \mathbf {m} _ {j} ^ {i} / 4 \tag {5}
+$$
+
+The visual features are leveraged twice in the forward and look-back paths, so this module is named by Look-Twice Feature Matching (LTFM), following the nature of two-step human visual inspection shown in Figure 4.
+
+# 3.4. Visual Token Selector
+
+Under the image cropping strategy widely applied in recent MLLMs, there will be a large number of visual tokens for a high-resolution image, e.g., 7290 tokens for an image with  $1152 \times 1152$  resolution in LLaVA-OneVision. While these tokens provide rich visual details, the LLM is required to pick the most useful information when adapting to a specific task. When the LLM lacks enough knowledge in this domain, the token-picking process will become complicated. Thus, our solution is to introduce a specialist or expert who knows which token is crucial or not and assist the LLM in selecting and emphasizing (zooming in) the crucial tokens.
+
+Given the encoded visual tokens  $\{\mathbf{v}_j^o\}$  for each image crop in  $\mathcal{I}$  and the corresponding significance map  $\mathbf{m}_j$ , the suspicious tokens are emphasized by direct multiplication of the two tensors. Then, the normal tokens will be scaled to zero while the anomalous tokens will be maintained. After that, spatial average pooling  $\mathcal{P}$  is applied to reduce the number of tokens. This process can be written as:
+
+$$
+\mathbf {q} _ {j} = \mathcal {P} \left(\mathbf {v} _ {j} ^ {o} \odot \mathbf {m} _ {j}\right) \tag {6}
+$$
+
+where  $\mathbf{q}_j\in \mathbb{R}^{h\times w\times D}$  refers to the pooled query tokens. Empirically, setting  $h = w = 2$  provides a better trade-off than other options. Then, a Q-Former  $\mathcal{Q}$  [49] is leveraged to aggregate the correlated tokens in the original output by forwarding  $\mathbf{q}_j$  as the query and  $\mathbf{v}_j^o$  as the key and value:
+
+$$
+\mathbf {v} _ {j} ^ {s} = \mathcal {Q} \left(\mathbf {q} _ {j}, \mathbf {v} _ {j} ^ {o}, \mathbf {v} _ {j} ^ {o}\right) \tag {7}
+$$
+
+The Visual Token Selector (VT Selector) serves as a tool for the anomaly expert to hand-pick visual tokens that contain the most suspicious semantics for a given image.
+
+Figure 5. Composition of the instruction data in Anomaly-Instruct-125k. There are four main types of image samples: in-the-wild, industrial, medical, and  $3D$  (in the format of multi-view images), covering most image anomaly detection tasks and enabling the possibility of a unified assistant for visual inspection. The reasoning words are highlighted in blue. For more information about dataset establishment, statistics, and the data collection pipeline, please refer to Section A1 in the supplementary.  
+![](images/27870aa2e157a6e267299eafa2aaa3056f439cbc70d1a89ab877c999c695d5e7.jpg)  
+Q: Do you notice any anomalies in the image? A: Yes, the anomaly in the image is a fine, dark crack visible on the ceramic cup, starting from the rim and extending downward.  
+Q: How can I prevent similar anomalies in the future? A: Avoid rapid temperature changes, handle the cup carefully to minimize impacts, and consider using cups made with thicker, more durable materials.
+
+![](images/1cae4e97724f842ce372d6f3700a1026cebd396a15bf77bfa7c68dac0e191126.jpg)  
+Q: Are there any anomalies in this cable image? A: Yes, the green wire is missing its internal copper conductors, unlike the blue and brown wires, which contain complete sets of bundled copper strands.  
+Q: What could be the reason behind this anomaly? A: The absence of copper in the green wire could be due to manufacturing defects such as failure to insert the strands or a mechanical issue where the strands were removed or misplaced during production or handling.
+
+![](images/236f9ae2d43bd439000f023802369ecec417947d1e99b3582af1ac9171fb9000.jpg)  
+Q: What does the image show? A: The image shows a densely cellular tissue sample with features indicating potential metastatic breast cancer.  
+Q: What suggests that there might be malignant activity in the tissue sample? A: Malignant activity is suggested by irregularly shaped cells, hyperchromatic nuclei, and an increased nuclear-to-cytoplasmic ratio, indicative of abnormal and increased DNA content typical in cancerous cells.
+
+![](images/2e43f15b428805c298907b2430f159ad648f7a757bfd215f12ca344cd3fa05b1.jpg)  
+Q: What kind of object are you examining in the images? A: I'm examining a round cookie with a golden-brown surface and visible chocolate chips embedded across its surface. Q: Are there any significant anomalies detected on this cookie? A: Yes, there are two significant anomalies on the cookie: a central defect featuring a noticeable hole and an edge defect showing a gap or crack.
+
+# 3.5. Inference and Loss
+
+Anomaly Prediction. In the traditional anomaly detection task, the model predicts the possibility of the image being abnormal. To achieve anomaly score prediction, we aggregate the anomaly information from all the image crops by an average operation weighted on the significance maps:
+
+$$
+\mathbf {r} (\mathcal {I}) = \frac {\sum_ {j , k} \mathbf {v} _ {j} ^ {\mathrm {s}} [ k ] \cdot \mathcal {P} (\mathbf {m} _ {j}) [ k ]}{\sum_ {j , k} \mathcal {P} (\mathbf {m} _ {j}) [ k ]} \tag {8}
+$$
+
+where  $\mathcal{P}$  is the same spatial pooling in VT Selector and  $\mathbf{r}(\mathcal{I})$  is a vector containing the global anomaly information for the entire image. Then, the anomaly expert can calculate the image-level abnormal possibility by parsing  $\mathbf{r}(\mathcal{I})$ :
+
+$$
+\operatorname {s c o r e} (\mathcal {I}) = \operatorname {S i g m o i d} \left(\mathcal {G} ^ {o} (\mathbf {r} (\mathcal {I}))\right) \tag {9}
+$$
+
+where  $\mathcal{G}^o$  is an MLP for distinguishing normal and abnormal semantics. To handle the unbalanced sample distribution, we employ the balanced BCE loss as the professional training objective for the anomaly expert components.
+
+Text Generation. Instead of directly forwarding the concatenation of the original  $\{\mathbf{v}_j^o\}$  and the selected  $\{\mathbf{r}(\mathcal{I}),\mathbf{v}_j^s\}$  visual tokens into the LLM, we apply an indication prompt with <adv> suspicious feature: in the middle of the two series of tokens, which will highlight the selected tokens for LLM when handling anomaly-related instructions. This approach can be considered a form of prompt engineering in MLLMs. Besides, the <adv> is chosen from {highly, moderately, slightly} and is determined by score  $(\mathcal{I})$  and predefined thresholds  $\{\mathbf{s}_{low},\mathbf{s}_{high}\}$ . When the input image  $\mathcal{I}$  has a high likelihood of anomaly, the LLM will place greater emphasis on the selected tokens; otherwise, these tokens will have less significance. The text generation is implemented by the original auto-regressive token prediction mechanism of LLM:
+
+$$
+p \left(X _ {a} \mid \mathcal {I}, X _ {q}\right) = \prod_ {t = 1} ^ {L} p _ {\theta} \left(x _ {t} \mid \mathcal {I}, X _ {q, <   t}, X _ {a, <   t}\right) \tag {10}
+$$
+
+where  $X_{a, < t}$  and  $X_{q, < t}$  are the answer and instruction tokens from all prior turns before the current prediction token  $x_{t}$  for a sequence of length  $L$ . The entire model is parameterized by  $\theta$  and trained by the original language model cross-entropy loss for each predicted answer token  $x_{t}$ .
+
+# 4. Dataset and Benchmark
+
+The lack of multimodal instruction-following data for image anomaly detection and reasoning hinders the development of special intelligent assistants in this domain. Even though AnomalyGPT [28] introduces a prompt tuning dataset by simulating the anomalies, the scale of their dataset and the diversity of their instructions and answers are limited, only focusing on anomaly localization. To resolve the data scarcity issue, we establish the first large-scale instruction tuning dataset: Anomaly-Instruct-125k and the corresponding anomaly detection and reasoning benchmark: VisA-D&R.
+
+# 4.1. Anomaly-Instruct-125k
+
+LLaVA [57] builds its instruction tuning dataset by leveraging the image caption and bounding boxes available in the COCO dataset[55] to prompt the text-only GPT-4. ShareGPT4V [8] provides a higher-quality dataset by directly prompting GPT-4V [71]. However, there is no image caption provided in existing anomaly detection datasets [1, 2], and no matter GPT-4V [71] or most recent GPT-4o [72] cannot accurately locate and describe the anomalies in the image without explicit human involvement.
+
+To resolve these issues, we design a new prompt pipeline for accurate anomaly description generation. Since most of the datasets contain annotations for anomaly types, we manually combine the class name and anomaly type, such as a [capsule] with [poke] on surface. If the anomaly masks are provided, we draw bounding boxes on the images to highlight the anomalous area. The short description and the image with (or w/o) bounding boxes are used to prompt GPT-4o to generate the detailed image and anomaly descriptions. Then, we employ an in-context learning strategy similar to LLaVA to create the instructions.
+
+# Detection:
+
+Q: Are there any defects for the object in the image? Please reply with 'Yes' or 'No'.
+
+# Reasoning:
+
+Q1: Do you observe any anomalies in the image?  
+Q2: Can you describe the anomalies you observed?  
+Q3: What is the potential cause for the anomalies?  
+Q4: How can such anomalies be prevented in the future?
+
+Figure 6. Prompt examples in VisA-D&R for detection and reasoning. The complex reasoning instructions are highlighted.
+
+For a unified visual inspection dataset, precise instruction data is collected from MVtec AD [2], the training set of BMAD [1], Anomaly-ShapeNet [50], Real3D-AD [60], and MVtec-3D AD [4], covering both 2D to 3D data across industry to medical domains. The 3D point cloud data are converted into 9 multi-view images, and the corresponding masks are rendered using predefined camera positions. However, the diversities and scales of these datasets are relatively limited, probably due to the difficulty of collecting anomaly images. To scale up the instruction data, we introduce an automatic anomaly data collection pipeline combining GPT4o [72] and Google Image Search [26] for image collection, data cleaning, and instruction generation. Finally, 72k in-the-wild images (named as WebAD) targeting anomaly detection are collected, significantly enriching our instruction dataset. Several samples from Anomaly-Instruct-125k are shown in Figure 5. The instructions are mainly in the format of multi-round conversations, covering anomaly detection and description in low-level reasoning and potential cause and future suggestions for complex understanding.
+
+# 4.2.VisA-D&R
+
+VisA [115] is a classic but challenging industrial anomaly detection dataset, providing fine-grained anomaly type and segmentation for each image. For evaluation of the anomaly detection and reasoning performance on existing and future methods, we select 10 classes from VisA and follow a similar data generation pipeline of Anomaly-Instruct-125k to create the benchmark. Differently, significant human effort has been invested in meticulously reviewing all generated images and anomaly descriptions. Wrong descriptions are picked out and re-annotated by humans before utilizing them for Q&A generation. Totally, the benchmark consists of 761 normal samples and 1000 anomalous ones.
+
+For evaluating detection performance, questions designed to elicit a one-word answer are used to prompt the MLLMs (Figure 6), with results quantified using Accuracy, Precision, Recall, and F1-score. We divide the reasoning performance into two parts: low-level reasoning that focuses on the description of visual defects or anomalies and complex reasoning requiring the MLLMs to provide the potential cause and future improvement strategies for the detected anomalies, where ROUGE-L [54], Sentence-BERT (SBERT) [75], and
+
+GPT-Score (GPT-4 as the judge [57]) are utilized to quantify the similarity between generated text and ground truth. Note that low-level reasoning is highly correlated to detection performance, while anomaly-type descriptions of low-level reasoning determine the output of complex reasoning.
+
+# 5. Experiments
+
+# 5.1. Training & Evaluation
+
+There are two independent training stages for Anomaly-OV. In Stage 1, the components of the anomaly expert are trained to obtain the token selection capability, targeting traditional ZSAD. This stage utilizes all of the data with anomaly labels in Anomaly-Instruct-125k. Similar to previous works [6, 110], when evaluating the model on the datasets contained in the training set, the corresponding datasets are replaced by VisA [115]. In Stage 2, the anomaly expert and visual encoder are frozen, while the projector and LLM are trainable. In addition to our instruction dataset, we sample around  $350\mathrm{k}$  data from the original training recipe of  $LLaVA-OneVision$  to maintain the generalization ability. For more details on training, please refer to the supplementary.
+
+The ZSAD performance for the anomaly expert is evaluated on nine benchmarks, including MVTec AD [2], VisA [115], AITEX [81], ELPV [16], BTAD [68], and MPDD [39] for industrial inspection, and BrainMRI [40], HeadCT [41], and Br35H [30] for medical diagnosis. AUROC (Area Under the Receiver Operating Characteristic) is leveraged to quantify the image-level AD performance. For text-based anomaly detection, both normal and anomaly data are employed to assess the accuracy by examining if the generated text contains the word Yes. Differently, only anomaly data are utilized to prompt the MLLMs to determine their anomaly reasoning capabilities since the justifications of normality are similar for different models.
+
+<table><tr><td>Method</td><td>MVTec</td><td>VisA</td><td>HeadCT</td><td>BrainMRI</td></tr><tr><td>Full Model</td><td>94.0</td><td>91.1</td><td>97.6</td><td>93.9</td></tr><tr><td>w/o. Look-back</td><td>92.8</td><td>90.5</td><td>96.6</td><td>93.5</td></tr><tr><td>w/o. e+ &amp; e-</td><td>92.1</td><td>90.1</td><td>94.7</td><td>92.9</td></tr><tr><td>w/o. Q-Former</td><td>91.7</td><td>89.9</td><td>92.8</td><td>95.1</td></tr><tr><td>w/o. WebAD</td><td>88.5</td><td>88.9</td><td>91.2</td><td>93.4</td></tr></table>
+
+Table 1. Ablation study for the anomaly expert of Anomaly-OV. w/o. Look-back refers to the removal of  $\mathbf{v}_0^o$  in LTFM.
+
+# 5.2. Zero-Shot Anomaly Detection
+
+As shown in Table 2, compared with existing methods, the anomaly expert of Anomaly-OV achieves significant image-level AUROC improvements on most of the ZSAD benchmarks, which demonstrates that the text encoder widely applied in existing models is not necessary. The success of our model mainly originates from the extra data of WebAD (Table 1), which enables the model to learn more generic
+
+<table><tr><td rowspan="2">Model</td><td colspan="6">Industrial Defects</td><td colspan="3">Medical Anomalies</td><td rowspan="2">Average</td></tr><tr><td>MVTec AD</td><td>VisA</td><td>AITEX</td><td>ELPV</td><td>BTAD</td><td>MPDD</td><td>BrainMRI</td><td>HeadCT</td><td>Br35H</td></tr><tr><td>CLIP [73]</td><td>74.1</td><td>66.4</td><td>71.0</td><td>59.2</td><td>34.5</td><td>54.3</td><td>73.9</td><td>56.5</td><td>78.4</td><td>63.1</td></tr><tr><td>CoOp [108]</td><td>88.8</td><td>62.8</td><td>66.2</td><td>73.0</td><td>66.8</td><td>55.1</td><td>61.3</td><td>78.4</td><td>86.0</td><td>70.9</td></tr><tr><td>WinCLIP [38]</td><td>91.8</td><td>78.8</td><td>73.0</td><td>74.0</td><td>68.2</td><td>63.6</td><td>92.6</td><td>90.0</td><td>80.5</td><td>79.2</td></tr><tr><td>APRIL-GAN [11]</td><td>86.2</td><td>78.0</td><td>57.6</td><td>65.5</td><td>73.6</td><td>73.0</td><td>89.3</td><td>89.1</td><td>93.1</td><td>78.4</td></tr><tr><td>AnoVL [19]</td><td>92.5</td><td>79.2</td><td>72.5</td><td>70.6</td><td>80.3</td><td>68.9</td><td>88.7</td><td>81.6</td><td>88.4</td><td>80.3</td></tr><tr><td>AnomalyCLIP [110]</td><td>91.5</td><td>82.1</td><td>62.2</td><td>81.5</td><td>88.3</td><td>77.0</td><td>90.3</td><td>93.4</td><td>94.6</td><td>84.5</td></tr><tr><td>AdaCLIP [6]</td><td>89.2</td><td>85.8</td><td>64.5</td><td>79.7</td><td>88.6</td><td>76.0</td><td>94.8</td><td>91.4</td><td>97.7</td><td>85.3</td></tr><tr><td>Ours</td><td>94.0</td><td>91.1</td><td>72.0</td><td>83.0</td><td>89.0</td><td>81.7</td><td>93.9</td><td>97.6</td><td>95.5</td><td>88.6</td></tr></table>
+
+semantics for normality and abnormality from the data distribution in the absence of the text encoder. This observation also reveals that large-scale in-the-wild online data can benefit zero-shot performance in anomaly detection.
+
+While the Q-Former reduces the model performance on BrainMRI, it shows effectiveness on most benchmarks, indicating the importance of token aggregation. Similarly, the look-back information and two learnable embeddings are required for describing class-awareness abnormality and distinguishing positive and negative features, respectively. As previously discussed, the anomaly expert is responsible for selecting suspicious visual tokens for the LLM, and the significance maps shown in Figure 7 demonstrate the interpretable token selection mechanism. The high intensities are automatically distributed around the anomalous areas even without any supervision of the anomaly masks.
+
+![](images/64f50cfa285d9f428edf3462c2cab1e9bdb560c5289d050d2012e58827b12991.jpg)
+
+![](images/0981baf11f7c8786c1a1a03b1ed604d3cfd7543e706fbc85a5b2ca1336e324fb.jpg)  
+Figure 7. Visualization of the significance map on VisA samples.
+
+![](images/67319f985b80c7763a079f5228d21f9acfc0ba797304cbc6bb0961b1e5991746.jpg)
+
+![](images/fc0d1a80b23d087fe076abffd6415aab48435be08723deaadd71ed472a66b3e1.jpg)
+
+![](images/7c5d82ad91c401f62339a88ed2869ef5c6f00df1516237dd8d3ed6a69835e079.jpg)
+
+![](images/447de3945edc43d51dfcbac8faf9e4d3338c3b8041cb0ddd06aba3772e56e6b4.jpg)
+
+# 5.3. Anomaly Detection & Reasoning
+
+With the strong capabilities of the anomaly expert for zero-shot detection and suspicious token selection, Anomaly-OV accomplishes significant improvement in text-based anomaly detection and reasoning over other open-sourced generalist MLLMs, as shown in Table 4. Here are a few observations: i) While a larger language model cannot guarantee better detection performance, it always provides greater reasoning ability; ii) Most of the existing MLLMs present much lower recall than precision, indicating their insensitivity to visual anomalies; iii) GPT-4o shows stronger reasoning ability
+
+Table 2. Quantitative comparison of Image-level AUROC on different ZSAD methods (some of the results are borrowed from [6, 110, 114]). The best and the second-best results are bolded and underlined, respectively. Please refer to the supplementary for more detailed results.  
+
+<table><tr><td colspan="2">PCB Example: Excessive Solder</td></tr><tr><td>User</td><td>Are there any anomalies visible in the image?</td></tr><tr><td>GPT-4o</td><td>The image shows an HC-SR04 ultrasonic sensor. There don’t appear to be any visible anomalies with the sensor itself. The components and pins seem intact and properly aligned.</td></tr><tr><td>LLaVA-OV</td><td>The image shows an ultrasonic sensor module, specifically the HC-SR04 model. There are no visible anomalies in the image; it appears to be a standard representation of this type of sensor module.</td></tr><tr><td>Ours-7B</td><td>Yes, there is a noticeable anomaly in the image: a short circuit on the PCB, indicated by a white, crusty residue near the ‘Echo’ pin.</td></tr></table>
+
+Table 3. Anomaly-OV presents more accurate anomaly detection.
+
+compared to other open-sourced models. Table 3 and Table 5 provide the qualitative comparison of our Anomaly-OV with its base model LLaVA-OV-7B [44] and GPT-4o [72]. Both GPT-4o and LLaVA-OV show insensitivity to anomalous features and cannot accurately detect the anomaly in the image. Sometimes, GPT-4o knows the image is anomalous but fails to describe the anomalies precisely.
+
+We provide the fine-tuned version of the base model LLaVA-OV-0.5B on Anomaly-Instruct-125k, which presents much higher accuracy and more balanced precision and recall than its original version. This demonstrates the effectiveness of our instruction-tuning dataset. By integrating the anomaly expert with the base model, our Anomaly-OV-0.5B achieves 0.08 accuracy and 0.06 F1-score improvements in text-based anomaly detection and better reasoning capability in low-level and complex settings. Equipped with a larger language model, Anomaly-OV-7B provides the best detection performance among all the existing MLLMs and shows comparable reasoning ability with GPT-4o. Notably, we observe that the anomaly expert restricts the detection perfor
+
+<table><tr><td rowspan="2">Model</td><td colspan="4">Anomaly Detection</td><td colspan="3">Low-level Reasoning</td><td colspan="2">Complex Reasoning</td></tr><tr><td>Accuracy</td><td>Precision</td><td>Recall</td><td>F1-score</td><td>ROUGE-L</td><td>SBERT</td><td>GPT-Score</td><td>SBERT</td><td>GPT-Score</td></tr><tr><td>GPT-4V [71]</td><td>0.68</td><td>0.90</td><td>0.49</td><td>0.55</td><td>0.16</td><td>0.65</td><td>3.31</td><td>0.77</td><td>5.64</td></tr><tr><td>GPT-4o [72]</td><td>0.70</td><td>0.83</td><td>0.71</td><td>0.68</td><td>0.24</td><td>0.71</td><td>4.84</td><td>0.81</td><td>6.89</td></tr><tr><td>Qwen2-VL-2B [87]</td><td>0.65</td><td>0.87</td><td>0.55</td><td>0.59</td><td>0.22</td><td>0.55</td><td>1.94</td><td>0.74</td><td>4.26</td></tr><tr><td>Qwen2-VL-7B [87]</td><td>0.76</td><td>0.91</td><td>0.69</td><td>0.75</td><td>0.25</td><td>0.61</td><td>3.09</td><td>0.68</td><td>4.62</td></tr><tr><td>InternVL-2-8B [13]</td><td>0.74</td><td>0.78</td><td>0.81</td><td>0.76</td><td>0.23</td><td>0.73</td><td>3.69</td><td>0.80</td><td>5.08</td></tr><tr><td>InternVL-2-26B [13]</td><td>0.73</td><td>0.86</td><td>0.66</td><td>0.68</td><td>0.21</td><td>0.74</td><td>4.13</td><td>0.80</td><td>5.49</td></tr><tr><td>IXC-2.5-7B [101]</td><td>0.72</td><td>0.88</td><td>0.63</td><td>0.67</td><td>0.21</td><td>0.58</td><td>2.45</td><td>0.77</td><td>5.14</td></tr><tr><td>LLaVA-OV-0.5B [44]</td><td>0.54</td><td>0.70</td><td>0.19</td><td>0.28</td><td>0.20</td><td>0.63</td><td>2.54</td><td>0.81</td><td>4.34</td></tr><tr><td>LLaVA-OV-7B [44]</td><td>0.71</td><td>0.95</td><td>0.56</td><td>0.63</td><td>0.24</td><td>0.66</td><td>3.57</td><td>0.79</td><td>5.44</td></tr><tr><td>LLaVA-OV-0.5B*</td><td>0.71</td><td>0.77</td><td>0.84</td><td>0.76</td><td>0.31</td><td>0.70</td><td>3.69</td><td>0.82</td><td>5.31</td></tr><tr><td>Anomaly-OV-0.5B</td><td>0.79</td><td>0.86</td><td>0.83</td><td>0.82</td><td>0.33</td><td>0.72</td><td>3.87</td><td>0.83</td><td>5.67</td></tr><tr><td>Anomaly-OV-7B</td><td>0.79</td><td>0.83</td><td>0.86</td><td>0.83</td><td>0.34</td><td>0.73</td><td>4.26</td><td>0.84</td><td>6.34</td></tr></table>
+
+Table 4. Quantitative comparison of text-based anomaly detection and reasoning for MLLMs. Notably, the Accuracy and F1-score for the anomaly expert of Anomaly-OV can be calculated as  $\{0.78, 0.77\}$  with threshold 0.5. * indicates the model is fine-tuned on our dataset.
+
+mance of Anomaly-OV. Therefore, the design of a stronger anomaly expert is suggested for future works.
+
+Macaroni Example: Yellowish Spot
+
+![](images/88fd3c7885aa08856f47465b158091edfd8ec145c9f1f282665e988598f21199.jpg)
+
+User  
+Do you see any anomalies in the image?  
+
+<table><tr><td>GPT-4o</td><td>The image shows four pieces of elbow macaroni on a green background. The anomaly is that the macaroni pieces are not whole; they are cut in half.</td></tr><tr><td>LLaVA-OV</td><td>The image shows four pieces of pasta, specifically macaroni shells, arranged on a green textured surface. The pasta appears to be uniformly colored and shaped, with no visible defects or anomalies.</td></tr><tr><td>Ours-7B</td><td>Yes, there is an anomaly in the image. The bottom right pasta piece has a noticeable yellowish discoloration on its surface.</td></tr></table>
+
+Table 5. Anomaly-OV presents more precise anomaly reasoning.
+
+# 5.4. Extension
+
+With the generalization and multi-image processing capabilities of MLLMs, it is possible to build a unified assistant for visual inspection. Table 6 demonstrates the comprehensive knowledge of Anomaly-OV (without using Anomaly-ShapeNet [50] for training) on 3D and medical (testing set of BMAD [1]) AD & reasoning. More data, benchmarks, and investigation on a unified model are meaningful.
+
+# 6. Conclusion
+
+In this paper, we establish the first large-scale visual instruction tuning dataset, Anomaly-Instruct-125k, and the corresponding benchmark, VisA-D&R, to address the data scarcity issue for visual anomaly detection and reasoning.
+
+3D Example: Bulge
+
+![](images/04a92329325abad74a691dc739a11fb8b70be3374afe8a3918381a73b2a06cd4.jpg)
+
+![](images/56fbef7823f2c266934b4fbe1b91fbc04875b8ebd7da10dcf49856622d26e33d.jpg)  
+···
+
+User
+
+Is there any anomaly present in the bowl?
+
+Ours-7B
+
+Yes, there is a bulge on the side of the bowl, which disrupts its otherwise smooth and symmetrical surface.
+
+Medical Example: Pneumonia
+
+![](images/02534aa62c713e51f9815a7cd628b8ba19ef8c12742b7262babaea7aca80ebc9.jpg)
+
+User
+
+What is the most likely diagnosis for this patient?
+
+Ours-7B
+
+Based on the chest X-ray, the most likely diagnosis for this patient is pneumonia. The presence of bilateral patchy opacities in the lower lung fields is a common finding in pneumonia, which is an infection that causes inflammation in the air sacs of the lungs.
+
+User
+
+What should the patient do next?
+
+Ours-7B
+
+The patient should consult a healthcare professional for further evaluation and appropriate treatment. This may include antibiotics to treat the infection, supportive care such as oxygen therapy, and monitoring of the patient's condition.
+
+Table 6. Extension to 3D and medical AD & reasoning.
+
+Then, a specialist MLLM, Anomaly-OV, targeting visual inspection is introduced to serve as the baseline in this domain. Anomaly-OV leverages an anomaly expert to assist the LLM with suspicious visual token selection and presents significant improvements on both traditional ZSAD and text-based anomaly detection and reasoning tasks over existing methods. Extension to 3D and medical domains is demonstrated.
+
+# References
+
+[1] Jinan Bao, Hanshi Sun, Hanqiu Deng, Yinsheng He, Zhaoxiang Zhang, and Xingyu Li. Bmad: Benchmarks for medical anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4042-4053, 2024. 1, 5, 6, 8, 14  
+[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 5, 6, 14, 15  
+[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4183–4192, 2020. 1  
+[4] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. SCITEPRESS - Science and Technology Publications, 2022. 6  
+[5] Tri Cao, Jiawen Zhu, and Guansong Pang. Anomaly detection under distribution shift. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6511-6523, 2023. 1  
+[6] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, pages 55-72. Springer, 2025. 1, 2, 3, 6, 7  
+[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 1  
+[8] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions, 2023. 5, 14  
+[9] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. arXiv preprint arXiv:2407.09359, 2024. 1  
+[10] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023. 2  
+[11] Xuhai Chen, Yue Han, and Jiangning Zhang. April-gan: A zero-/few-shot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad, 2023. 2, 7
+
+[12] Yuanhong Chen, Yu Tian, Guansong Pang, and Gustavo Carneiro. Deep one-class classification via interpolated gaussian descriptor. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 383-392, 2022. 1  
+[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, Ji Ma, Jiaqi Wang, Xiaoyi Dong, Hang Yan, Hewei Guo, Conghui He, Botian Shi, Zhenjiang Jin, Chao Xu, Bin Wang, Xingjian Wei, Wei Li, Wenjian Zhang, Bo Zhang, Pinlong Cai, Licheng Wen, Xiangchao Yan, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang. How far are we to gpt-4v? closing the gap to commercial multimodal models with opensource suites, 2024. 8  
+[14] Zebang Cheng, Zhi-Qi Cheng, Jun-Yan He, Jingdong Sun, Kai Wang, Yuxiang Lin, Zheng Lian, Xiaojiang Peng, and Alexander Hauptmann. Emotion-llama: Multimodal emotion recognition and reasoning with instruction tuning. In Conference on Neural Information Processing Systems, 2024. 1, 2  
+[15] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2  
+[16] Sergiu Deitsch, Vincent Christlein, Stephan Berger, Claudia Buerhop-Lutz, Andreas Maier, Florian Gallwitz, and Christian Riess. Automatic classification of defective photovoltaic module cells in electroluminescence images. Solar Energy, 185:455-468, 2019. 6  
+[17] Chenghao Deng, Haote Xu, Xiaolu Chen, Haodi Xu, Xiaotong Tu, Xinghao Ding, and Yue Huang. Simclip: Refining image-text alignment with simple prompts for zero-/few-shot anomaly detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 1761-1770, 2024. 1  
+[18] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In IEEE/CVF conference on computer vision and pattern recognition, 2022. 2  
+[19] Hanqiu Deng, Zhaoxiang Zhang, Jinan Bao, and Xingyu Li. Anovl: Adapting vision-language models for unified zero-shot anomaly localization. arXiv preprint arXiv:2308.15939, 2023. 1, 2, 7  
+[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 2  
+[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3  
+[22] Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fastrecon: Few-shot industrial
+
+anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481-17490, 2023. 2  
+[23] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Chatpose: Chatting about 3d human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2093-2103, 2024. 3  
+[24] Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection-a survey. ACM Computing Surveys (CSUR), 54(7):1-37, 2021. 1  
+[25] Matic Fučka, Vitjan Zavrtanik, and Danijel Skočaj. Transfusion-a transparency-based diffusion model for anomaly detection. In European conference on computer vision, pages 91-108. Springer, 2025. 1  
+[26] Google. Google-images-search 1.4.7, 2024. https://pypi.org/project/Google-Images-Search.6, 14  
+[27] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Hao Li, Ming Tang, and Jinqiao Wang. Filo: Zero-shot anomaly detection by fine-grained description and high-quality localization. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2041-2049, 2024. 1  
+[28] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1932-1940, 2024. 1, 2, 3, 5  
+[29] Yuxiang Guo, Faizan Siddiqui, Yang Zhao, Rama Chellappa, and Shao-Yuan Lo. Stimuvar: Spatiotemporal stimuli-aware video affective reasoning with multimodal large language models. arXiv preprint arXiv:2409.00304, 2024. 1, 2  
+[30] Ahmed Hamada. Br35h: Brain tumor detection 2020, 2020. 6  
+[31] Changhee Han, Leonardo Rundo, Kohei Murao, Tomoyuki Noguchi, Yuki Shimahara, Zoltan Adam Milacski, Saori Koshino, Evis Sala, Hideki Nakayama, and Shin'ichi Satoh. Madgan: Unsupervised medical anomaly detection gan using multiple adjacent brain mri slice reconstruction. BMC bioinformatics, 22:1-20, 2021. 1  
+[32] Liren He, Zhengkai Jiang, Jinlong Peng, Liang Liu, Qianggang Du, Xiaobin Hu, Wenbing Zhu, Mingmin Chi, Yabiao Wang, and Chengjie Wang. Learning unified reference representation for unsupervised multi-class anomaly detection. arXiv preprint arXiv:2403.11561, 2024. 1  
+[33] Chih-Hui Ho, Kuan-Chuan Peng, and Nuno Vasconcelos. Long-tailed anomaly detection with learnable class names. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12435-12446, 2024.  
+[34] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8791-8800, 2021. 1  
+[35] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based
+
+few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1, 2  
+[36] Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xin chao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in medical images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11375-11385, 2024. 1, 2  
+[37] Brian KS Isaac-Medina, Yona Falinie A Gaus, Neelanjan Bhowmik, and Toby P Breckon. Towards open-world object-based anomaly detection via self-supervised outlier synthesis. In European Conference on Computer Vision (ECCV), 2024. 1  
+[38] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 1, 2, 7  
+[39] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International congress on ultra modern telecommunications and control systems and workshops (ICUMT), pages 66-71. IEEE, 2021. 6  
+[40] Pranita Balaji Kanade and PP Gumaste. Brain tumor detection using mri images. *Brain*, 3(2):146-150, 2015. 6  
+[41] Felipe Campos Kitamura. Head ct - hemorrhage, 2018. 6  
+[42] Mingyu Lee and Jongwon Choi. Text-guided variational image generation for industrial anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26519-26528, 2024. 1  
+[43] Aodong Li, Chen Qiu, Marius Kloft, Padhraic Smyth, Maja Rudolph, and Stephan Mandt. Zero-shot anomaly detection via batch normalization. Advances in Neural Information Processing Systems, 36, 2024. 1  
+[44] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 1, 3, 7, 8, 15  
+[45] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. LLaVA-med: Training a large language-and-vision assistant for biomedicine in one day. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 1, 15  
+[46] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In IEEE/CVF conference on computer vision and pattern recognition, 2021. 2  
+[47] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1, 3  
+[48] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Interna-
+
+tional conference on machine learning, pages 12888-12900.  
+PMLR, 2022. 1, 2  
+[49] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2, 4  
+[50] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22207-22216, 2024. 6, 8  
+[51] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompt: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16838-16848, 2024. 1, 2  
+[52] Yiting Li, Adam Goodge, Fayao Liu, and Chuan-Sheng Foo. Promptad: Zero-shot anomaly detection using text prompts. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1093-1102, 2024. 1  
+[53] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2  
+[54] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6  
+[55] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, 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  
+[56] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 2  
+[57] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 3, 5, 6, 16  
+[58] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1  
+[59] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024.3  
+[60] Jiaqi Liu, Guoyang Xie, ruitao chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. Real3d
+
+AD: A dataset of point cloud anomaly detection. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6  
+[61] Jie Liu, Yixiao Zhang, Jie-Neng Chen, Junfei Xiao, Yongyi Lu, Bennett A Landman, Yixuan Yuan, Alan Yuille, Yucheng Tang, and Zongwei Zhou. Clip-driven universal model for organ segmentation and tumor detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21152-21164, 2023. 2  
+[62] 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, 2019. 2  
+[63] Shao-Yuan Lo, Poojan Oza, and Vishal M Patel. Adversarially robust one-class novelty detection. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2  
+[64] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations, 2017. 15  
+[65] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 15  
+[66] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 2  
+[67] Hui Lv and Qianru Sun. Video anomaly detection and explanation via large language models. arXiv preprint arXiv:2401.05702, 2024. 1, 2  
+[68] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adt: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE), pages 01-06. IEEE, 2021. 6  
+[69] Shancong Mou, Xiaoyi Gu, Meng Cao, Haoping Bai, Ping Huang, Jiulong Shan, and Jianjun Shi. RGI: robust GAN-inversion for mask-free image inpainting and unsupervised pixel-wise anomaly detection. In The Eleventh International Conference on Learning Representations, 2023. 1, 2  
+[70] Ming Nie, Renyuan Peng, Chunwei Wang, Xinyue Cai, Jianhua Han, Hang Xu, and Li Zhang. Reason2drive: Towards interpretable and chain-based reasoning for autonomous driving. In European Conference on Computer Vision, pages 292-308. Springer, 2025. 1  
+[71] OpenAI. Gpt-4v(isdiction) system card, 2023. https://openai.com/index/gpt-4v-system-card.5,8  
+[72] OpenAI. Gpt-4o system card, 2024. https://openai.com/index/gpt-4o-system-card. 2, 5, 6, 7, 8, 14, 15  
+[73] 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, 7, 14
+
+[74] 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, 2023. 2  
+[75] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks, 2019. 6  
+[76] Tal Reiss and Yedid Hoshen. Mean-shifted contrastive loss for anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2155-2162, 2023. 1  
+[77] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2022. 1, 2  
+[78] Fumiaki Sato, Ryo Hachiuma, and Taiki Sekii. Prompt-guided zero-shot anomaly action recognition using pretrained deep skeleton features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6471-6480, 2023. 1  
+[79] Eli Schwartz, Assaf Arbelle, Leonid Karlinsky, Sivan Harary, Florian Scheidegger, Sivan Doveh, and Raja Giryes. Mae-day: Mae for few-and zero-shot anomaly-detection. Computer Vision and Image Understanding, 241:103958, 2024. 1  
+[80] Pierre Sermanet, Tianli Ding, Jeffrey Zhao, Fei Xia, Debidatta Dwibedi, Keerthana Gopalakrishnan, Christine Chan, Gabriel Dulac-Arnold, Sharath Maddineni, Nikhil J Joshi, et al. Robovqa: Multimodal long-horizon reasoning for robotics. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 645-652. IEEE, 2024. 1  
+[81] Javier Silvestre-Blanes, Teresa Albero-Albero, Ignacio Miralles, Rubén Pérez-Llorens, and Jorge Moreno. A public fabric database for defect detection methods and results. Autex Research Journal, 19(4):363–374, 2019. 6  
+[82] Luc PJ Sträter, Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Generalad: Anomaly detection across domains by attending to distorted features. arXiv preprint arXiv:2407.12427, 2024. 1  
+[83] Haomiao Sun, Mingjie He, Tianheng Lian, Hu Han, and Shiguang Shan. Face-mllm: A large face perception model, 2024. 3  
+[84] Jiaqi Tang, Hao Lu, Xiaogang Xu, Ruizheng Wu, Sixing Hu, Tong Zhang, Tsz Wa Cheng, Ming Ge, Ying-Cong Chen, and Fugee Tsung. An incremental unified framework for small defect inspection. In European Conference on Computer Vision, pages 307-324. Springer, 2025. 1  
+[85] Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan Duong, Chanh D Tr Nguyen, and Steven QH Truong. Revisiting reverse distillation for anomaly detection. In IEEE/CVF conference on computer vision and pattern recognition, 2023. 2  
+[86] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1802-1812, 2023. 2
+
+[87] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution, 2024. 8  
+[88] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In Conference on Neural Information Processing Systems, 2022. 2  
+[89] Qi Wei, Yinhao Ren, Rui Hou, Bibo Shi, Joseph Y Lo, and Lawrence Carin. Anomaly detection for medical images based on a one-class classification. In Medical Imaging 2018: Computer-Aided Diagnosis, pages 375-380. SPIE, 2018. 1  
+[90] Julia Wolleb, Florentin Bieder, Robin Sandkühler, and Philippe C Cattin. Diffusion models for medical anomaly detection. In International Conference on Medical image computing and computer-assisted intervention, pages 35-45. Springer, 2022. 1  
+[91] Binzhu Xie, Sicheng Zhang, Zitang Zhou, Bo Li, Yuanhan Zhang, Jack Hessel, Jingkang Yang, and Ziwei Liu. Funqa: Towards surprising video comprehension. In European Conference on Computer Vision, pages 39-57. Springer, 2025. 1  
+[92] Guoyang Xie, Jinbao Wang, Jiaqi Liu, Yaochu Jin, and Feng Zheng. Pushing the limits of fewshot anomaly detection in industry vision: Graphcore. In The Eleventh International Conference on Learning Representations, 2023. 1  
+[93] Hongxia Xie, Chu-Jun Peng, Yu-Wen Tseng, Hung-Jen Chen, Chan-Feng Hsu, Hong-Han Shuai, and Wen-Huang Cheng. Emovit: Revolutionizing emotion insights with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26596-26605, 2024. 3  
+[94] Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: reasoning for video anomaly detection with large language models. arXiv preprint arXiv:2407.10299, 2024. 1, 2  
+[95] Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. Glad: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection. arXiv preprint arXiv:2406.07487, 2024. 1  
+[96] Huanjin Yao, Wenhao Wu, Taojiannan Yang, YuXin Song, Mengxi Zhang, Haocheng Feng, Yifan Sun, Zhiheng Li, Wanli Ouyang, and Jingdong Wang. Dense connector for mllms, 2024. 17  
+[97] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. arXiv preprint arXiv:2403.13349, 2024. 1  
+[98] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Advances in Neural Information Processing Systems, 2022. 1
+
+[99] Vitjan Zavrtanik, Matej Kristan, and Danijel Skoçaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In IEEE/CVF international conference on computer vision, 2021. 2  
+[100] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3  
+[101] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output, 2024. 8  
+[102] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising studentteacher for anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2  
+[103] Ximiao Zhang, Min Xu, Dehui Qiu, Ruixin Yan, Ning Lang, and Xiuzhuang Zhou. Mediclip: Adapting clip for few-shot medical image anomaly detection. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 458-468. Springer, 2024. 1  
+[104] Ximiao Zhang, Min Xu, and Xiuzhuang Zhou. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16699-16708, 2024. 1  
+[105] Zhuosheng Zhang, Aston Zhang, Mu Li, hai zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. Transactions on Machine Learning Research, 2024. 2  
+[106] He Zhao, Yuexiang Li, Nanjun He, Kai Ma, Leyuan Fang, Huiqi Li, and Yefeng Zheng. Anomaly detection for medical images using self-supervised and translation-consistent features. IEEE Transactions on Medical Imaging, 40(12): 3641-3651, 2021. 1  
+[107] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 1  
+[108] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 7  
+[109] Kaiwen Zhou, Kwonjoon Lee, Teruhisa Mitsu, and Xin Eric Wang. Vicor: Bridging visual understanding and commonsense reasoning with large language models. In Findings of the Association for Computational Linguistics, 202'. 1, 2  
+[110] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 6, 7  
+[111] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language
+
+understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2  
+[112] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17826-17836, 2024. 1, 2  
+[113] Jiaqi Zhu, Shaofeng Cai, Fang Deng, Beng Chin Ooi, and Junran Wu. Do llms understand visual anomalies? uncovering llm's capabilities in zero-shot anomaly detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 48-57, 2024. 1  
+[114] Jiawen Zhu, Yew-Soon Ong, Chunhua Shen, and Guansong Pang. Fine-grained abnormality prompt learning for zero-shot anomaly detection, 2024. 7  
+[115] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 6, 15

2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00ac7d3028648fc096f40f073b0c3cac593afe12897bbcc63e354807ae83ebf2
+size 623487

2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/49360ca2-47e3-4c7b-81b1-6a87621996c4_content_list.json

@@ -0,0 +1,1635 @@
+[
+    {
+        "type": "text",
+        "text": "Towards a Universal Synthetic Video Detector: From Face or Background Manipulations to Fully AI-Generated Content",
+        "text_level": 1,
+        "bbox": [
+            125,
+            128,
+            870,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Rohit Kundu $^{1,2}$ , Hao Xiong $^{1}$ , Vishal Mohanty $^{1}$ , Athula Balachandran $^{1}$ , Amit K. Roy-Chowdhury $^{2}$ ,  $^{1}$ Google LLC,  $^{2}$ University of California, Riverside",
+        "bbox": [
+            112,
+            202,
+            883,
+            239
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{rohit.kundu@email, amertc@ece}.ucr.edu; {rohitkun, haoxg, vishal, athula}@google.com",
+        "bbox": [
+            120,
+            241,
+            870,
+            258
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            291,
+            326,
+            306
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Existing DeepFake detection techniques primarily focus on facial manipulations, such as face-swapping or lip-syncing. However, advancements in text-to-video (T2V) and image-to-video (I2V) generative models now allow fully AI-generated synthetic content and seamless background alterations, challenging face-centric detection methods and demanding more versatile approaches.",
+        "bbox": [
+            89,
+            324,
+            483,
+            430
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address this, we introduce the Universal Network for Identifying Tampered and synthEtic videos (UNITE) model, which, unlike traditional detectors, captures full-frame manipulations. UNITE extends detection capabilities to scenarios without faces, non-human subjects, and complex background modifications. It leverages a transformer-based architecture that processes domain-agnostic features extracted from videos via the SigLIP-So400M foundation model. Given limited datasets encompassing both facial/background alterations and T2V/I2V content, we integrate task-irrelevant data alongside standard DeepFake datasets in training. We further mitigate the model's tendency to over-focus on faces by incorporating an attention-diversity (AD) loss, which promotes diverse spatial attention across video frames. Combining AD loss with cross-entropy improves detection performance across varied contexts. Comparative evaluations demonstrate that UNITE outperforms state-of-the-art detectors on datasets featuring face/background manipulations and fully synthetic T2V/I2V videos, showcasing its adaptability and generalizable detection capabilities.",
+        "bbox": [
+            89,
+            431,
+            483,
+            750
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            91,
+            782,
+            220,
+            797
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The rise of synthetic media, especially DeepFakes, has transformed content perception and interaction online. Hyper-realistic images produced by technologies like  $\\mathrm{FLUX}_{1.1}$  [1] are challenging for even humans to identify as fake, and with the development of their video model underway, this underscores the critical need for effective detec",
+        "bbox": [
+            89,
+            809,
+            483,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/7f4d537dbea8ce1cfa0f687526832537988d457109714d1137feb379a094bc2b.jpg",
+        "image_caption": [
+            "Figure 1. Problem Overview: Existing DeepFake detection methods focus on identifying face-manipulated videos, most of which cannot perform inference unless there is a face detected in the video. However, with advancements like seamless background modifications (e.g., AVID [61]) and hyper-realistic content from games like GTA-V [24] and T2V/I2V models [9], a more comprehensive approach is needed. A model trained with only cross-entropy (CE) loss, using full frames, automatically focuses on the face, capturing temporal discontinuities through its transformer architecture, performing better than random  $(\\approx)$  on T2V/I2V content but struggling with background manipulations. UNITE, with its attention-diversity (AD) loss, effectively detects both face/background manipulations and fully synthetic content."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            535,
+            290,
+            885,
+            503
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tion methods for media generated by deep neural networks. Traditional DeepFake generators [28, 49] focus on manipulating human faces through face-swapping and lip-syncing. However, powerful text-to-video (T2V) and image-to-video (I2V) models [42, 52, 63] have expanded manipulation possibilities beyond faces.",
+        "bbox": [
+            511,
+            700,
+            905,
+            792
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "While conventional detectors [10, 45, 57] perform well on older, face-centric DeepFake datasets [17, 30, 43], they often struggle with newer manipulations involving full scenes or backgrounds. DeepFake-O-Meter [26] is an open-source tool (consisting of several state-of-the-art models) for DeepFake detection, but it cannot run inference unless a human face is visible in the image/video. The rapid spread",
+        "bbox": [
+            511,
+            794,
+            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": "28050",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "of misinformation, particularly during critical periods such as elections, highlights the need for generalizable detection models capable of identifying diverse manipulations, including face, background, and fully AI-generated T2V/I2V content with/without human subjects (overview in Fig. 1).",
+        "bbox": [
+            89,
+            90,
+            480,
+            167
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address these challenges, we present a Universal Network for Identifying Tampered and synthEtic videos or UNITE model, which detects both partially manipulated (foreground/background) and fully synthetic videos. Unlike detectors focused solely on face detection, UNITE analyzes entire frames, regardless of whether a human subject is present in the videos.",
+        "bbox": [
+            89,
+            167,
+            480,
+            272
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Given the inherent domain gaps in DeepFake datasets [44, 64], even when generated by similar techniques, we leverage the SigLIP-So400m [4] foundation model to extract domain-agnostic features. This serves as inputs to a learnable transformer with multi-head attention, enabling effective detection by capturing temporal inconsistencies in synthetic content. However, preliminary experiments show that training a transformer architecture solely with cross-entropy (CE) loss often leads to the attention heads converging on the face region (Fig. 3). As a result, the model struggles during inference when handling videos where a real human subject is placed in a manipulated background or content generated by T2V [8, 52] and I2V [35, 63] models. To address this, we introduce an \"attention-diversity\" (AD) loss that encourages attention heads to focus on different spatial regions, enhancing the model's ability to capture critical cues from both foreground and background.",
+        "bbox": [
+            91,
+            275,
+            482,
+            531
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Due to the limited availability of open-source datasets covering face or background manipulations and fully AI-generated content, we employ innovative training strategies for UNITE. This includes integrating task-irrelevant data with standard DeepFake datasets to simulate AI-generated synthetic media. Beyond the popularly used FaceForensics++ [43] dataset, we utilize the SAIL-VOS-3D [24] dataset, originally designed for 3D video object segmentation in the game GTA-V. As this dataset is fully synthetic, it helps simulate AI-generated media, enhancing our model's ability to detect diverse forms of synthetic manipulation.",
+        "bbox": [
+            89,
+            532,
+            482,
+            698
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions can be summarized as follows:",
+        "bbox": [
+            109,
+            700,
+            434,
+            714
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose UNITE, a model for detecting partially manipulated (foreground/background) and fully synthetic videos, moving beyond face-centric DeepFake detection.",
+            "- Unlike detectors relying on face detection and cropping, our model can process full video frames and can detect fakes even if there are no human subjects.",
+            "- Using the SigLIP-So400m foundation model [4], we extract domain-agnostic features, enabling generalization to in-the-wild DeepFakes.",
+            "- We introduce an attention-diversity loss, encouraging attention heads to focus on diverse spatial regions, enhancing detection beyond face manipulations."
+        ],
+        "bbox": [
+            91,
+            719,
+            482,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- Unlike existing methods that evaluate on a few datasets, we comprehensively assess UNITE on a broad range of face and synthetic datasets with various T2V/I2V generators, outperforming detection in foreground, background, and T2V/I2V manipulations.",
+        "bbox": [
+            513,
+            90,
+            903,
+            167
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            513,
+            180,
+            653,
+            195
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Face-centric DeepFake Detection: Methods such as [13, 19, 40] focus on modeling spatial inconsistencies for DeepFake detection. Concas et al. [13] track specific facial features (eyes, nose, mouth) in a high-frequency domain, as fake videos exhibit distinct behaviors in this space compared to real ones. However, due to their dataset-centric architectures, it is not a generalizable approach. PUDD [40] learns person-specific prototypes by modeling patterns from real videos and assessing how closely inference videos match these prototypes, making it less effective for detecting in-the-wild DeepFakes. LAA-Net [39] attempts to improve generalizability through a multi-task attention module that detects artifacts via heatmap and self-consistency regression, but it works by analyzing every still frame of a video, thus limiting its scalability. The authors in [36] tackled the challenging problem of detecting and localizing expression swaps, where the number of manipulated pixels is even fewer compared to identity swaps. Although the authors obtained good localization performance, their reliance on a CNN-based architecture limits their ability to capture temporal inconsistencies effectively.",
+        "bbox": [
+            511,
+            205,
+            903,
+            522
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "DPNet [50] and ID-Reveal [16] are identity-aware temporal artifact modeling methods: DPNet [50] focuses on interpretable, prototype-based detection of unnatural movements through dynamic feature representations, while IDReveal [16] leverages metric learning on real data to detect biometric inconsistencies. However these methods assume access to authentic reference videos, making them unfit for in-the-wild DeepFake detection. Shifting from identity-specific methods, TALL [57] constructs composite \"thumbnail\" images from four consecutive frames taken at random timestamps for temporal analysis. TI2Net [32] captures temporal anomalies by subtracting consecutive frame features but relies on the assumption that FaceSwap manipulations show pronounced inter-frame discrepancies.",
+        "bbox": [
+            511,
+            523,
+            903,
+            734
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Choi et al. [11] approach DeepFake detection by targeting suppressed variance in style-based latent temporal features via a StyleGRU module. However, this method is limited to cropped face regions, making it ineffective for T2V/I2V or background-manipulated videos.",
+        "bbox": [
+            511,
+            734,
+            903,
+            809
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Synthetic Video Detection: While there has been considerable effort in synthetic image detection [14, 15, 55], synthetic video detection is a scarcely explored area. DeMamba [9], a recently proposed synthetic video detector, analyzes small zones of video frames to capture how pixels change spatiotemporally. Its continuous scanning approach",
+        "bbox": [
+            511,
+            810,
+            903,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "28051",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/293854bf54d05f4417715994a0c7ca60da622532c26dfd731035498dd8ac84a2.jpg",
+        "image_caption": [
+            "Figure 2. UNITE architecture overview: We extract domain-agnostic features  $(\\xi)$  using the SigLIP-So400m foundation model [4] to mitigate domain gaps between DeepFake datasets (Sec. 3.2). These embeddings, combined with positional encodings, are input to a transformer with multi-head attention and MLP layers (Sec. 3.3), culminating in a classifier for final predictions. AD-loss (Sec. 3.4) encourages the model attention to span diverse spatial regions."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            133,
+            88,
+            441,
+            335
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "better tracks subtle changes and patterns, making it easier to spot manipulated or fake content. The authors also propose a million scale T2V/I2V dataset (which we call the DeMamba dataset hereforth) to evaluate their performance. However, the method was not evaluated on typical DeepFake datasets [30, 43, 53] and thus is not fit for human-centric DeepFake detection. In contrast our UNITE model is built to detect partially manipulated as well as fully AI-generated videos. A key reason for the scarcity of synthetic video detection methods in the literature is the lack of standardized datasets for training and evaluation.",
+        "bbox": [
+            89,
+            449,
+            483,
+            614
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Proposed Method",
+        "text_level": 1,
+        "bbox": [
+            89,
+            628,
+            264,
+            646
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section we first describe our problem setup (Sec. 3.1), then we describe the foundation model-based video encoding (Sec. 3.2), our trainable transformer architecture (Sec. 3.3) and finally the novel loss function for training the UNITE model (Sec. 3.4). The overview of the UNITE architecture is shown in Fig. 2.",
+        "bbox": [
+            89,
+            655,
+            483,
+            747
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Problem Setup",
+        "text_level": 1,
+        "bbox": [
+            89,
+            756,
+            241,
+            772
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For each video in the dataset, we implement a frame sampling strategy where every other frame is extracted. From the resulting collection of frames, we construct video segment samples consisting of  $n_f = 64$  consecutive frames, which are defined as a single data sample, denoted as \"v\". This selection aligns with the context window established for our video transformer model, ensuring that each sample retains the temporal coherence necessary for effective train-",
+        "bbox": [
+            89,
+            780,
+            483,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ing. In instances where the video duration results in fewer than  $n_f$  frames, we apply padding (ablation in supplementary) to ensure that each sample remains uniform in size.",
+        "bbox": [
+            511,
+            90,
+            903,
+            136
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given that the original video-level classification labels are known and fake videos contain manipulations throughout their entirety (i.e., no videos have manipulations limited to specific frames), each  $n_f$ -frame segment generated from a given video is assigned the same label (denoted by  $l \\in 0, \\dots, n_c$ ) as the source video, with number of possible classes as  $n_c$ . The aggregation of these samples across all videos forms our final dataset,  $v_i, l_{i=1}^N \\in \\mathcal{V}$ , where  $N$  is the total number of samples obtained. This approach serves as a data augmentation technique, increasing the number of training samples without additional data collection.",
+        "bbox": [
+            511,
+            137,
+            906,
+            303
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Domain-Agnostic Feature Extraction",
+        "text_level": 1,
+        "bbox": [
+            511,
+            313,
+            833,
+            329
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our objective is to leverage the trained UNITE model for detecting in-the-wild DeepFakes, necessitating careful consideration of the domain gap [7, 34] between the train and test datasets. To effectively mitigate the impact of domain discrepancies, we utilize the SigLIP foundation model [59] to extract domain-agnostic features. Specifically, we employ the \"shape-optimized\" ViT [4] image encoder, whose architecture is rooted in [18]. The SigLIP-So400m [4] model was contrastively pretrained on 3B diverse examples using a sigmoid loss, enabling it to capture robust, generalizable features while maintaining a compact size (400M parameters). This extensive pretraining makes SigLIP particularly adept at extracting domain-invariant representations, crucial for handling the significant variability encountered in DeepFake datasets and real-world scenarios.",
+        "bbox": [
+            511,
+            337,
+            906,
+            561
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For every frame  $f_{j}$  (resized to  $\\mathbb{R}^{384\\times 384\\times 3}$ ), of a video sample  $v_{i}\\in \\mathcal{V}$ , we obtain an image encoding from the frozen SigLIP-So400m [4] model as  $e_j = SigLIP(f_j)\\in \\mathbb{R}^{t_s\\times d_s}$ , where  $j\\in \\{1,2,\\dots ,n_f\\}$ , since there are  $n_f$  frames per video sample (as explained in Sec. 3.1),  $t_s = 729$  represents the token length and  $d_{s} = 1152$  represents the feature dimension. We concatenate these frame-level features for the video sample  $v_{i}$ , while maintaining the order of the frames, to obtain a video segment encoding denoted by  $\\xi_{i}\\in \\mathbb{R}^{n_{f}\\times t_{s}\\times d_{s}}$ , which are used as input to our trainable transformer architecture. This encoded video dataset can be denoted as  $\\{\\xi_i,l_i\\}_{i = 1}^N\\in \\mathcal{V}_\\xi$ .",
+        "bbox": [
+            511,
+            564,
+            908,
+            748
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3. Learnable Transformer Architecture",
+        "text_level": 1,
+        "bbox": [
+            511,
+            756,
+            833,
+            771
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The UNITE architecture is a multi-head self-attention (MHSA) based transformer model designed specifically for video classification tasks. It leverages the strengths of Transformer networks to process sequences of video frame embeddings effectively, allowing for robust classification performance. The UNITE architecture allows for adjustable depth, signifying the number of encoder blocks stacked within the model. A deeper architecture can capture more",
+        "bbox": [
+            511,
+            780,
+            908,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "28052",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "intricate features and dependencies at the cost of increased computational complexity. For our architecture the depth is set to 4. Ablation experiments with different transformer depths are provided in Sec. 4.4.",
+        "bbox": [
+            89,
+            90,
+            483,
+            151
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3.1 Encoder Block",
+        "text_level": 1,
+        "bbox": [
+            89,
+            172,
+            246,
+            186
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Each encoder block consists of the following components:",
+        "bbox": [
+            89,
+            196,
+            478,
+            212
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Multi-Head Self-Attention Layer: This layer allows the model to focus on different parts of the input sequence simultaneously. With 12 attention heads in our architecture, it captures diverse interactions and dependencies among frames. Each head computes attention scores independently, learning multiple representations of the input data. This enhances the model's ability to detect subtle temporal variations, which is crucial for video classification. The use of scaled dot-product attention further helps mitigate large gradients and stabilizes training.",
+        "bbox": [
+            89,
+            213,
+            483,
+            364
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Layer Normalization and Residual Connections: To aid gradient flow during training, each sub-layer incorporates residual connections that add the input back to the output, mitigating vanishing gradient issues in deep networks. Layer normalization follows to stabilize activations and enhance convergence speed. Additionally, dropout is applied post-attention to reduce overfitting by randomly deactivating a portion of neurons during training.",
+        "bbox": [
+            89,
+            364,
+            483,
+            486
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Feed-Forward Network (MLP): The second sub-layer is a point-wise feed-forward network with two dense layers separated by a GELU activation [23]. This MLP enables nonlinear feature transformations, enhancing the model's expressiveness. By projecting attention outputs into a higher-dimensional space before reducing them, it captures complex interactions beyond linear transformations.",
+        "bbox": [
+            89,
+            486,
+            483,
+            590
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Attention maps are derived from the outputs of the encoder blocks to be used for computing the Attention-Diversity loss (Sec. 3.4) and to illustrate how the model prioritizes different spatial regions in the frames during the classification process (Fig. 3). This interpretability is valuable for understanding the model's decision-making, particularly in complex video classification scenarios.",
+        "bbox": [
+            89,
+            592,
+            483,
+            698
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3.2 Positional Encoding",
+        "text_level": 1,
+        "bbox": [
+            89,
+            719,
+            284,
+            734
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We utilize a sine-cosine positional encoding scheme following the original transformer [51] to provide a unique position identifier for each input token. For every frame  $f_{j}$  in the video sample with encoded feature dimension  $d_{s}$  (as per Sec. 3.2), we compute the positional encoding for odd and even indexed feature dimensions as,",
+        "bbox": [
+            89,
+            743,
+            483,
+            834
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP E _ {(j, 2 i + 1)} = \\cos \\left(\\frac {j}{1 0 0 0 0 ^ {\\frac {2 i}{d _ {s}}}}\\right), \\quad P E _ {(j, 2 i)} = \\sin \\left(\\frac {j}{1 0 0 0 0 ^ {\\frac {2 i}{d _ {s}}}}\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            89,
+            857,
+            488,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where,  $i$  denotes the feature index in the encoded feature dimension  $d_{s}$ . We embed these encodings directly into the input tokens of our video transformer to provide positional context, for better temporal modeling during the attention computation with minimal computational overhead.",
+        "bbox": [
+            511,
+            90,
+            905,
+            167
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4. Attention-Diversity Loss",
+        "text_level": 1,
+        "bbox": [
+            511,
+            174,
+            738,
+            191
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Training a multi-head attention transformer with only cross-entropy loss makes the attention maps focus only on the face regions of the frame (as evident in Fig. 3). However, we aim to detect fake videos where the manipulation might not be in the face at all, and instead be in the background (like video inpainting with the AVID model [61]). To ensure that the attention heads focus on diverse spatial regions of the video frames, we devise the \"Attention-Diversity\" (AD) loss.",
+        "bbox": [
+            511,
+            198,
+            905,
+            318
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "AD-loss is designed to minimize overlap among attention maps (obtained from the first encoder block of the trainable transformer architecture in 3.3.1) while maintaining consistency across different input video samples. The attention outputs from the video transformer model  $\\mathcal{A} \\in \\mathbb{R}^{n_h \\times t_s \\times d_s}$  are used to pool the input SigLIP-So400m features  $\\xi \\in \\mathbb{R}^{n_f \\times t_s \\times d_s}$ , resulting in a pooled feature tensor  $\\mathcal{P} \\in \\mathbb{R}^{n_h \\times n_f}$  according to,",
+        "bbox": [
+            511,
+            319,
+            906,
+            439
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {P} = \\sum_ {j = 1} ^ {t _ {s}} \\sum_ {k = 1} ^ {d _ {s}} \\mathcal {A} _ {h, j, k} \\cdot \\xi_ {f, j, k}, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            614,
+            450,
+            905,
+            493
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $f$  indexes the frames  $(n_f)$ ,  $h$  indexes the number of attention heads  $(n_h)$ , and  $j, k$  are spatial positions.",
+        "bbox": [
+            511,
+            497,
+            905,
+            527
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We compute \"feature centers\" which are points in the feature space that serve as anchors for the learned representations of different classes. These centers represent the average features of samples from a particular class, allowing the model to capture essential characteristics of that class. In the context of the AD-loss, feature centers  $\\mathcal{C} \\in \\mathbb{R}^{n_h \\times n_f}$  are dynamically updated in each training iteration  $\\tau$  based on the pooled feature vectors  $\\mathcal{P}$ , following,",
+        "bbox": [
+            511,
+            527,
+            905,
+            648
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {C} ^ {\\tau} = \\mathcal {C} ^ {\\tau - 1} - \\eta \\left(\\mathcal {C} ^ {\\tau - 1} - \\frac {1}{B} \\sum_ {b = 1} ^ {B} \\mathcal {P} _ {b}\\right), \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            659,
+            905,
+            700
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\eta$  represents the learning rate (set to 0.05 in our experiments) for feature center updates and  $B$  is the batch dimension. The feature centers are initialized with zeros  $(\\mathcal{C}^0 = [0]_{n_h\\times n_f})$  at the beginning of the model training.",
+        "bbox": [
+            511,
+            704,
+            905,
+            765
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "AD-loss consists of two distinct components: within-class loss and between-class loss inspired by the Fisher Discriminant Analysis for deep networks [22]. The within-class term calculates the distance between the pooled feature vectors  $\\mathcal{P}$  and their corresponding feature centers  $\\mathcal{C}$ , promoting closeness among similar samples. The between-class term measures the distance between feature centers of different classes, encouraging them to be spaced apart in the feature space, encouraging separability.",
+        "bbox": [
+            511,
+            766,
+            906,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "28053",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The within-class loss term is calculated as",
+        "bbox": [
+            109,
+            90,
+            392,
+            104
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {w i t h i n}} = \\max  \\left(\\| \\mathcal {P} - \\mathcal {C} \\| _ {2} - \\delta_ {\\text {w i t h i n}}, 0\\right), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            156,
+            119,
+            480,
+            137
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\delta_{\\mathrm{within}} \\in \\mathbb{R}^{n_c}$  is a hyperparameter controlling the allowable distance between feature vectors and their respective feature centers for inputs of the same class, encouraging tighter clustering and  $\\| \\cdot \\|_2$  represents  $L_2$  normalization.",
+        "bbox": [
+            89,
+            140,
+            482,
+            200
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The between-class loss is calculated as",
+        "bbox": [
+            109,
+            200,
+            369,
+            215
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {b e t w e e n}} = \\sum_ {\\substack {k \\neq l \\\\ (k, l) \\in \\left(n _ {h}, n _ {h}\\right)}} \\max  \\left(\\delta_ {\\text {b e t w e e n}} - \\| \\mathcal {C} _ {k} - \\mathcal {C} _ {l} \\| _ {2}, 0\\right), \\tag{5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            99,
+            234,
+            480,
+            290
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\delta_{\\mathrm{between}}$  is a predefined hyperparameter that ensures a minimum distance between feature centers of different classes. Thus the AD-loss is a simple addition of these two components  $\\mathcal{L}_{AD} = \\mathcal{L}_{\\mathrm{within}} + \\mathcal{L}_{\\mathrm{between}}$ , making the final objective function for training the UNITE model,",
+        "bbox": [
+            89,
+            291,
+            483,
+            367
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {U N I T E}} = \\lambda_ {1} \\cdot \\mathcal {L} _ {C E} + \\lambda_ {2} \\cdot \\mathcal {L} _ {A D}, \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            176,
+            381,
+            482,
+            397
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{L}_{CE}$  is the traditional cross-entropy loss, and  $\\lambda_{1}$  and  $\\lambda_{2}$  are loss-weighting hyperparameters.",
+        "bbox": [
+            89,
+            400,
+            483,
+            431
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            441,
+            223,
+            458
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Datasets",
+        "text_level": 1,
+        "bbox": [
+            89,
+            465,
+            192,
+            481
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Most DeepFake datasets focus on face manipulations, such as face-swaps or lip-syncing, and lack fully AI-generated video content. The DeMamba dataset [9] features videos from T2V/I2V models, including powerful diffusion models like SORA [5]. It is not a human-centric dataset and primarily contains videos from natural scenes. We use its validation split to evaluate UNITE.",
+        "bbox": [
+            89,
+            489,
+            483,
+            595
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To train UNITE, we employ the FaceForensics++ (FF++) [43] dataset (c23 i.e., medium-quality) and the SAIL-VOS-3D dataset [24], which includes synthetic (although not AI-generated) GTA-V game videos featuring human subjects suitable for DeepFake detection. To evaluate the UNITE model, we use:",
+        "bbox": [
+            89,
+            595,
+            483,
+            686
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Face manipulated data:  $\\mathrm{FF}++$  [43] (in-domain evaluation), CelebDF [30], DeeperForensics [25], Deepfake-TIMIT [27], HifiFace [53], UADFV [58].",
+            "- Background manipulated data: Sample videos from the AVID [61] model provided publicly by the authors.",
+            "- Fully synthetic data: GTA-V [24] (in-domain evaluation), DeMamba [9].",
+            "- In-the-wild DeepFakes: Publicly available videos from the New York Times DeepFake quiz [48]."
+        ],
+        "bbox": [
+            89,
+            686,
+            482,
+            823
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Results and Discussion",
+        "text_level": 1,
+        "bbox": [
+            89,
+            832,
+            303,
+            847
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metrics: Although most DeepFake detection methods only report accuracy, we evaluate UNITE using additional metrics: (1) AUC (Area Under the Precision-Recall",
+        "bbox": [
+            89,
+            854,
+            483,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Curve); (2) Precision@0.5 and (3) Recall@0.5 at a 0.5 confidence threshold; (4) Precs@Rec=0.8 (precision at 0.8 recall); (5) Rec@Precs=0.8 (recall at 0.8 precision).",
+        "bbox": [
+            511,
+            90,
+            903,
+            136
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training Details: UNITE is trained using an AdamW optimizer [33] with an initial learning rate of 0.0001 and a decay rate of 0.5 every 1000 steps. For AD-loss,  $\\delta_{\\mathrm{between}} = 0.5$  and  $\\delta_{\\mathrm{within}}$  is set to  $[0.01, -2]$  for binary and  $[0.01, -2, 1]$  for fine-grained classification in Sec. 4.3 (see supplementary for sensitivity). Loss weights are  $\\lambda_1 = \\lambda_2 = 0.5$ . Training uses a batch size of 32 for 25 epochs on 8 TPUv3 chips, with the framework implemented in TensorFlow.",
+        "bbox": [
+            511,
+            137,
+            905,
+            258
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Quantitative Results: The results obtained by the UNITE model when trained on  $\\mathrm{FF}++$  [43] alone, and when trained with both  $\\mathrm{FF}++$  [43] and GTA-V [24] data and evaluated on the various datasets mentioned in Sec. 4.1 are shown in Table 1. On the AVID [61] and DeMamba datasets [9], the model performs poorly when trained with only face manipulated data, but the performance enhances several-fold when the GTA-V [24] synthetic data, even though it is not AI-generated, is used in training.",
+        "bbox": [
+            511,
+            261,
+            906,
+            397
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Interestingly, when evaluating on face-manipulated datasets such as CelebDF [30] and DeeperForensics [25], we observe a performance boost, particularly for CelebDF [30], when GTA-V [24] data is included in the training set. Intuitively, one would expect that training solely on  $\\mathrm{FF}++$  [43] would yield similar results since the inference set consists of similar face-manipulated data. This deviation from the expected behavior comes from the contribution of the AD-Loss, as the ablation study (Fig. 4) using only CE-loss compared to both CE and AD-losses reveals that the performance is enhanced with GTA-V [24] data in training only in the latter (detailed discussion in Sec. 4.4).",
+        "bbox": [
+            511,
+            398,
+            905,
+            579
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To evaluate UNITE on in-the-wild DeepFakes, we attempted the recent New York Times quiz [48] (denoted by NYTimes) which provides 10 videos (4 real and 6 fake) for testing DeepFake detectors. UNITE got 8 of those 10 videos correct, even though some of the face-swap videos were indistinguishable from real videos, even by humans. On all the AI-generated videos, UNITE made the correct prediction (when trained on  $\\mathrm{FF} + +$  [43] and GTA-V [24]).",
+        "bbox": [
+            511,
+            582,
+            905,
+            702
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Attention Heatmaps: To analyze the impact of CE and AD-losses on attention distribution, we extracted attention features from the first encoder of the UNITE model, comparing models trained with CE loss alone versus that trained with both CE and AD-losses. The resulting heatmaps, shown in Fig. 3, reveal that the model trained with only CE loss tends to concentrate primarily on the face region. In contrast, the model trained with the combined  $\\mathrm{CE + AD}$  loss demonstrates a broader attention span across the frame, as evidenced by a lighter, more distributed bluish tone in the heatmaps, indicating increased spatial diversity in the model's focus. This is why on the example shown in the first row (face-manipulated video) both models give the correct",
+        "bbox": [
+            511,
+            704,
+            906,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "28054",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/42739651164fe353323c6279da5dce08e66572740d45c8b707abe18c536a5005.jpg",
+        "table_caption": [
+            "Table 1. Results from the UNITE model trained with (1) FF++ [43] only and (2) FF++ [43] combined with GTA-V [24]. All other results reflect cross-dataset evaluations except for FF++ and GTA-V (when trained). Performance gains are highlighted in green."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Train</td><td colspan=\"7\">Test</td></tr><tr><td>FF++</td><td>GTA-V</td><td>Dataset</td><td>Accuracy</td><td>AUC</td><td>Precision@0.5</td><td>Recall@0.5</td><td>Precs@Rec=0.8</td><td>Rec@Precs=0.8</td></tr><tr><td colspan=\"9\">Face Manipulated Data</td></tr><tr><td>✓</td><td></td><td>FF++</td><td>99.53%</td><td>99.77%</td><td>99.94%</td><td>99.49%</td><td>99.94%</td><td>99.94%</td></tr><tr><td>✓</td><td></td><td>CelebDF</td><td>72.61%</td><td>94.05%</td><td>96.45%</td><td>61.22%</td><td>80.45%</td><td>61.22%</td></tr><tr><td>✓</td><td></td><td>DeeperForensics</td><td>91.35%</td><td>100.00%</td><td>100.00%</td><td>91.35%</td><td>100.00%</td><td>91.35%</td></tr><tr><td>✓</td><td></td><td>DeepfakeTIMIT</td><td>86.90%</td><td>86.46%</td><td>83.61%</td><td>83.97%</td><td>88.90%</td><td>81.33%</td></tr><tr><td>✓</td><td></td><td>HiFiFace</td><td>63.63%</td><td>62.47%</td><td>67.12%</td><td>63.63%</td><td>59.30%</td><td>63.63%</td></tr><tr><td>✓</td><td></td><td>UADFV</td><td>94.12%</td><td>94.38%</td><td>95.68%</td><td>97.11%</td><td>93.79%</td><td>94.38%</td></tr><tr><td>✓</td><td>✓</td><td>FF++</td><td>99.96%(+0.43)</td><td>99.89%(+0.12)</td><td>100.00%(+0.06)</td><td>99.84%(+0.35)</td><td>100.00%(+0.06)</td><td>99.96%(+0.02)</td></tr><tr><td>✓</td><td>✓</td><td>CelebDF</td><td>95.11%(+22.50)</td><td>94.36%(+0.31)</td><td>96.82%(+0.37)</td><td>68.75%(+7.53)</td><td>96.53%(+16.08)</td><td>68.75%(+7.53)</td></tr><tr><td>✓</td><td>✓</td><td>DeeperForensics</td><td>99.62%(+8.27)</td><td>100.00%(+0.00)</td><td>100.00%(+0.00)</td><td>99.62%(+8.27)</td><td>100.00%(+0.00)</td><td>99.63%(+8.28)</td></tr><tr><td>✓</td><td>✓</td><td>DeepfakeTIMIT</td><td>91.90%(+5.00)</td><td>91.33%(+4.87)</td><td>90.45%(+6.84)</td><td>88.39%(+4.42)</td><td>100.00%(+11.10)</td><td>91.95%(+10.62)</td></tr><tr><td>✓</td><td>✓</td><td>HiFiFace</td><td>75.62%(+11.99)</td><td>81.24%(+18.77)</td><td>79.55%(+12.43)</td><td>71.71%(+8.08)</td><td>75.62%(+16.32)</td><td>72.47%(+8.84)</td></tr><tr><td>✓</td><td>✓</td><td>UADFV</td><td>97.01%(+2.89)</td><td>94.95%(+0.57)</td><td>96.89%(+1.21)</td><td>100.00%(+2.89)</td><td>94.12%(+0.33)</td><td>100.00%(+5.62)</td></tr><tr><td colspan=\"9\">Background Manipulated Data</td></tr><tr><td>✓</td><td></td><td>AVID</td><td>41.67%</td><td>33.33%</td><td>33.33%</td><td>41.67%</td><td>41.67%</td><td>33.33%</td></tr><tr><td>✓</td><td>✓</td><td>AVID</td><td>100.00%(+58.33)</td><td>100.00%(+66.67)</td><td>100.00%(+66.67)</td><td>100.00%(+58.33)</td><td>100.00%(+58.33)</td><td>100.00%(+66.67)</td></tr><tr><td colspan=\"9\">Fully Synthetic Data</td></tr><tr><td>✓</td><td></td><td>GTA-V</td><td>60.16%</td><td>61.52%</td><td>60.16%</td><td>58.73%</td><td>63.29%</td><td>58.73%</td></tr><tr><td>✓</td><td></td><td>DeMamba</td><td>61.47%</td><td>57.38%</td><td>67.73%</td><td>33.01%</td><td>62.15%</td><td>54.16%</td></tr><tr><td>✓</td><td>✓</td><td>GTA-V</td><td>100.00%(+39.84)</td><td>100.00%(+38.48)</td><td>100.00%(+39.84)</td><td>100.00%(+41.27)</td><td>100.00%(+36.71)</td><td>100.00%(+41.27)</td></tr><tr><td>✓</td><td>✓</td><td>DeMamba</td><td>87.12%(+25.65)</td><td>93.75%(+36.67)</td><td>92.76%(+25.03)</td><td>89.60%(+56.59)</td><td>89.81%(+27.66)</td><td>92.12%(+37.96)</td></tr><tr><td colspan=\"9\">In-the-wild DeepFakes</td></tr><tr><td>✓</td><td></td><td>NYTimes [48]</td><td>50.00%</td><td>53.74%</td><td>50.00%</td><td>25.00%</td><td>0.00%</td><td>0.00%</td></tr><tr><td>✓</td><td>✓</td><td>NYTimes [48]</td><td>80.00%(+30.00)</td><td>97.42%(+43.68)</td><td>83.33%(+33.33)</td><td>83.33%(+58.33)</td><td>60.00%(+60.00)</td><td>100.00%(+100.00)</td></tr></table>",
+        "bbox": [
+            171,
+            117,
+            821,
+            358
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/026a53759fa5e9c90ff0e7f220988e58d42c475cf1f306716a2082362aedd22d.jpg",
+        "table_caption": [
+            "Table 2. SOTA Comparison on Face-Manipulated Data: We compared the performance of UNITE with recent DeepFake detectors, in terms of detection accuracy on various face manipulated datasets. UNITE outperforms the existing methods. Bold shows the current best results and the previous best and second-best results are highlighted in red and blue respectively."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>FF++</td><td>CelebDF</td><td>DeeperForensics</td><td>UADFV</td></tr><tr><td>TALL [57]</td><td>98.65%</td><td>90.79%</td><td>99.62%</td><td>-</td></tr><tr><td>ISTVT [62]</td><td>99.00%</td><td>84.10%</td><td>98.60%</td><td>-</td></tr><tr><td>Concas et al. [13]</td><td>99.49%</td><td>-</td><td>-</td><td>-</td></tr><tr><td>PUDD [40]</td><td>-</td><td>95.10%</td><td>-</td><td>-</td></tr><tr><td>TI2Net [32]</td><td>99.95%</td><td>68.22%</td><td>76.08%</td><td>-</td></tr><tr><td>LRNet [46]</td><td>99.89%</td><td>53.20%</td><td>56.77%</td><td>-</td></tr><tr><td>Choi et al. [11]</td><td>-</td><td>89.00%</td><td>99.00%</td><td>-</td></tr><tr><td>Lin et al. [31]</td><td>98.28%</td><td>74.42%</td><td>-</td><td>-</td></tr><tr><td>Guo et al. [21]</td><td>99.24%</td><td>84.97%</td><td>-</td><td>-</td></tr><tr><td>Li et al. (Res-152) [29]</td><td>-</td><td>-</td><td>-</td><td>93.80%</td></tr><tr><td>HeadPose [58]</td><td>-</td><td>-</td><td>-</td><td>89.00%</td></tr><tr><td>CViT [56]</td><td>93.00%</td><td>-</td><td>-</td><td>93.75%</td></tr><tr><td>FakeCatcher [12]</td><td>94.65%</td><td>91.50%</td><td>-</td><td>-</td></tr><tr><td>MesoNet [3]</td><td>-</td><td>-</td><td>-</td><td>82.40%</td></tr><tr><td>UNITE (Ours)</td><td>99.96%</td><td>95.11%</td><td>99.62%</td><td>97.01%</td></tr></table>",
+        "bbox": [
+            109,
+            449,
+            457,
+            607
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "prediction (\"fake\") with high confidence, but on the row-2 example (video generated from the Sora [6] T2V model) the model trained with only CE loss predicted \"real\" with a confidence score of  $99.25\\%$  (which is wrong), but the model trained with  $\\mathrm{CE + AD}$  losses predicted \"fake\" with a confidence of  $100.00\\%$ , which is correct.",
+        "bbox": [
+            89,
+            611,
+            483,
+            700
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Comparison on Face Manipulated Data: UNITE is compared with recent state-of-the-art (SOTA) DeepFake detection models in Table 2. The methods we compared against were specifically designed for detecting face manipulations and often crop faces from the videos, during both training and inference of the models. However, UNITE, which is designed to detect a diverse range of fake videos, still outperforms these detectors.",
+        "bbox": [
+            89,
+            703,
+            483,
+            823
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Comparison on Synthetic Data: We evaluated UNITE against state-of-the-art synthetic video detectors on the DeMamba dataset [9] (validation split), with results in Table 3. While existing detectors were trained on the DeMamba train split and validated on the DeMamba validation split,",
+        "bbox": [
+            89,
+            825,
+            483,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/c22e4507b864d24b58be75ce23f856e165498a1d17c85f7886055912690090c3.jpg",
+        "image_caption": [
+            "(a) Original Frame"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            526,
+            359,
+            647,
+            411
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/9a7f55b96974cf7b358b430d9ef5d44c715dd9f04ae58d9d55339da946fc69ed.jpg",
+        "image_caption": [
+            "(b) CE Loss only"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            650,
+            359,
+            769,
+            411
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/a73dc7545cdac92a63c283887cef990df632ac6489198467c2b382e0b2d6c15f.jpg",
+        "image_caption": [
+            "(c)  $\\mathrm{CE + AD}$  Loss"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            359,
+            893,
+            411
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/66b04e8a8f21be849327484594cfd6b4ecf64172824ab28252ce8e38960c5b84.jpg",
+        "image_caption": [
+            "(d) Original Frame"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            527,
+            424,
+            647,
+            477
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/c6f70e908b9eb1083ccd7c19c4d8aafcce441fe1ac95d4a3093f9532bc7b1cf2.jpg",
+        "image_caption": [
+            "(e) CE Loss only"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            650,
+            424,
+            769,
+            477
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/f78bfdc6431dce203fc1738fdbc24287c4c17126ed02ee3507b3c365f47ad81c.jpg",
+        "image_caption": [
+            "(f)  $\\mathrm{CE + AD}$  Loss",
+            "Figure 3. Comparison of attention heatmaps when UNITE is trained with only CE loss (second column) versus CE and AD-losses (third column). The CE-loss heatmap predominantly focuses on the face region, whereas the  $\\mathrm{CE + AD}$  loss heatmap demonstrates more distributed attention across the entire frame, as indicated by a broader bluish tone in the heatmaps. Both samples are \"fake\" videos- (a) is taken from Celeb-DF [30] and (d) is taken from DeMamba [9] and generated by OpenAI's Sora [6] (Best viewed as GIFs provided in the supplementary material). Yellow to blue shades show increasing attention in the plot."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            424,
+            893,
+            477
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "UNITE was trained on FF++ [43] and GTA-V [24]. Despite not being trained on DeMamba, UNITE's average performance (including real videos) surpassed current SOTA detectors, with competitive generator-wise results.",
+        "bbox": [
+            511,
+            638,
+            906,
+            699
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Thus, UNITE reliably detects a diverse range of fake videos (even in cross-domain settings), including T2V/I2V videos and background manipulations, while consistently maintaining state-of-the-art performance on traditional face-manipulated data. UNITE eliminates the need to have separate DeepFake and T2V/I2V video detector models by by handling both tasks within a single trained model.",
+        "bbox": [
+            511,
+            700,
+            906,
+            806
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Finegrained Evaluation",
+        "text_level": 1,
+        "bbox": [
+            511,
+            816,
+            732,
+            832
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In fake media detection, flagging videos partially or fully manipulated by AI also necessitates fine-grained classification into three classes: (1) real videos, (2) partially manipulated videos, and (3) fully synthetic videos. While binary",
+        "bbox": [
+            511,
+            839,
+            906,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "28055",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/285d59e5e6e07288f95b3fdfee0aaabba479d0a223322ee43e05064bdcead95c.jpg",
+        "table_caption": [
+            "Table 3. SOTA Comparison on Synthetic Data: On the DeMamba dataset (validation split), we compare the performance of UNITE, which was NOT trained on DeMamba train split, against state-of-the-art detectors which were trained on DeMamba train split (results taken from Chen et al. [9]). We report the results ( $P = \\text{Precision}$  and  $R = \\text{Recall}$ ) on the individual T2V/I2V generators and the average performance across the entire validation set ( $Avg$ , which also includes real videos). Although the direct comparison is unfair against UNITE which was trained with FF++ [43] and GTA-V [24], our method still outperforms these synthetic video detectors. Bold shows the current best results and the previous best and second-best results are highlighted in red and blue respectively."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Metrics</td><td>Sora [6]</td><td>Morph Studio [2]</td><td>Runway ML (Gen2) [42]</td><td>HotShot [38]</td><td>Lavie [54]</td><td>Show-1 [60]</td><td>Moon Valley [37]</td><td>Crafter [8]</td><td>Model Scope [52]</td><td>Wild Scrape [9]</td><td>Avg</td></tr><tr><td rowspan=\"2\">TALL [57]</td><td>P</td><td>71.15%</td><td>96.89%</td><td>98.51%</td><td>79.38%</td><td>84.59%</td><td>79.38%</td><td>98.79%</td><td>99.02%</td><td>92.70%</td><td>76.47%</td><td>87.91%</td></tr><tr><td>R</td><td>91.07%</td><td>98.28%</td><td>97.83%</td><td>83.00%</td><td>76.57%</td><td>79.57%</td><td>99.52%</td><td>98.93%</td><td>94.14%</td><td>66.31%</td><td>88.52%</td></tr><tr><td rowspan=\"2\">F3Net [41]</td><td>P</td><td>68.27%</td><td>99.89%</td><td>99.67%</td><td>89.35%</td><td>57.00%</td><td>36.57%</td><td>99.52%</td><td>99.71%</td><td>93.80%</td><td>88.41%</td><td>88.73%</td></tr><tr><td>R</td><td>83.93%</td><td>99.71%</td><td>98.62%</td><td>77.57%</td><td>85.24%</td><td>63.17%</td><td>99.58%</td><td>99.89%</td><td>89.43%</td><td>76.78%</td><td>81.88%</td></tr><tr><td rowspan=\"2\">NPR [47]</td><td>P</td><td>91.07%</td><td>99.57%</td><td>99.49%</td><td>24.29%</td><td>89.64%</td><td>57.71%</td><td>97.12%</td><td>99.86%</td><td>94.29%</td><td>87.80%</td><td>82.45%</td></tr><tr><td>R</td><td>91.07%</td><td>99.57%</td><td>99.49%</td><td>24.29%</td><td>89.64%</td><td>57.71%</td><td>97.12%</td><td>99.86%</td><td>94.29%</td><td>87.80%</td><td>84.08%</td></tr><tr><td rowspan=\"2\">STIL [20]</td><td>P</td><td>57.21%</td><td>99.08%</td><td>99.32%</td><td>86.19%</td><td>82.24%</td><td>70.43%</td><td>99.25%</td><td>98.96%</td><td>97.18%</td><td>81.32%</td><td>87.12%</td></tr><tr><td>R</td><td>67.86%</td><td>96.00%</td><td>98.41%</td><td>96.14%</td><td>77.14%</td><td>80.43%</td><td>97.44%</td><td>96.93%</td><td>96.29%</td><td>68.36%</td><td>82.22%</td></tr><tr><td rowspan=\"2\">MINTIME-CLIP-B [9]</td><td>P</td><td>83.21%</td><td>99.99%</td><td>99.67%</td><td>50.84%</td><td>99.20%</td><td>99.27%</td><td>99.76%</td><td>99.99%</td><td>91.83%</td><td>91.77%</td><td>91.55%</td></tr><tr><td>R</td><td>89.29%</td><td>100.00%</td><td>98.99%</td><td>26.43%</td><td>96.79%</td><td>98.14%</td><td>99.84%</td><td>100.00%</td><td>84.29%</td><td>82.38%</td><td>87.62%</td></tr><tr><td rowspan=\"2\">FTCN-CLIP-B [9]</td><td>P</td><td>91.79%</td><td>99.99%</td><td>99.79%</td><td>45.94%</td><td>99.76%</td><td>97.80%</td><td>99.99%</td><td>99.99%</td><td>94.69%</td><td>92.32%</td><td>92.21%</td></tr><tr><td>R</td><td>87.50%</td><td>100.00%</td><td>98.91%</td><td>17.71%</td><td>97.71%</td><td>91.86%</td><td>100.00%</td><td>100.00%</td><td>85.29%</td><td>82.83%</td><td>86.18%</td></tr><tr><td rowspan=\"2\">CLIP-B-PT [9]</td><td>P</td><td>67.80%</td><td>43.56%</td><td>70.88%</td><td>29.97%</td><td>52.97%</td><td>35.36%</td><td>55.52%</td><td>66.03%</td><td>44.23%</td><td>42.99%</td><td>44.83%</td></tr><tr><td>R</td><td>85.71%</td><td>82.43%</td><td>90.36%</td><td>71.00%</td><td>79.29%</td><td>75.43%</td><td>89.62%</td><td>86.29%</td><td>82.14%</td><td>75.16%</td><td>81.74%</td></tr><tr><td rowspan=\"2\">DeMamba-CLIP-PT [9]</td><td>P</td><td>25.87%</td><td>95.14%</td><td>96.23%</td><td>73.43%</td><td>83.31%</td><td>75.49%</td><td>90.17%</td><td>95.06%</td><td>95.05%</td><td>69.95%</td><td>79.97%</td></tr><tr><td>R</td><td>58.93%</td><td>96.43%</td><td>93.12%</td><td>68.00%</td><td>69.36%</td><td>69.00%</td><td>89.14%</td><td>91.86%</td><td>96.14%</td><td>56.59%</td><td>78.86%</td></tr><tr><td rowspan=\"2\">XCLIP-B-PT [9]</td><td>P</td><td>16.39%</td><td>72.16%</td><td>87.77%</td><td>39.86%</td><td>65.57%</td><td>54.26%</td><td>75.23%</td><td>84.80%</td><td>61.60%</td><td>55.28%</td><td>61.29%</td></tr><tr><td>R</td><td>81.34%</td><td>82.15%</td><td>83.35%</td><td>80.98%</td><td>81.82%</td><td>81.55%</td><td>82.14%</td><td>82.98%</td><td>81.93%</td><td>81.10%</td><td>81.93%</td></tr><tr><td rowspan=\"2\">DeMamba-XCLIP-PT [9]</td><td>P</td><td>18.26%</td><td>93.50%</td><td>94.72%</td><td>69.94%</td><td>78.08%</td><td>71.50%</td><td>83.95%</td><td>92.23%</td><td>93.54%</td><td>68.10%</td><td>76.38%</td></tr><tr><td>R</td><td>66.07%</td><td>95.86%</td><td>94.64%</td><td>77.86%</td><td>75.36%</td><td>80.29%</td><td>90.89%</td><td>92.50%</td><td>96.00%</td><td>66.41%</td><td>83.59%</td></tr><tr><td rowspan=\"2\">XCLIP-B-FT [9]</td><td>P</td><td>64.42%</td><td>99.73%</td><td>96.78%</td><td>70.98%</td><td>90.35%</td><td>77.28%</td><td>97.34%</td><td>99.84%</td><td>82.01%</td><td>88.97%</td><td>86.77%</td></tr><tr><td>R</td><td>82.14%</td><td>99.57%</td><td>93.62%</td><td>61.29%</td><td>79.36%</td><td>69.71%</td><td>97.92%</td><td>99.79%</td><td>77.14%</td><td>83.59%</td><td>84.41%</td></tr><tr><td rowspan=\"2\">UNITE (Ours)</td><td>P</td><td>88.57%</td><td>100.00%</td><td>100.00%</td><td>90.16%</td><td>89.91%</td><td>98.34%</td><td>99.52%</td><td>100.00%</td><td>98.96%</td><td>92.56%</td><td>92.76%</td></tr><tr><td>R</td><td>92.11%</td><td>100.00%</td><td>94.62%</td><td>96.93%</td><td>98.12%</td><td>99.86%</td><td>98.69%</td><td>100.00%</td><td>96.29%</td><td>89.89%</td><td>89.60%</td></tr></table>",
+        "bbox": [
+            109,
+            172,
+            879,
+            411
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/06feaeed88e23c1a4a5632e9ccdd54161f908fe2f012c75ed66e2b745f5ebb78.jpg",
+        "table_caption": [
+            "Table 4. Results obtained by the UNITE model on finegrained classes: detecting whether a video is real, partially manipulated or fully AI-generated. Performance gains are mentioned in green."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Train</td><td colspan=\"2\">Test</td></tr><tr><td>FF++</td><td>GTA-V</td><td>Dataset</td><td>Accuracy</td></tr><tr><td>✓</td><td></td><td>FF++</td><td>96.46%</td></tr><tr><td>✓</td><td></td><td>CelebDF</td><td>60.40%</td></tr><tr><td>✓</td><td></td><td>DeeperForensics</td><td>71.04%</td></tr><tr><td>✓</td><td></td><td>DeepFakeTIMIT</td><td>80.68%</td></tr><tr><td>✓</td><td></td><td>HifiFace</td><td>43.27%</td></tr><tr><td>✓</td><td></td><td>UADFV</td><td>90.37%</td></tr><tr><td>✓</td><td></td><td>AVID</td><td>50.00%</td></tr><tr><td>✓</td><td></td><td>GTA-V</td><td>0.00%</td></tr><tr><td>✓</td><td></td><td>DeMamba</td><td>38.29%</td></tr><tr><td>✓</td><td>✓</td><td>FF++</td><td>97.70%(+1.24)</td></tr><tr><td>✓</td><td>✓</td><td>CelebDF</td><td>70.17%(+9.77)</td></tr><tr><td>✓</td><td>✓</td><td>DeeperForensics</td><td>80.59%(+9.55)</td></tr><tr><td>✓</td><td>✓</td><td>DeepFakeTIMIT</td><td>81.04%(+0.36)</td></tr><tr><td>✓</td><td>✓</td><td>HifiFace</td><td>64.20%(+20.93)</td></tr><tr><td>✓</td><td>✓</td><td>UADFV</td><td>92.61%(+1.24)</td></tr><tr><td>✓</td><td>✓</td><td>AVID</td><td>62.50%(+12.50)</td></tr><tr><td>✓</td><td>✓</td><td>GTA-V</td><td>100.00%(+100.00)</td></tr><tr><td>✓</td><td>✓</td><td>DeMamba</td><td>65.84%(+27.55)</td></tr></table>",
+        "bbox": [
+            148,
+            463,
+            418,
+            665
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "real vs. fake classification may suffice to curb misinformation, fine-grained classification offers greater explainability in otherwise black-box models.",
+        "bbox": [
+            89,
+            671,
+            482,
+            715
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Using this defined convention of finegrained classes, all the face manipulated datasets (CelebDF [30], DeeperForensics [25], etc.) and the videos from the AVID [61] model (background manipulations) fall under class-2, and the videos from the GTA-V [24] and the DeMamba [9] datasets fall in class-3. In this setup, only the final classification layer of the UNITE transformer architecture is changed to 3 neurons with softmax activation. The results obtained are presented in Table 4. As expected, when the model is trained only with  $\\mathrm{FF}++$  [43], the performance on fully synthetic data is poor (since the model has not seen any synthetic videos). When the GTA-V [24] data is also",
+        "bbox": [
+            89,
+            719,
+            482,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4e0ee692c4c5de2ef0684ec2901942f259d3d5e21d2f9aed9f0a466693c70f68.jpg",
+        "image_caption": [
+            "(a) UNITE trained on  $\\mathrm{FF} + +$  [43] only"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            540,
+            420,
+            880,
+            508
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/731f76544b4eb7ad0882ec6ef0bec3dce866c8aa404905fbfc7f87320a267252.jpg",
+        "image_caption": [
+            "(b) UNITE trained on  $\\mathrm{FF} + +$  [43] and GTA-V [24]",
+            "Figure 4. Ablation Results to show the effect of changing the loss functions used to train the UNITE model. The combination of the cross-entropy (CE) and Attention Diversity (AD) losses always performs better, and the contribution from the AD-loss component increases significantly when using fully synthetic data for training."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            540,
+            527,
+            879,
+            616
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "used in training, the performance is much higher.",
+        "bbox": [
+            511,
+            710,
+            836,
+            726
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            511,
+            739,
+            666,
+            756
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We conducted an analysis to assess the contribution of individual loss functions to the training of the UNITE model. The outcomes of this ablation study are presented in Fig. 4. Specifically, we evaluated the impact of the AD-loss in two scenarios: (1) when UNITE was trained solely on the FF++ dataset [43] (Fig. 4(a)), and (2) when UNITE was trained using a combination of the FF++ [43] and the GTA-V [24] datasets (Fig. 4(b)). The combination of the two losses always performed better than the individual losses.",
+        "bbox": [
+            511,
+            763,
+            906,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "28056",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e2581f9420b0e9c89632475fbb1f99e00a2b12ff8998a84ceffb81c5cc3e3920.jpg",
+        "image_caption": [
+            "Figure 5. No. of frames vs. Performance: Performance analysis of UNITE based on the number of frames sampled per video segment. The results illustrate that as the number of frames increases from 1 to 64 (context window), the detection accuracy improves, showcasing UNITE's ability to effectively capture temporal inconsistencies in fake videos."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            140,
+            93,
+            436,
+            229
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect on Synthetic Data: UNITE performed significantly better on the fully synthetic datasets GTA-V [24] and DeMamba [9] when the model was training includes GTA-V [24] data. However, for the DeMamba dataset (cross-data evaluation), the performance with only CE-loss is similar regardless of the training data. It is the AD-loss component that provides the accuracy boost in Fig 4(b).",
+        "bbox": [
+            89,
+            323,
+            482,
+            429
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect on Background-Manipulated Data: In Fig. 4(a), we observe that when UNITE was trained solely on  $\\mathrm{FF}++$  [43] using only the CE-loss, its performance on the AVID dataset [61] was  $0.00\\%$ . This outcome can be attributed to the fact that the fake data in  $\\mathrm{FF}++$  [43] contains exclusively face-manipulated videos, causing the UNITE transformer's attention heads to focus predominantly on the facial regions. As AVID [61] comprises only background-manipulated videos, UNITE failed to detect any of the fakes under such conditions. However, when the AD-loss was introduced, the attention heads of UNITE were encouraged to diversify their spatial focus, significantly improving performance. Incorporating synthetic data, as shown in Fig. 4(b), further enhanced this effect, achieving  $100\\%$  accuracy on AVID when using both CE-loss and AD-loss.",
+        "bbox": [
+            89,
+            430,
+            482,
+            656
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect on Face-Manipulated Data: For the evaluation on face-manipulated datasets such as CelebDF [30] and UADFV [58], one might expect comparable results between training solely on  $\\mathrm{FF} + +$  [43] and incorporating GTA-V [24] due to their perceived irrelevance in evaluating face-centric manipulations. However, as shown in Fig. 4(b), training with fully synthetic data leads to a substantial performance improvement. Notably, in both Fig. 4(a) and (b), using only CE loss yields similar results, but the inclusion of the AD-loss component significantly enhances overall detection accuracy. The diversity in the training data makes the UNITE model learn more discriminative attention map features through the AD-loss component, which in turn helps even in detecting face-manipulated videos.",
+        "bbox": [
+            89,
+            657,
+            482,
+            868
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Frames vs. Performance: For inference with UNITE, we evaluated the impact of temporal context by vary-",
+        "bbox": [
+            89,
+            869,
+            482,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/69c7bd286988af9c27ba996e301c20fc5c1e44c982413c17694203378e6c34a8.jpg",
+        "image_caption": [
+            "Figure 6. Transformer Depth Evaluation: Performance comparison of UNITE with varying the number of encoder blocks (depth). UNITE performs optimally in cross-domain settings when the depth is 4. Greater depths overfit to the training domains  $(\\mathrm{FF} + +$  [43] and GTA-V [24]), while a depth of 2 is insufficient to capture the complexity of the data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            563,
+            94,
+            859,
+            231
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "ing the number of frames sampled from video segments to  $\\{1,8,16,32,64\\}$ , with 64 frames corresponding to our model's context window. As shown in Fig. 5, performance improves as more frames are included, demonstrating that UNITE effectively captures temporal discontinuities in fake videos. This highlights the model's ability to leverage temporal cues for enhanced detection accuracy.",
+        "bbox": [
+            511,
+            329,
+            906,
+            434
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Transformer Depth Ablation: We evaluate the impact of varying the depth of the transformer architecture in UNITE by adjusting the number of encoder blocks, with results shown in Fig. 6. We tested depths of  $\\{2,4,6,8\\}$  and observed that the model achieved the best performance with 4 encoder blocks, which is the depth used in our final architecture. This suggests that a moderate depth strikes the optimal balance between model complexity and performance, especially in cross-dataset settings, providing robust detection capabilities without overfitting.",
+        "bbox": [
+            511,
+            435,
+            908,
+            587
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            511,
+            601,
+            633,
+            617
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Traditional DeepFake detectors, focused solely on face-manipulated content, struggle against newer, more sophisticated forgery methods involving T2V, I2V, and background manipulations. To address this, we proposed UNITE, leveraging the fully synthetic GTA-V [24] dataset for training. While this synthetic data is not AI-generated, our results show it effectively enhances UNITE's ability to detect AI-generated content and background manipulations.",
+        "bbox": [
+            511,
+            627,
+            905,
+            748
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "A key innovation in our approach is the AD-loss, which encourages the model to explore diverse spatial regions, enhancing its detection of manipulated areas beyond just the face region. This allows UNITE to excel not only in detecting AI-generated and background-altered videos but also in identifying traditional face-manipulated content. Our ablation studies demonstrate the robustness of UNITE, making it a valuable tool in synthetic video detection.",
+        "bbox": [
+            511,
+            750,
+            906,
+            869
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements: This work was supported in part by funding from YouTube (Google).",
+        "bbox": [
+            511,
+            869,
+            905,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "28057",
+        "bbox": [
+            478,
+            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] Flux 1.1. https://blackforestlabs.ai/, 2024.1",
+            "[2] Morph studio. https://www.morphstudio.com/, 2024.7",
+            "[3] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In 2018 IEEE international workshop on information forensics and security (WIFS), pages 1-7. IEEE, 2018. 6",
+            "[4] Ibrahim M Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. Advances in Neural Information Processing Systems, 36, 2024. 2, 3",
+            "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 5",
+            "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 6, 7",
+            "[7] Baoying Chen and Shunquan Tan. Featuretransfer: Unsupervised domain adaptation for cross-domain deepfake detection. Security and Communication Networks, 2021(1): 9942754, 2021. 3",
+            "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 7",
+            "[9] Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, et al. Demamba: Ai-generated video detection on million-scale genvideo benchmark. arXiv preprint arXiv:2405.19707, 2024. 1, 2, 5, 6, 7, 8",
+            "[10] Jikang Cheng, Zhiyuan Yan, Ying Zhang, Yuhao Luo, Zhongyuan Wang, and Chen Li. Can we leave deepfake data behind in training deepfake detector? arXiv preprint arXiv:2408.17052, 2024. 1",
+            "[11] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1133-1143, 2024. 2, 6",
+            "[12] Umur Aybars Ciftci, Ilke Demir, and Lijun Yin. Fakecatcher: Detection of synthetic portrait videos using biological signals. IEEE transactions on pattern analysis and machine intelligence, 2020. 6",
+            "[13] Sara Concas, Simone Maurizio La Cava, Roberto Casula, Giulia Orrù, Giovanni Puglisi, and Gian Luca Marcialis. Quality-based artifact modeling for facial deepfake detection in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3845–3854, 2024. 2, 6"
+        ],
+        "bbox": [
+            93,
+            114,
+            482,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[14] Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 973-982, 2023. 2",
+            "[15] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 2",
+            "[16] Davide Cozzolino, Andreas Rössler, Justus Thies, Matthias Nießner, and Luisa Verdoliva. Id-reveal: Identity-aware deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15108-15117, 2021. 2",
+            "[17] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 1",
+            "[18] 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 ICLR, 2021. 3",
+            "[19] Ricard Durall, Margret Keuper, Franz-Josef Pfundt, and Janis Keuper. Unmasking deepfakes with simple features. arXiv preprint arXiv:1911.00686, 2019. 2",
+            "[20] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. In Proceedings of the 29th ACM international conference on multimedia, pages 3473-3481, 2021. 7",
+            "[21] Ying Guo, Cheng Zhen, and Pengfei Yan. Controllable guide-space for generalizable face forgery detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20818-20827, 2023. 6",
+            "[22] Harald Hanselmann, Shen Yan, and Hermann Ney. Deep fisher faces. In BMVC, 2017. 4",
+            "[23] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4",
+            "[24] Yuan-Ting Hu, Jiahong Wang, Raymond A Yeh, and Alexander G Schwing. Sail-vos 3d: A synthetic dataset and baselines for object detection and 3d mesh reconstruction from video data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1418-1428, 2021. 1, 2, 5, 6, 7, 8",
+            "[25] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2889–2898, 2020. 5, 7",
+            "[26] Yan Ju, Chengzhe Sun, Shan Jia, Shuwei Hou, Zhaofeng Si, Soumyya Kanti Datta, Lipeng Ke, Riky Zhou, Anita Nikolich, and Siwei Lyu. Deepfake-o-meter v2.0: An open platform for deepfake detection. In 2024 IEEE 7th Inter-"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "28058",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "national Conference on Multimedia Information Processing and Retrieval (MIPR), pages 439-445. IEEE, 2024. 1",
+            "[27] Pavel Korshunov and Sébastien Marcel. Deepfakes: a new threat to face recognition? assessment and detection. arXiv preprint arXiv:1812.08685, 2018. 5",
+            "[28] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 1",
+            "[29] Y Li. Exposing deepfake videos by detecting face warping artifact acts. arXiv preprint arXiv:1811.00656, 2018. 6",
+            "[30] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3207-3216, 2020. 1, 3, 5, 6, 7, 8",
+            "[31] Li Lin, Xinan He, Yan Ju, Xin Wang, Feng Ding, and Shu Hu. Preserving fairness generalization in deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16815-16825, 2024. 6",
+            "[32] Baoping Liu, Bo Liu, Ming Ding, Tianqing Zhu, and Xin Yu. Ti2net: temporal identity inconsistency network for deepfake detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4691-4700, 2023. 2, 6",
+            "[33] I Loshchilov and F Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 5",
+            "[34] Qingxuan Lv, Yuezun Li, Junyu Dong, Sheng Chen, Hui Yu, Huiyu Zhou, and Shu Zhang. Domainforensics: Exposing face forgery across domains via bi-directional adaptation. IEEE Transactions on Information Forensics and Security, 2024. 3",
+            "[35] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2",
+            "[36] Ghazal Mazaheri and Amit K Roy-Chowdhury. Detection and localization of facial expression manipulations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1035-1045, 2022. 2",
+            "[37] moonvalley.ai. moonvalley.ai. https://moonvalley.ai/, 2022.7",
+            "[38] John Mullan, Duncan Crawbuck, and Aakash Sastry. Hotshot-XL, 2023. 7",
+            "[39] Dat Nguyen, Nesryne Mejri, Inder Pal Singh, Polina Kuleshova, Marcella Astrid, Anis Kacem, Enjie Ghorbel, and Djamila Aouada. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17395–17405, 2024. 2",
+            "[40] Alvaro Lopez Pellicer, Yi Li, and Plamen Angelov. Pudd: Towards robust multi-modal prototype-based deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3809-3817, 2024. 2, 6"
+        ],
+        "bbox": [
+            91,
+            92,
+            482,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[41] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 7",
+            "[42] Runway Research. Text driven video generation. https://research.runwayml.com/gen2, 2023.1,7",
+            "[43] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 1, 2, 3, 5, 6, 7, 8",
+            "[44] Md Shamim Seraj, Ankita Singh, and Shayok Chakraborty. Semi-supervised deep domain adaptation for deepfake detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1061-1071, 2024. 2",
+            "[45] Ke Sun, Shen Chen, Taiping Yao, Hong Liu, Xiaoshuai Sun, Shouhong Ding, and Rongrong Ji. Diffusionfake: Enhancing generalization in deepfake detection via guided stable diffusion. NeurIPS, 2024. 1",
+            "[46] Zekun Sun, Yujie Han, Zeyu Hua, Na Ruan, and Weijia Jia. Improving the efficiency and robustness of deepfakes detection through precise geometric features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3609-3618, 2021. 6",
+            "[47] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 7",
+            "[48] The New York Times. A.i. can now create lifelike videos. can you tell what's real? https://www.nytimes.com/interactive/2024/09/09/technology/ai-video-deepfake-runway-kling-quiz.html, 2024.5,6",
+            "[49] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 1",
+            "[50] Loc Trinh, Michael Tsang, Sirisha Rambhatla, and Yan Liu. Interpretable and trustworthy deepfake detection via dynamic prototypes. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1973-1983, 2021. 2",
+            "[51] A Vaswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A Gomez, L Kaiser, and I Polosukhin. Attention is all you need. NeurIPS, 2017. 4",
+            "[52] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2, 7",
+            "[53] Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, 2021. 3, 5"
+        ],
+        "bbox": [
+            516,
+            92,
+            906,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "28059",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[54] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 7",
+            "[55] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 2",
+            "[56] Deressa Wodajo and Solomon Atnafu. Deepfake video detection using convolutional vision transformer. arXiv preprint arXiv:2102.11126, 2021. 6",
+            "[57] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22658-22668, 2023. 1, 2, 6, 7",
+            "[58] Xin Yang, Yuezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8261-8265. IEEE, 2019. 5, 6, 8",
+            "[59] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3",
+            "[60] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision, pages 1–15, 2024. 7",
+            "[61] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 1, 4, 5, 7, 8",
+            "[62] Cairong Zhao, Chutian Wang, Guosheng Hu, Haonan Chen, Chun Liu, and Jinhui Tang. Istvt: interpretable spatial-temporal video transformer for deepfake detection. IEEE Transactions on Information Forensics and Security, 18: 1335-1348, 2023. 6",
+            "[63] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2",
+            "[64] Xinye Zhou, Hu Han, Shiguang Shan, and Xin Chen. Fine-grained open-set deepfake detection via unsupervised domain adaptation. IEEE Transactions on Information Forensics and Security, 2024. 2"
+        ],
+        "bbox": [
+            91,
+            90,
+            480,
+            796
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "28060",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/49360ca2-47e3-4c7b-81b1-6a87621996c4_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:353eb54ebe234c165fb2f8fff55dfd8e415c61627cd214944b17d4c881902ecd
+size 8006347

2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/full.md

@@ -0,0 +1,337 @@
+# Towards a Universal Synthetic Video Detector: From Face or Background Manipulations to Fully AI-Generated Content
+
+Rohit Kundu $^{1,2}$ , Hao Xiong $^{1}$ , Vishal Mohanty $^{1}$ , Athula Balachandran $^{1}$ , Amit K. Roy-Chowdhury $^{2}$ ,  $^{1}$ Google LLC,  $^{2}$ University of California, Riverside
+
+{rohit.kundu@email, amertc@ece}.ucr.edu; {rohitkun, haoxg, vishal, athula}@google.com
+
+# Abstract
+
+Existing DeepFake detection techniques primarily focus on facial manipulations, such as face-swapping or lip-syncing. However, advancements in text-to-video (T2V) and image-to-video (I2V) generative models now allow fully AI-generated synthetic content and seamless background alterations, challenging face-centric detection methods and demanding more versatile approaches.
+
+To address this, we introduce the Universal Network for Identifying Tampered and synthEtic videos (UNITE) model, which, unlike traditional detectors, captures full-frame manipulations. UNITE extends detection capabilities to scenarios without faces, non-human subjects, and complex background modifications. It leverages a transformer-based architecture that processes domain-agnostic features extracted from videos via the SigLIP-So400M foundation model. Given limited datasets encompassing both facial/background alterations and T2V/I2V content, we integrate task-irrelevant data alongside standard DeepFake datasets in training. We further mitigate the model's tendency to over-focus on faces by incorporating an attention-diversity (AD) loss, which promotes diverse spatial attention across video frames. Combining AD loss with cross-entropy improves detection performance across varied contexts. Comparative evaluations demonstrate that UNITE outperforms state-of-the-art detectors on datasets featuring face/background manipulations and fully synthetic T2V/I2V videos, showcasing its adaptability and generalizable detection capabilities.
+
+# 1. Introduction
+
+The rise of synthetic media, especially DeepFakes, has transformed content perception and interaction online. Hyper-realistic images produced by technologies like  $\mathrm{FLUX}_{1.1}$  [1] are challenging for even humans to identify as fake, and with the development of their video model underway, this underscores the critical need for effective detec
+
+![](images/7f4d537dbea8ce1cfa0f687526832537988d457109714d1137feb379a094bc2b.jpg)  
+Figure 1. Problem Overview: Existing DeepFake detection methods focus on identifying face-manipulated videos, most of which cannot perform inference unless there is a face detected in the video. However, with advancements like seamless background modifications (e.g., AVID [61]) and hyper-realistic content from games like GTA-V [24] and T2V/I2V models [9], a more comprehensive approach is needed. A model trained with only cross-entropy (CE) loss, using full frames, automatically focuses on the face, capturing temporal discontinuities through its transformer architecture, performing better than random  $(\approx)$  on T2V/I2V content but struggling with background manipulations. UNITE, with its attention-diversity (AD) loss, effectively detects both face/background manipulations and fully synthetic content.
+
+tion methods for media generated by deep neural networks. Traditional DeepFake generators [28, 49] focus on manipulating human faces through face-swapping and lip-syncing. However, powerful text-to-video (T2V) and image-to-video (I2V) models [42, 52, 63] have expanded manipulation possibilities beyond faces.
+
+While conventional detectors [10, 45, 57] perform well on older, face-centric DeepFake datasets [17, 30, 43], they often struggle with newer manipulations involving full scenes or backgrounds. DeepFake-O-Meter [26] is an open-source tool (consisting of several state-of-the-art models) for DeepFake detection, but it cannot run inference unless a human face is visible in the image/video. The rapid spread
+
+of misinformation, particularly during critical periods such as elections, highlights the need for generalizable detection models capable of identifying diverse manipulations, including face, background, and fully AI-generated T2V/I2V content with/without human subjects (overview in Fig. 1).
+
+To address these challenges, we present a Universal Network for Identifying Tampered and synthEtic videos or UNITE model, which detects both partially manipulated (foreground/background) and fully synthetic videos. Unlike detectors focused solely on face detection, UNITE analyzes entire frames, regardless of whether a human subject is present in the videos.
+
+Given the inherent domain gaps in DeepFake datasets [44, 64], even when generated by similar techniques, we leverage the SigLIP-So400m [4] foundation model to extract domain-agnostic features. This serves as inputs to a learnable transformer with multi-head attention, enabling effective detection by capturing temporal inconsistencies in synthetic content. However, preliminary experiments show that training a transformer architecture solely with cross-entropy (CE) loss often leads to the attention heads converging on the face region (Fig. 3). As a result, the model struggles during inference when handling videos where a real human subject is placed in a manipulated background or content generated by T2V [8, 52] and I2V [35, 63] models. To address this, we introduce an "attention-diversity" (AD) loss that encourages attention heads to focus on different spatial regions, enhancing the model's ability to capture critical cues from both foreground and background.
+
+Due to the limited availability of open-source datasets covering face or background manipulations and fully AI-generated content, we employ innovative training strategies for UNITE. This includes integrating task-irrelevant data with standard DeepFake datasets to simulate AI-generated synthetic media. Beyond the popularly used FaceForensics++ [43] dataset, we utilize the SAIL-VOS-3D [24] dataset, originally designed for 3D video object segmentation in the game GTA-V. As this dataset is fully synthetic, it helps simulate AI-generated media, enhancing our model's ability to detect diverse forms of synthetic manipulation.
+
+Our contributions can be summarized as follows:
+
+- We propose UNITE, a model for detecting partially manipulated (foreground/background) and fully synthetic videos, moving beyond face-centric DeepFake detection.  
+- Unlike detectors relying on face detection and cropping, our model can process full video frames and can detect fakes even if there are no human subjects.  
+- Using the SigLIP-So400m foundation model [4], we extract domain-agnostic features, enabling generalization to in-the-wild DeepFakes.  
+- We introduce an attention-diversity loss, encouraging attention heads to focus on diverse spatial regions, enhancing detection beyond face manipulations.
+
+- Unlike existing methods that evaluate on a few datasets, we comprehensively assess UNITE on a broad range of face and synthetic datasets with various T2V/I2V generators, outperforming detection in foreground, background, and T2V/I2V manipulations.
+
+# 2. Related Work
+
+Face-centric DeepFake Detection: Methods such as [13, 19, 40] focus on modeling spatial inconsistencies for DeepFake detection. Concas et al. [13] track specific facial features (eyes, nose, mouth) in a high-frequency domain, as fake videos exhibit distinct behaviors in this space compared to real ones. However, due to their dataset-centric architectures, it is not a generalizable approach. PUDD [40] learns person-specific prototypes by modeling patterns from real videos and assessing how closely inference videos match these prototypes, making it less effective for detecting in-the-wild DeepFakes. LAA-Net [39] attempts to improve generalizability through a multi-task attention module that detects artifacts via heatmap and self-consistency regression, but it works by analyzing every still frame of a video, thus limiting its scalability. The authors in [36] tackled the challenging problem of detecting and localizing expression swaps, where the number of manipulated pixels is even fewer compared to identity swaps. Although the authors obtained good localization performance, their reliance on a CNN-based architecture limits their ability to capture temporal inconsistencies effectively.
+
+DPNet [50] and ID-Reveal [16] are identity-aware temporal artifact modeling methods: DPNet [50] focuses on interpretable, prototype-based detection of unnatural movements through dynamic feature representations, while IDReveal [16] leverages metric learning on real data to detect biometric inconsistencies. However these methods assume access to authentic reference videos, making them unfit for in-the-wild DeepFake detection. Shifting from identity-specific methods, TALL [57] constructs composite "thumbnail" images from four consecutive frames taken at random timestamps for temporal analysis. TI2Net [32] captures temporal anomalies by subtracting consecutive frame features but relies on the assumption that FaceSwap manipulations show pronounced inter-frame discrepancies.
+
+Choi et al. [11] approach DeepFake detection by targeting suppressed variance in style-based latent temporal features via a StyleGRU module. However, this method is limited to cropped face regions, making it ineffective for T2V/I2V or background-manipulated videos.
+
+Synthetic Video Detection: While there has been considerable effort in synthetic image detection [14, 15, 55], synthetic video detection is a scarcely explored area. DeMamba [9], a recently proposed synthetic video detector, analyzes small zones of video frames to capture how pixels change spatiotemporally. Its continuous scanning approach
+
+![](images/293854bf54d05f4417715994a0c7ca60da622532c26dfd731035498dd8ac84a2.jpg)  
+Figure 2. UNITE architecture overview: We extract domain-agnostic features  $(\xi)$  using the SigLIP-So400m foundation model [4] to mitigate domain gaps between DeepFake datasets (Sec. 3.2). These embeddings, combined with positional encodings, are input to a transformer with multi-head attention and MLP layers (Sec. 3.3), culminating in a classifier for final predictions. AD-loss (Sec. 3.4) encourages the model attention to span diverse spatial regions.
+
+better tracks subtle changes and patterns, making it easier to spot manipulated or fake content. The authors also propose a million scale T2V/I2V dataset (which we call the DeMamba dataset hereforth) to evaluate their performance. However, the method was not evaluated on typical DeepFake datasets [30, 43, 53] and thus is not fit for human-centric DeepFake detection. In contrast our UNITE model is built to detect partially manipulated as well as fully AI-generated videos. A key reason for the scarcity of synthetic video detection methods in the literature is the lack of standardized datasets for training and evaluation.
+
+# 3. Proposed Method
+
+In this section we first describe our problem setup (Sec. 3.1), then we describe the foundation model-based video encoding (Sec. 3.2), our trainable transformer architecture (Sec. 3.3) and finally the novel loss function for training the UNITE model (Sec. 3.4). The overview of the UNITE architecture is shown in Fig. 2.
+
+# 3.1. Problem Setup
+
+For each video in the dataset, we implement a frame sampling strategy where every other frame is extracted. From the resulting collection of frames, we construct video segment samples consisting of  $n_f = 64$  consecutive frames, which are defined as a single data sample, denoted as "v". This selection aligns with the context window established for our video transformer model, ensuring that each sample retains the temporal coherence necessary for effective train-
+
+ing. In instances where the video duration results in fewer than  $n_f$  frames, we apply padding (ablation in supplementary) to ensure that each sample remains uniform in size.
+
+Given that the original video-level classification labels are known and fake videos contain manipulations throughout their entirety (i.e., no videos have manipulations limited to specific frames), each  $n_f$ -frame segment generated from a given video is assigned the same label (denoted by  $l \in 0, \dots, n_c$ ) as the source video, with number of possible classes as  $n_c$ . The aggregation of these samples across all videos forms our final dataset,  $v_i, l_{i=1}^N \in \mathcal{V}$ , where  $N$  is the total number of samples obtained. This approach serves as a data augmentation technique, increasing the number of training samples without additional data collection.
+
+# 3.2. Domain-Agnostic Feature Extraction
+
+Our objective is to leverage the trained UNITE model for detecting in-the-wild DeepFakes, necessitating careful consideration of the domain gap [7, 34] between the train and test datasets. To effectively mitigate the impact of domain discrepancies, we utilize the SigLIP foundation model [59] to extract domain-agnostic features. Specifically, we employ the "shape-optimized" ViT [4] image encoder, whose architecture is rooted in [18]. The SigLIP-So400m [4] model was contrastively pretrained on 3B diverse examples using a sigmoid loss, enabling it to capture robust, generalizable features while maintaining a compact size (400M parameters). This extensive pretraining makes SigLIP particularly adept at extracting domain-invariant representations, crucial for handling the significant variability encountered in DeepFake datasets and real-world scenarios.
+
+For every frame  $f_{j}$  (resized to  $\mathbb{R}^{384\times 384\times 3}$ ), of a video sample  $v_{i}\in \mathcal{V}$ , we obtain an image encoding from the frozen SigLIP-So400m [4] model as  $e_j = SigLIP(f_j)\in \mathbb{R}^{t_s\times d_s}$ , where  $j\in \{1,2,\dots ,n_f\}$ , since there are  $n_f$  frames per video sample (as explained in Sec. 3.1),  $t_s = 729$  represents the token length and  $d_{s} = 1152$  represents the feature dimension. We concatenate these frame-level features for the video sample  $v_{i}$ , while maintaining the order of the frames, to obtain a video segment encoding denoted by  $\xi_{i}\in \mathbb{R}^{n_{f}\times t_{s}\times d_{s}}$ , which are used as input to our trainable transformer architecture. This encoded video dataset can be denoted as  $\{\xi_i,l_i\}_{i = 1}^N\in \mathcal{V}_\xi$ .
+
+# 3.3. Learnable Transformer Architecture
+
+The UNITE architecture is a multi-head self-attention (MHSA) based transformer model designed specifically for video classification tasks. It leverages the strengths of Transformer networks to process sequences of video frame embeddings effectively, allowing for robust classification performance. The UNITE architecture allows for adjustable depth, signifying the number of encoder blocks stacked within the model. A deeper architecture can capture more
+
+intricate features and dependencies at the cost of increased computational complexity. For our architecture the depth is set to 4. Ablation experiments with different transformer depths are provided in Sec. 4.4.
+
+# 3.3.1 Encoder Block
+
+Each encoder block consists of the following components:
+
+Multi-Head Self-Attention Layer: This layer allows the model to focus on different parts of the input sequence simultaneously. With 12 attention heads in our architecture, it captures diverse interactions and dependencies among frames. Each head computes attention scores independently, learning multiple representations of the input data. This enhances the model's ability to detect subtle temporal variations, which is crucial for video classification. The use of scaled dot-product attention further helps mitigate large gradients and stabilizes training.
+
+Layer Normalization and Residual Connections: To aid gradient flow during training, each sub-layer incorporates residual connections that add the input back to the output, mitigating vanishing gradient issues in deep networks. Layer normalization follows to stabilize activations and enhance convergence speed. Additionally, dropout is applied post-attention to reduce overfitting by randomly deactivating a portion of neurons during training.
+
+Feed-Forward Network (MLP): The second sub-layer is a point-wise feed-forward network with two dense layers separated by a GELU activation [23]. This MLP enables nonlinear feature transformations, enhancing the model's expressiveness. By projecting attention outputs into a higher-dimensional space before reducing them, it captures complex interactions beyond linear transformations.
+
+Attention maps are derived from the outputs of the encoder blocks to be used for computing the Attention-Diversity loss (Sec. 3.4) and to illustrate how the model prioritizes different spatial regions in the frames during the classification process (Fig. 3). This interpretability is valuable for understanding the model's decision-making, particularly in complex video classification scenarios.
+
+# 3.3.2 Positional Encoding
+
+We utilize a sine-cosine positional encoding scheme following the original transformer [51] to provide a unique position identifier for each input token. For every frame  $f_{j}$  in the video sample with encoded feature dimension  $d_{s}$  (as per Sec. 3.2), we compute the positional encoding for odd and even indexed feature dimensions as,
+
+$$
+P E _ {(j, 2 i + 1)} = \cos \left(\frac {j}{1 0 0 0 0 ^ {\frac {2 i}{d _ {s}}}}\right), \quad P E _ {(j, 2 i)} = \sin \left(\frac {j}{1 0 0 0 0 ^ {\frac {2 i}{d _ {s}}}}\right), \tag {1}
+$$
+
+where,  $i$  denotes the feature index in the encoded feature dimension  $d_{s}$ . We embed these encodings directly into the input tokens of our video transformer to provide positional context, for better temporal modeling during the attention computation with minimal computational overhead.
+
+# 3.4. Attention-Diversity Loss
+
+Training a multi-head attention transformer with only cross-entropy loss makes the attention maps focus only on the face regions of the frame (as evident in Fig. 3). However, we aim to detect fake videos where the manipulation might not be in the face at all, and instead be in the background (like video inpainting with the AVID model [61]). To ensure that the attention heads focus on diverse spatial regions of the video frames, we devise the "Attention-Diversity" (AD) loss.
+
+AD-loss is designed to minimize overlap among attention maps (obtained from the first encoder block of the trainable transformer architecture in 3.3.1) while maintaining consistency across different input video samples. The attention outputs from the video transformer model  $\mathcal{A} \in \mathbb{R}^{n_h \times t_s \times d_s}$  are used to pool the input SigLIP-So400m features  $\xi \in \mathbb{R}^{n_f \times t_s \times d_s}$ , resulting in a pooled feature tensor  $\mathcal{P} \in \mathbb{R}^{n_h \times n_f}$  according to,
+
+$$
+\mathcal {P} = \sum_ {j = 1} ^ {t _ {s}} \sum_ {k = 1} ^ {d _ {s}} \mathcal {A} _ {h, j, k} \cdot \xi_ {f, j, k}, \tag {2}
+$$
+
+where  $f$  indexes the frames  $(n_f)$ ,  $h$  indexes the number of attention heads  $(n_h)$ , and  $j, k$  are spatial positions.
+
+We compute "feature centers" which are points in the feature space that serve as anchors for the learned representations of different classes. These centers represent the average features of samples from a particular class, allowing the model to capture essential characteristics of that class. In the context of the AD-loss, feature centers  $\mathcal{C} \in \mathbb{R}^{n_h \times n_f}$  are dynamically updated in each training iteration  $\tau$  based on the pooled feature vectors  $\mathcal{P}$ , following,
+
+$$
+\mathcal {C} ^ {\tau} = \mathcal {C} ^ {\tau - 1} - \eta \left(\mathcal {C} ^ {\tau - 1} - \frac {1}{B} \sum_ {b = 1} ^ {B} \mathcal {P} _ {b}\right), \tag {3}
+$$
+
+where  $\eta$  represents the learning rate (set to 0.05 in our experiments) for feature center updates and  $B$  is the batch dimension. The feature centers are initialized with zeros  $(\mathcal{C}^0 = [0]_{n_h\times n_f})$  at the beginning of the model training.
+
+AD-loss consists of two distinct components: within-class loss and between-class loss inspired by the Fisher Discriminant Analysis for deep networks [22]. The within-class term calculates the distance between the pooled feature vectors  $\mathcal{P}$  and their corresponding feature centers  $\mathcal{C}$ , promoting closeness among similar samples. The between-class term measures the distance between feature centers of different classes, encouraging them to be spaced apart in the feature space, encouraging separability.
+
+The within-class loss term is calculated as
+
+$$
+\mathcal {L} _ {\text {w i t h i n}} = \max  \left(\| \mathcal {P} - \mathcal {C} \| _ {2} - \delta_ {\text {w i t h i n}}, 0\right), \tag {4}
+$$
+
+where  $\delta_{\mathrm{within}} \in \mathbb{R}^{n_c}$  is a hyperparameter controlling the allowable distance between feature vectors and their respective feature centers for inputs of the same class, encouraging tighter clustering and  $\| \cdot \|_2$  represents  $L_2$  normalization.
+
+The between-class loss is calculated as
+
+$$
+\mathcal {L} _ {\text {b e t w e e n}} = \sum_ {\substack {k \neq l \\ (k, l) \in \left(n _ {h}, n _ {h}\right)}} \max  \left(\delta_ {\text {b e t w e e n}} - \| \mathcal {C} _ {k} - \mathcal {C} _ {l} \| _ {2}, 0\right), \tag{5}
+$$
+
+where  $\delta_{\mathrm{between}}$  is a predefined hyperparameter that ensures a minimum distance between feature centers of different classes. Thus the AD-loss is a simple addition of these two components  $\mathcal{L}_{AD} = \mathcal{L}_{\mathrm{within}} + \mathcal{L}_{\mathrm{between}}$ , making the final objective function for training the UNITE model,
+
+$$
+\mathcal {L} _ {\text {U N I T E}} = \lambda_ {1} \cdot \mathcal {L} _ {C E} + \lambda_ {2} \cdot \mathcal {L} _ {A D}, \tag {6}
+$$
+
+where  $\mathcal{L}_{CE}$  is the traditional cross-entropy loss, and  $\lambda_{1}$  and  $\lambda_{2}$  are loss-weighting hyperparameters.
+
+# 4. Experiments
+
+# 4.1. Datasets
+
+Most DeepFake datasets focus on face manipulations, such as face-swaps or lip-syncing, and lack fully AI-generated video content. The DeMamba dataset [9] features videos from T2V/I2V models, including powerful diffusion models like SORA [5]. It is not a human-centric dataset and primarily contains videos from natural scenes. We use its validation split to evaluate UNITE.
+
+To train UNITE, we employ the FaceForensics++ (FF++) [43] dataset (c23 i.e., medium-quality) and the SAIL-VOS-3D dataset [24], which includes synthetic (although not AI-generated) GTA-V game videos featuring human subjects suitable for DeepFake detection. To evaluate the UNITE model, we use:
+
+- Face manipulated data:  $\mathrm{FF}++$  [43] (in-domain evaluation), CelebDF [30], DeeperForensics [25], Deepfake-TIMIT [27], HifiFace [53], UADFV [58].  
+- Background manipulated data: Sample videos from the AVID [61] model provided publicly by the authors.  
+- Fully synthetic data: GTA-V [24] (in-domain evaluation), DeMamba [9].  
+- In-the-wild DeepFakes: Publicly available videos from the New York Times DeepFake quiz [48].
+
+# 4.2. Results and Discussion
+
+Evaluation Metrics: Although most DeepFake detection methods only report accuracy, we evaluate UNITE using additional metrics: (1) AUC (Area Under the Precision-Recall
+
+Curve); (2) Precision@0.5 and (3) Recall@0.5 at a 0.5 confidence threshold; (4) Precs@Rec=0.8 (precision at 0.8 recall); (5) Rec@Precs=0.8 (recall at 0.8 precision).
+
+Training Details: UNITE is trained using an AdamW optimizer [33] with an initial learning rate of 0.0001 and a decay rate of 0.5 every 1000 steps. For AD-loss,  $\delta_{\mathrm{between}} = 0.5$  and  $\delta_{\mathrm{within}}$  is set to  $[0.01, -2]$  for binary and  $[0.01, -2, 1]$  for fine-grained classification in Sec. 4.3 (see supplementary for sensitivity). Loss weights are  $\lambda_1 = \lambda_2 = 0.5$ . Training uses a batch size of 32 for 25 epochs on 8 TPUv3 chips, with the framework implemented in TensorFlow.
+
+Quantitative Results: The results obtained by the UNITE model when trained on  $\mathrm{FF}++$  [43] alone, and when trained with both  $\mathrm{FF}++$  [43] and GTA-V [24] data and evaluated on the various datasets mentioned in Sec. 4.1 are shown in Table 1. On the AVID [61] and DeMamba datasets [9], the model performs poorly when trained with only face manipulated data, but the performance enhances several-fold when the GTA-V [24] synthetic data, even though it is not AI-generated, is used in training.
+
+Interestingly, when evaluating on face-manipulated datasets such as CelebDF [30] and DeeperForensics [25], we observe a performance boost, particularly for CelebDF [30], when GTA-V [24] data is included in the training set. Intuitively, one would expect that training solely on  $\mathrm{FF}++$  [43] would yield similar results since the inference set consists of similar face-manipulated data. This deviation from the expected behavior comes from the contribution of the AD-Loss, as the ablation study (Fig. 4) using only CE-loss compared to both CE and AD-losses reveals that the performance is enhanced with GTA-V [24] data in training only in the latter (detailed discussion in Sec. 4.4).
+
+To evaluate UNITE on in-the-wild DeepFakes, we attempted the recent New York Times quiz [48] (denoted by NYTimes) which provides 10 videos (4 real and 6 fake) for testing DeepFake detectors. UNITE got 8 of those 10 videos correct, even though some of the face-swap videos were indistinguishable from real videos, even by humans. On all the AI-generated videos, UNITE made the correct prediction (when trained on  $\mathrm{FF} + +$  [43] and GTA-V [24]).
+
+Attention Heatmaps: To analyze the impact of CE and AD-losses on attention distribution, we extracted attention features from the first encoder of the UNITE model, comparing models trained with CE loss alone versus that trained with both CE and AD-losses. The resulting heatmaps, shown in Fig. 3, reveal that the model trained with only CE loss tends to concentrate primarily on the face region. In contrast, the model trained with the combined  $\mathrm{CE + AD}$  loss demonstrates a broader attention span across the frame, as evidenced by a lighter, more distributed bluish tone in the heatmaps, indicating increased spatial diversity in the model's focus. This is why on the example shown in the first row (face-manipulated video) both models give the correct
+
+Table 1. Results from the UNITE model trained with (1) FF++ [43] only and (2) FF++ [43] combined with GTA-V [24]. All other results reflect cross-dataset evaluations except for FF++ and GTA-V (when trained). Performance gains are highlighted in green.  
+
+<table><tr><td colspan="2">Train</td><td colspan="7">Test</td></tr><tr><td>FF++</td><td>GTA-V</td><td>Dataset</td><td>Accuracy</td><td>AUC</td><td>Precision@0.5</td><td>Recall@0.5</td><td>Precs@Rec=0.8</td><td>Rec@Precs=0.8</td></tr><tr><td colspan="9">Face Manipulated Data</td></tr><tr><td>✓</td><td></td><td>FF++</td><td>99.53%</td><td>99.77%</td><td>99.94%</td><td>99.49%</td><td>99.94%</td><td>99.94%</td></tr><tr><td>✓</td><td></td><td>CelebDF</td><td>72.61%</td><td>94.05%</td><td>96.45%</td><td>61.22%</td><td>80.45%</td><td>61.22%</td></tr><tr><td>✓</td><td></td><td>DeeperForensics</td><td>91.35%</td><td>100.00%</td><td>100.00%</td><td>91.35%</td><td>100.00%</td><td>91.35%</td></tr><tr><td>✓</td><td></td><td>DeepfakeTIMIT</td><td>86.90%</td><td>86.46%</td><td>83.61%</td><td>83.97%</td><td>88.90%</td><td>81.33%</td></tr><tr><td>✓</td><td></td><td>HiFiFace</td><td>63.63%</td><td>62.47%</td><td>67.12%</td><td>63.63%</td><td>59.30%</td><td>63.63%</td></tr><tr><td>✓</td><td></td><td>UADFV</td><td>94.12%</td><td>94.38%</td><td>95.68%</td><td>97.11%</td><td>93.79%</td><td>94.38%</td></tr><tr><td>✓</td><td>✓</td><td>FF++</td><td>99.96%(+0.43)</td><td>99.89%(+0.12)</td><td>100.00%(+0.06)</td><td>99.84%(+0.35)</td><td>100.00%(+0.06)</td><td>99.96%(+0.02)</td></tr><tr><td>✓</td><td>✓</td><td>CelebDF</td><td>95.11%(+22.50)</td><td>94.36%(+0.31)</td><td>96.82%(+0.37)</td><td>68.75%(+7.53)</td><td>96.53%(+16.08)</td><td>68.75%(+7.53)</td></tr><tr><td>✓</td><td>✓</td><td>DeeperForensics</td><td>99.62%(+8.27)</td><td>100.00%(+0.00)</td><td>100.00%(+0.00)</td><td>99.62%(+8.27)</td><td>100.00%(+0.00)</td><td>99.63%(+8.28)</td></tr><tr><td>✓</td><td>✓</td><td>DeepfakeTIMIT</td><td>91.90%(+5.00)</td><td>91.33%(+4.87)</td><td>90.45%(+6.84)</td><td>88.39%(+4.42)</td><td>100.00%(+11.10)</td><td>91.95%(+10.62)</td></tr><tr><td>✓</td><td>✓</td><td>HiFiFace</td><td>75.62%(+11.99)</td><td>81.24%(+18.77)</td><td>79.55%(+12.43)</td><td>71.71%(+8.08)</td><td>75.62%(+16.32)</td><td>72.47%(+8.84)</td></tr><tr><td>✓</td><td>✓</td><td>UADFV</td><td>97.01%(+2.89)</td><td>94.95%(+0.57)</td><td>96.89%(+1.21)</td><td>100.00%(+2.89)</td><td>94.12%(+0.33)</td><td>100.00%(+5.62)</td></tr><tr><td colspan="9">Background Manipulated Data</td></tr><tr><td>✓</td><td></td><td>AVID</td><td>41.67%</td><td>33.33%</td><td>33.33%</td><td>41.67%</td><td>41.67%</td><td>33.33%</td></tr><tr><td>✓</td><td>✓</td><td>AVID</td><td>100.00%(+58.33)</td><td>100.00%(+66.67)</td><td>100.00%(+66.67)</td><td>100.00%(+58.33)</td><td>100.00%(+58.33)</td><td>100.00%(+66.67)</td></tr><tr><td colspan="9">Fully Synthetic Data</td></tr><tr><td>✓</td><td></td><td>GTA-V</td><td>60.16%</td><td>61.52%</td><td>60.16%</td><td>58.73%</td><td>63.29%</td><td>58.73%</td></tr><tr><td>✓</td><td></td><td>DeMamba</td><td>61.47%</td><td>57.38%</td><td>67.73%</td><td>33.01%</td><td>62.15%</td><td>54.16%</td></tr><tr><td>✓</td><td>✓</td><td>GTA-V</td><td>100.00%(+39.84)</td><td>100.00%(+38.48)</td><td>100.00%(+39.84)</td><td>100.00%(+41.27)</td><td>100.00%(+36.71)</td><td>100.00%(+41.27)</td></tr><tr><td>✓</td><td>✓</td><td>DeMamba</td><td>87.12%(+25.65)</td><td>93.75%(+36.67)</td><td>92.76%(+25.03)</td><td>89.60%(+56.59)</td><td>89.81%(+27.66)</td><td>92.12%(+37.96)</td></tr><tr><td colspan="9">In-the-wild DeepFakes</td></tr><tr><td>✓</td><td></td><td>NYTimes [48]</td><td>50.00%</td><td>53.74%</td><td>50.00%</td><td>25.00%</td><td>0.00%</td><td>0.00%</td></tr><tr><td>✓</td><td>✓</td><td>NYTimes [48]</td><td>80.00%(+30.00)</td><td>97.42%(+43.68)</td><td>83.33%(+33.33)</td><td>83.33%(+58.33)</td><td>60.00%(+60.00)</td><td>100.00%(+100.00)</td></tr></table>
+
+Table 2. SOTA Comparison on Face-Manipulated Data: We compared the performance of UNITE with recent DeepFake detectors, in terms of detection accuracy on various face manipulated datasets. UNITE outperforms the existing methods. Bold shows the current best results and the previous best and second-best results are highlighted in red and blue respectively.  
+
+<table><tr><td>Method</td><td>FF++</td><td>CelebDF</td><td>DeeperForensics</td><td>UADFV</td></tr><tr><td>TALL [57]</td><td>98.65%</td><td>90.79%</td><td>99.62%</td><td>-</td></tr><tr><td>ISTVT [62]</td><td>99.00%</td><td>84.10%</td><td>98.60%</td><td>-</td></tr><tr><td>Concas et al. [13]</td><td>99.49%</td><td>-</td><td>-</td><td>-</td></tr><tr><td>PUDD [40]</td><td>-</td><td>95.10%</td><td>-</td><td>-</td></tr><tr><td>TI2Net [32]</td><td>99.95%</td><td>68.22%</td><td>76.08%</td><td>-</td></tr><tr><td>LRNet [46]</td><td>99.89%</td><td>53.20%</td><td>56.77%</td><td>-</td></tr><tr><td>Choi et al. [11]</td><td>-</td><td>89.00%</td><td>99.00%</td><td>-</td></tr><tr><td>Lin et al. [31]</td><td>98.28%</td><td>74.42%</td><td>-</td><td>-</td></tr><tr><td>Guo et al. [21]</td><td>99.24%</td><td>84.97%</td><td>-</td><td>-</td></tr><tr><td>Li et al. (Res-152) [29]</td><td>-</td><td>-</td><td>-</td><td>93.80%</td></tr><tr><td>HeadPose [58]</td><td>-</td><td>-</td><td>-</td><td>89.00%</td></tr><tr><td>CViT [56]</td><td>93.00%</td><td>-</td><td>-</td><td>93.75%</td></tr><tr><td>FakeCatcher [12]</td><td>94.65%</td><td>91.50%</td><td>-</td><td>-</td></tr><tr><td>MesoNet [3]</td><td>-</td><td>-</td><td>-</td><td>82.40%</td></tr><tr><td>UNITE (Ours)</td><td>99.96%</td><td>95.11%</td><td>99.62%</td><td>97.01%</td></tr></table>
+
+prediction ("fake") with high confidence, but on the row-2 example (video generated from the Sora [6] T2V model) the model trained with only CE loss predicted "real" with a confidence score of  $99.25\%$  (which is wrong), but the model trained with  $\mathrm{CE + AD}$  losses predicted "fake" with a confidence of  $100.00\%$ , which is correct.
+
+Comparison on Face Manipulated Data: UNITE is compared with recent state-of-the-art (SOTA) DeepFake detection models in Table 2. The methods we compared against were specifically designed for detecting face manipulations and often crop faces from the videos, during both training and inference of the models. However, UNITE, which is designed to detect a diverse range of fake videos, still outperforms these detectors.
+
+Comparison on Synthetic Data: We evaluated UNITE against state-of-the-art synthetic video detectors on the DeMamba dataset [9] (validation split), with results in Table 3. While existing detectors were trained on the DeMamba train split and validated on the DeMamba validation split,
+
+![](images/c22e4507b864d24b58be75ce23f856e165498a1d17c85f7886055912690090c3.jpg)  
+(a) Original Frame
+
+![](images/9a7f55b96974cf7b358b430d9ef5d44c715dd9f04ae58d9d55339da946fc69ed.jpg)  
+(b) CE Loss only
+
+![](images/a73dc7545cdac92a63c283887cef990df632ac6489198467c2b382e0b2d6c15f.jpg)  
+(c)  $\mathrm{CE + AD}$  Loss
+
+![](images/66b04e8a8f21be849327484594cfd6b4ecf64172824ab28252ce8e38960c5b84.jpg)  
+(d) Original Frame
+
+![](images/c6f70e908b9eb1083ccd7c19c4d8aafcce441fe1ac95d4a3093f9532bc7b1cf2.jpg)  
+(e) CE Loss only
+
+![](images/f78bfdc6431dce203fc1738fdbc24287c4c17126ed02ee3507b3c365f47ad81c.jpg)  
+(f)  $\mathrm{CE + AD}$  Loss  
+Figure 3. Comparison of attention heatmaps when UNITE is trained with only CE loss (second column) versus CE and AD-losses (third column). The CE-loss heatmap predominantly focuses on the face region, whereas the  $\mathrm{CE + AD}$  loss heatmap demonstrates more distributed attention across the entire frame, as indicated by a broader bluish tone in the heatmaps. Both samples are "fake" videos- (a) is taken from Celeb-DF [30] and (d) is taken from DeMamba [9] and generated by OpenAI's Sora [6] (Best viewed as GIFs provided in the supplementary material). Yellow to blue shades show increasing attention in the plot.
+
+UNITE was trained on FF++ [43] and GTA-V [24]. Despite not being trained on DeMamba, UNITE's average performance (including real videos) surpassed current SOTA detectors, with competitive generator-wise results.
+
+Thus, UNITE reliably detects a diverse range of fake videos (even in cross-domain settings), including T2V/I2V videos and background manipulations, while consistently maintaining state-of-the-art performance on traditional face-manipulated data. UNITE eliminates the need to have separate DeepFake and T2V/I2V video detector models by by handling both tasks within a single trained model.
+
+# 4.3. Finegrained Evaluation
+
+In fake media detection, flagging videos partially or fully manipulated by AI also necessitates fine-grained classification into three classes: (1) real videos, (2) partially manipulated videos, and (3) fully synthetic videos. While binary
+
+Table 3. SOTA Comparison on Synthetic Data: On the DeMamba dataset (validation split), we compare the performance of UNITE, which was NOT trained on DeMamba train split, against state-of-the-art detectors which were trained on DeMamba train split (results taken from Chen et al. [9]). We report the results ( $P = \text{Precision}$  and  $R = \text{Recall}$ ) on the individual T2V/I2V generators and the average performance across the entire validation set ( $Avg$ , which also includes real videos). Although the direct comparison is unfair against UNITE which was trained with FF++ [43] and GTA-V [24], our method still outperforms these synthetic video detectors. Bold shows the current best results and the previous best and second-best results are highlighted in red and blue respectively.  
+
+<table><tr><td>Method</td><td>Metrics</td><td>Sora [6]</td><td>Morph Studio [2]</td><td>Runway ML (Gen2) [42]</td><td>HotShot [38]</td><td>Lavie [54]</td><td>Show-1 [60]</td><td>Moon Valley [37]</td><td>Crafter [8]</td><td>Model Scope [52]</td><td>Wild Scrape [9]</td><td>Avg</td></tr><tr><td rowspan="2">TALL [57]</td><td>P</td><td>71.15%</td><td>96.89%</td><td>98.51%</td><td>79.38%</td><td>84.59%</td><td>79.38%</td><td>98.79%</td><td>99.02%</td><td>92.70%</td><td>76.47%</td><td>87.91%</td></tr><tr><td>R</td><td>91.07%</td><td>98.28%</td><td>97.83%</td><td>83.00%</td><td>76.57%</td><td>79.57%</td><td>99.52%</td><td>98.93%</td><td>94.14%</td><td>66.31%</td><td>88.52%</td></tr><tr><td rowspan="2">F3Net [41]</td><td>P</td><td>68.27%</td><td>99.89%</td><td>99.67%</td><td>89.35%</td><td>57.00%</td><td>36.57%</td><td>99.52%</td><td>99.71%</td><td>93.80%</td><td>88.41%</td><td>88.73%</td></tr><tr><td>R</td><td>83.93%</td><td>99.71%</td><td>98.62%</td><td>77.57%</td><td>85.24%</td><td>63.17%</td><td>99.58%</td><td>99.89%</td><td>89.43%</td><td>76.78%</td><td>81.88%</td></tr><tr><td rowspan="2">NPR [47]</td><td>P</td><td>91.07%</td><td>99.57%</td><td>99.49%</td><td>24.29%</td><td>89.64%</td><td>57.71%</td><td>97.12%</td><td>99.86%</td><td>94.29%</td><td>87.80%</td><td>82.45%</td></tr><tr><td>R</td><td>91.07%</td><td>99.57%</td><td>99.49%</td><td>24.29%</td><td>89.64%</td><td>57.71%</td><td>97.12%</td><td>99.86%</td><td>94.29%</td><td>87.80%</td><td>84.08%</td></tr><tr><td rowspan="2">STIL [20]</td><td>P</td><td>57.21%</td><td>99.08%</td><td>99.32%</td><td>86.19%</td><td>82.24%</td><td>70.43%</td><td>99.25%</td><td>98.96%</td><td>97.18%</td><td>81.32%</td><td>87.12%</td></tr><tr><td>R</td><td>67.86%</td><td>96.00%</td><td>98.41%</td><td>96.14%</td><td>77.14%</td><td>80.43%</td><td>97.44%</td><td>96.93%</td><td>96.29%</td><td>68.36%</td><td>82.22%</td></tr><tr><td rowspan="2">MINTIME-CLIP-B [9]</td><td>P</td><td>83.21%</td><td>99.99%</td><td>99.67%</td><td>50.84%</td><td>99.20%</td><td>99.27%</td><td>99.76%</td><td>99.99%</td><td>91.83%</td><td>91.77%</td><td>91.55%</td></tr><tr><td>R</td><td>89.29%</td><td>100.00%</td><td>98.99%</td><td>26.43%</td><td>96.79%</td><td>98.14%</td><td>99.84%</td><td>100.00%</td><td>84.29%</td><td>82.38%</td><td>87.62%</td></tr><tr><td rowspan="2">FTCN-CLIP-B [9]</td><td>P</td><td>91.79%</td><td>99.99%</td><td>99.79%</td><td>45.94%</td><td>99.76%</td><td>97.80%</td><td>99.99%</td><td>99.99%</td><td>94.69%</td><td>92.32%</td><td>92.21%</td></tr><tr><td>R</td><td>87.50%</td><td>100.00%</td><td>98.91%</td><td>17.71%</td><td>97.71%</td><td>91.86%</td><td>100.00%</td><td>100.00%</td><td>85.29%</td><td>82.83%</td><td>86.18%</td></tr><tr><td rowspan="2">CLIP-B-PT [9]</td><td>P</td><td>67.80%</td><td>43.56%</td><td>70.88%</td><td>29.97%</td><td>52.97%</td><td>35.36%</td><td>55.52%</td><td>66.03%</td><td>44.23%</td><td>42.99%</td><td>44.83%</td></tr><tr><td>R</td><td>85.71%</td><td>82.43%</td><td>90.36%</td><td>71.00%</td><td>79.29%</td><td>75.43%</td><td>89.62%</td><td>86.29%</td><td>82.14%</td><td>75.16%</td><td>81.74%</td></tr><tr><td rowspan="2">DeMamba-CLIP-PT [9]</td><td>P</td><td>25.87%</td><td>95.14%</td><td>96.23%</td><td>73.43%</td><td>83.31%</td><td>75.49%</td><td>90.17%</td><td>95.06%</td><td>95.05%</td><td>69.95%</td><td>79.97%</td></tr><tr><td>R</td><td>58.93%</td><td>96.43%</td><td>93.12%</td><td>68.00%</td><td>69.36%</td><td>69.00%</td><td>89.14%</td><td>91.86%</td><td>96.14%</td><td>56.59%</td><td>78.86%</td></tr><tr><td rowspan="2">XCLIP-B-PT [9]</td><td>P</td><td>16.39%</td><td>72.16%</td><td>87.77%</td><td>39.86%</td><td>65.57%</td><td>54.26%</td><td>75.23%</td><td>84.80%</td><td>61.60%</td><td>55.28%</td><td>61.29%</td></tr><tr><td>R</td><td>81.34%</td><td>82.15%</td><td>83.35%</td><td>80.98%</td><td>81.82%</td><td>81.55%</td><td>82.14%</td><td>82.98%</td><td>81.93%</td><td>81.10%</td><td>81.93%</td></tr><tr><td rowspan="2">DeMamba-XCLIP-PT [9]</td><td>P</td><td>18.26%</td><td>93.50%</td><td>94.72%</td><td>69.94%</td><td>78.08%</td><td>71.50%</td><td>83.95%</td><td>92.23%</td><td>93.54%</td><td>68.10%</td><td>76.38%</td></tr><tr><td>R</td><td>66.07%</td><td>95.86%</td><td>94.64%</td><td>77.86%</td><td>75.36%</td><td>80.29%</td><td>90.89%</td><td>92.50%</td><td>96.00%</td><td>66.41%</td><td>83.59%</td></tr><tr><td rowspan="2">XCLIP-B-FT [9]</td><td>P</td><td>64.42%</td><td>99.73%</td><td>96.78%</td><td>70.98%</td><td>90.35%</td><td>77.28%</td><td>97.34%</td><td>99.84%</td><td>82.01%</td><td>88.97%</td><td>86.77%</td></tr><tr><td>R</td><td>82.14%</td><td>99.57%</td><td>93.62%</td><td>61.29%</td><td>79.36%</td><td>69.71%</td><td>97.92%</td><td>99.79%</td><td>77.14%</td><td>83.59%</td><td>84.41%</td></tr><tr><td rowspan="2">UNITE (Ours)</td><td>P</td><td>88.57%</td><td>100.00%</td><td>100.00%</td><td>90.16%</td><td>89.91%</td><td>98.34%</td><td>99.52%</td><td>100.00%</td><td>98.96%</td><td>92.56%</td><td>92.76%</td></tr><tr><td>R</td><td>92.11%</td><td>100.00%</td><td>94.62%</td><td>96.93%</td><td>98.12%</td><td>99.86%</td><td>98.69%</td><td>100.00%</td><td>96.29%</td><td>89.89%</td><td>89.60%</td></tr></table>
+
+Table 4. Results obtained by the UNITE model on finegrained classes: detecting whether a video is real, partially manipulated or fully AI-generated. Performance gains are mentioned in green.  
+
+<table><tr><td colspan="2">Train</td><td colspan="2">Test</td></tr><tr><td>FF++</td><td>GTA-V</td><td>Dataset</td><td>Accuracy</td></tr><tr><td>✓</td><td></td><td>FF++</td><td>96.46%</td></tr><tr><td>✓</td><td></td><td>CelebDF</td><td>60.40%</td></tr><tr><td>✓</td><td></td><td>DeeperForensics</td><td>71.04%</td></tr><tr><td>✓</td><td></td><td>DeepFakeTIMIT</td><td>80.68%</td></tr><tr><td>✓</td><td></td><td>HifiFace</td><td>43.27%</td></tr><tr><td>✓</td><td></td><td>UADFV</td><td>90.37%</td></tr><tr><td>✓</td><td></td><td>AVID</td><td>50.00%</td></tr><tr><td>✓</td><td></td><td>GTA-V</td><td>0.00%</td></tr><tr><td>✓</td><td></td><td>DeMamba</td><td>38.29%</td></tr><tr><td>✓</td><td>✓</td><td>FF++</td><td>97.70%(+1.24)</td></tr><tr><td>✓</td><td>✓</td><td>CelebDF</td><td>70.17%(+9.77)</td></tr><tr><td>✓</td><td>✓</td><td>DeeperForensics</td><td>80.59%(+9.55)</td></tr><tr><td>✓</td><td>✓</td><td>DeepFakeTIMIT</td><td>81.04%(+0.36)</td></tr><tr><td>✓</td><td>✓</td><td>HifiFace</td><td>64.20%(+20.93)</td></tr><tr><td>✓</td><td>✓</td><td>UADFV</td><td>92.61%(+1.24)</td></tr><tr><td>✓</td><td>✓</td><td>AVID</td><td>62.50%(+12.50)</td></tr><tr><td>✓</td><td>✓</td><td>GTA-V</td><td>100.00%(+100.00)</td></tr><tr><td>✓</td><td>✓</td><td>DeMamba</td><td>65.84%(+27.55)</td></tr></table>
+
+real vs. fake classification may suffice to curb misinformation, fine-grained classification offers greater explainability in otherwise black-box models.
+
+Using this defined convention of finegrained classes, all the face manipulated datasets (CelebDF [30], DeeperForensics [25], etc.) and the videos from the AVID [61] model (background manipulations) fall under class-2, and the videos from the GTA-V [24] and the DeMamba [9] datasets fall in class-3. In this setup, only the final classification layer of the UNITE transformer architecture is changed to 3 neurons with softmax activation. The results obtained are presented in Table 4. As expected, when the model is trained only with  $\mathrm{FF}++$  [43], the performance on fully synthetic data is poor (since the model has not seen any synthetic videos). When the GTA-V [24] data is also
+
+![](images/4e0ee692c4c5de2ef0684ec2901942f259d3d5e21d2f9aed9f0a466693c70f68.jpg)  
+(a) UNITE trained on  $\mathrm{FF} + +$  [43] only
+
+![](images/731f76544b4eb7ad0882ec6ef0bec3dce866c8aa404905fbfc7f87320a267252.jpg)  
+(b) UNITE trained on  $\mathrm{FF} + +$  [43] and GTA-V [24]  
+Figure 4. Ablation Results to show the effect of changing the loss functions used to train the UNITE model. The combination of the cross-entropy (CE) and Attention Diversity (AD) losses always performs better, and the contribution from the AD-loss component increases significantly when using fully synthetic data for training.
+
+used in training, the performance is much higher.
+
+# 4.4. Ablation Study
+
+We conducted an analysis to assess the contribution of individual loss functions to the training of the UNITE model. The outcomes of this ablation study are presented in Fig. 4. Specifically, we evaluated the impact of the AD-loss in two scenarios: (1) when UNITE was trained solely on the FF++ dataset [43] (Fig. 4(a)), and (2) when UNITE was trained using a combination of the FF++ [43] and the GTA-V [24] datasets (Fig. 4(b)). The combination of the two losses always performed better than the individual losses.
+
+![](images/e2581f9420b0e9c89632475fbb1f99e00a2b12ff8998a84ceffb81c5cc3e3920.jpg)  
+Figure 5. No. of frames vs. Performance: Performance analysis of UNITE based on the number of frames sampled per video segment. The results illustrate that as the number of frames increases from 1 to 64 (context window), the detection accuracy improves, showcasing UNITE's ability to effectively capture temporal inconsistencies in fake videos.
+
+Effect on Synthetic Data: UNITE performed significantly better on the fully synthetic datasets GTA-V [24] and DeMamba [9] when the model was training includes GTA-V [24] data. However, for the DeMamba dataset (cross-data evaluation), the performance with only CE-loss is similar regardless of the training data. It is the AD-loss component that provides the accuracy boost in Fig 4(b).
+
+Effect on Background-Manipulated Data: In Fig. 4(a), we observe that when UNITE was trained solely on  $\mathrm{FF}++$  [43] using only the CE-loss, its performance on the AVID dataset [61] was  $0.00\%$ . This outcome can be attributed to the fact that the fake data in  $\mathrm{FF}++$  [43] contains exclusively face-manipulated videos, causing the UNITE transformer's attention heads to focus predominantly on the facial regions. As AVID [61] comprises only background-manipulated videos, UNITE failed to detect any of the fakes under such conditions. However, when the AD-loss was introduced, the attention heads of UNITE were encouraged to diversify their spatial focus, significantly improving performance. Incorporating synthetic data, as shown in Fig. 4(b), further enhanced this effect, achieving  $100\%$  accuracy on AVID when using both CE-loss and AD-loss.
+
+Effect on Face-Manipulated Data: For the evaluation on face-manipulated datasets such as CelebDF [30] and UADFV [58], one might expect comparable results between training solely on  $\mathrm{FF} + +$  [43] and incorporating GTA-V [24] due to their perceived irrelevance in evaluating face-centric manipulations. However, as shown in Fig. 4(b), training with fully synthetic data leads to a substantial performance improvement. Notably, in both Fig. 4(a) and (b), using only CE loss yields similar results, but the inclusion of the AD-loss component significantly enhances overall detection accuracy. The diversity in the training data makes the UNITE model learn more discriminative attention map features through the AD-loss component, which in turn helps even in detecting face-manipulated videos.
+
+Frames vs. Performance: For inference with UNITE, we evaluated the impact of temporal context by vary-
+
+![](images/69c7bd286988af9c27ba996e301c20fc5c1e44c982413c17694203378e6c34a8.jpg)  
+Figure 6. Transformer Depth Evaluation: Performance comparison of UNITE with varying the number of encoder blocks (depth). UNITE performs optimally in cross-domain settings when the depth is 4. Greater depths overfit to the training domains  $(\mathrm{FF} + +$  [43] and GTA-V [24]), while a depth of 2 is insufficient to capture the complexity of the data.
+
+ing the number of frames sampled from video segments to  $\{1,8,16,32,64\}$ , with 64 frames corresponding to our model's context window. As shown in Fig. 5, performance improves as more frames are included, demonstrating that UNITE effectively captures temporal discontinuities in fake videos. This highlights the model's ability to leverage temporal cues for enhanced detection accuracy.
+
+Transformer Depth Ablation: We evaluate the impact of varying the depth of the transformer architecture in UNITE by adjusting the number of encoder blocks, with results shown in Fig. 6. We tested depths of  $\{2,4,6,8\}$  and observed that the model achieved the best performance with 4 encoder blocks, which is the depth used in our final architecture. This suggests that a moderate depth strikes the optimal balance between model complexity and performance, especially in cross-dataset settings, providing robust detection capabilities without overfitting.
+
+# 5. Conclusion
+
+Traditional DeepFake detectors, focused solely on face-manipulated content, struggle against newer, more sophisticated forgery methods involving T2V, I2V, and background manipulations. To address this, we proposed UNITE, leveraging the fully synthetic GTA-V [24] dataset for training. While this synthetic data is not AI-generated, our results show it effectively enhances UNITE's ability to detect AI-generated content and background manipulations.
+
+A key innovation in our approach is the AD-loss, which encourages the model to explore diverse spatial regions, enhancing its detection of manipulated areas beyond just the face region. This allows UNITE to excel not only in detecting AI-generated and background-altered videos but also in identifying traditional face-manipulated content. Our ablation studies demonstrate the robustness of UNITE, making it a valuable tool in synthetic video detection.
+
+Acknowledgements: This work was supported in part by funding from YouTube (Google).
+
+# References
+
+[1] Flux 1.1. https://blackforestlabs.ai/, 2024.1  
+[2] Morph studio. https://www.morphstudio.com/, 2024.7  
+[3] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In 2018 IEEE international workshop on information forensics and security (WIFS), pages 1-7. IEEE, 2018. 6  
+[4] Ibrahim M Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. Advances in Neural Information Processing Systems, 36, 2024. 2, 3  
+[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 5  
+[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 6, 7  
+[7] Baoying Chen and Shunquan Tan. Featuretransfer: Unsupervised domain adaptation for cross-domain deepfake detection. Security and Communication Networks, 2021(1): 9942754, 2021. 3  
+[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 7  
+[9] Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, et al. Demamba: Ai-generated video detection on million-scale genvideo benchmark. arXiv preprint arXiv:2405.19707, 2024. 1, 2, 5, 6, 7, 8  
+[10] Jikang Cheng, Zhiyuan Yan, Ying Zhang, Yuhao Luo, Zhongyuan Wang, and Chen Li. Can we leave deepfake data behind in training deepfake detector? arXiv preprint arXiv:2408.17052, 2024. 1  
+[11] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1133-1143, 2024. 2, 6  
+[12] Umur Aybars Ciftci, Ilke Demir, and Lijun Yin. Fakecatcher: Detection of synthetic portrait videos using biological signals. IEEE transactions on pattern analysis and machine intelligence, 2020. 6  
+[13] Sara Concas, Simone Maurizio La Cava, Roberto Casula, Giulia Orrù, Giovanni Puglisi, and Gian Luca Marcialis. Quality-based artifact modeling for facial deepfake detection in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3845–3854, 2024. 2, 6
+
+[14] Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 973-982, 2023. 2  
+[15] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 2  
+[16] Davide Cozzolino, Andreas Rössler, Justus Thies, Matthias Nießner, and Luisa Verdoliva. Id-reveal: Identity-aware deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15108-15117, 2021. 2  
+[17] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 1  
+[18] 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 ICLR, 2021. 3  
+[19] Ricard Durall, Margret Keuper, Franz-Josef Pfundt, and Janis Keuper. Unmasking deepfakes with simple features. arXiv preprint arXiv:1911.00686, 2019. 2  
+[20] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. In Proceedings of the 29th ACM international conference on multimedia, pages 3473-3481, 2021. 7  
+[21] Ying Guo, Cheng Zhen, and Pengfei Yan. Controllable guide-space for generalizable face forgery detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20818-20827, 2023. 6  
+[22] Harald Hanselmann, Shen Yan, and Hermann Ney. Deep fisher faces. In BMVC, 2017. 4  
+[23] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4  
+[24] Yuan-Ting Hu, Jiahong Wang, Raymond A Yeh, and Alexander G Schwing. Sail-vos 3d: A synthetic dataset and baselines for object detection and 3d mesh reconstruction from video data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1418-1428, 2021. 1, 2, 5, 6, 7, 8  
+[25] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2889–2898, 2020. 5, 7  
+[26] Yan Ju, Chengzhe Sun, Shan Jia, Shuwei Hou, Zhaofeng Si, Soumyya Kanti Datta, Lipeng Ke, Riky Zhou, Anita Nikolich, and Siwei Lyu. Deepfake-o-meter v2.0: An open platform for deepfake detection. In 2024 IEEE 7th Inter-
+
+national Conference on Multimedia Information Processing and Retrieval (MIPR), pages 439-445. IEEE, 2024. 1  
+[27] Pavel Korshunov and Sébastien Marcel. Deepfakes: a new threat to face recognition? assessment and detection. arXiv preprint arXiv:1812.08685, 2018. 5  
+[28] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 1  
+[29] Y Li. Exposing deepfake videos by detecting face warping artifact acts. arXiv preprint arXiv:1811.00656, 2018. 6  
+[30] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3207-3216, 2020. 1, 3, 5, 6, 7, 8  
+[31] Li Lin, Xinan He, Yan Ju, Xin Wang, Feng Ding, and Shu Hu. Preserving fairness generalization in deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16815-16825, 2024. 6  
+[32] Baoping Liu, Bo Liu, Ming Ding, Tianqing Zhu, and Xin Yu. Ti2net: temporal identity inconsistency network for deepfake detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4691-4700, 2023. 2, 6  
+[33] I Loshchilov and F Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 5  
+[34] Qingxuan Lv, Yuezun Li, Junyu Dong, Sheng Chen, Hui Yu, Huiyu Zhou, and Shu Zhang. Domainforensics: Exposing face forgery across domains via bi-directional adaptation. IEEE Transactions on Information Forensics and Security, 2024. 3  
+[35] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2  
+[36] Ghazal Mazaheri and Amit K Roy-Chowdhury. Detection and localization of facial expression manipulations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1035-1045, 2022. 2  
+[37] moonvalley.ai. moonvalley.ai. https://moonvalley.ai/, 2022.7  
+[38] John Mullan, Duncan Crawbuck, and Aakash Sastry. Hotshot-XL, 2023. 7  
+[39] Dat Nguyen, Nesryne Mejri, Inder Pal Singh, Polina Kuleshova, Marcella Astrid, Anis Kacem, Enjie Ghorbel, and Djamila Aouada. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17395–17405, 2024. 2  
+[40] Alvaro Lopez Pellicer, Yi Li, and Plamen Angelov. Pudd: Towards robust multi-modal prototype-based deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3809-3817, 2024. 2, 6
+
+[41] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 7  
+[42] Runway Research. Text driven video generation. https://research.runwayml.com/gen2, 2023.1,7  
+[43] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 1, 2, 3, 5, 6, 7, 8  
+[44] Md Shamim Seraj, Ankita Singh, and Shayok Chakraborty. Semi-supervised deep domain adaptation for deepfake detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1061-1071, 2024. 2  
+[45] Ke Sun, Shen Chen, Taiping Yao, Hong Liu, Xiaoshuai Sun, Shouhong Ding, and Rongrong Ji. Diffusionfake: Enhancing generalization in deepfake detection via guided stable diffusion. NeurIPS, 2024. 1  
+[46] Zekun Sun, Yujie Han, Zeyu Hua, Na Ruan, and Weijia Jia. Improving the efficiency and robustness of deepfakes detection through precise geometric features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3609-3618, 2021. 6  
+[47] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 7  
+[48] The New York Times. A.i. can now create lifelike videos. can you tell what's real? https://www.nytimes.com/interactive/2024/09/09/technology/ai-video-deepfake-runway-kling-quiz.html, 2024.5,6  
+[49] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 1  
+[50] Loc Trinh, Michael Tsang, Sirisha Rambhatla, and Yan Liu. Interpretable and trustworthy deepfake detection via dynamic prototypes. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1973-1983, 2021. 2  
+[51] A Vaswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A Gomez, L Kaiser, and I Polosukhin. Attention is all you need. NeurIPS, 2017. 4  
+[52] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2, 7  
+[53] Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, 2021. 3, 5
+
+[54] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 7  
+[55] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 2  
+[56] Deressa Wodajo and Solomon Atnafu. Deepfake video detection using convolutional vision transformer. arXiv preprint arXiv:2102.11126, 2021. 6  
+[57] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22658-22668, 2023. 1, 2, 6, 7  
+[58] Xin Yang, Yuezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8261-8265. IEEE, 2019. 5, 6, 8  
+[59] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3  
+[60] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision, pages 1–15, 2024. 7  
+[61] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 1, 4, 5, 7, 8  
+[62] Cairong Zhao, Chutian Wang, Guosheng Hu, Haonan Chen, Chun Liu, and Jinhui Tang. Istvt: interpretable spatial-temporal video transformer for deepfake detection. IEEE Transactions on Information Forensics and Security, 18: 1335-1348, 2023. 6  
+[63] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2  
+[64] Xinye Zhou, Hu Han, Shiguang Shan, and Xin Chen. Fine-grained open-set deepfake detection via unsupervised domain adaptation. IEEE Transactions on Information Forensics and Security, 2024. 2

2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81b3f9eceaac4bfd814574ae6faa54c65568f2d06252b76ead4e3083ef9954c5
+size 586882

2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/d41e3854-0a29-4aee-85fc-0fcf3b667c89_content_list.json

@@ -0,0 +1,1515 @@
+[
+    {
+        "type": "text",
+        "text": "Tra-MoE: Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning",
+        "text_level": 1,
+        "bbox": [
+            130,
+            128,
+            869,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jiange Yang  $^{1,2}$  Haoyi Zhu  $^{2,3}$  Yating Wang  $^{2,4}$  Gangshan Wu  $^{1}$  Tong He  $^{2}$  Limin Wang  $^{1,2*}$",
+        "bbox": [
+            107,
+            200,
+            895,
+            220
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ State Key Laboratory for Novel Software Technology, Nanjing University, China  $^{2}$ Shanghai AI Laboratory,  $^{3}$ University of Science and Technology of China,  $^{4}$ Tongji University",
+        "bbox": [
+            125,
+            220,
+            870,
+            257
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "jiangyang.jgy@gmail.com, {gswu,lmwang}@nju.edu.cn",
+        "bbox": [
+            272,
+            258,
+            720,
+            273
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{zhuhaoyi, wangyating, hetong}@pjlab.org.cn",
+        "bbox": [
+            308,
+            277,
+            687,
+            291
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "https://github.com/MCG-NJU/Tra-MoE",
+        "bbox": [
+            326,
+            294,
+            665,
+            308
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            344,
+            326,
+            359
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Learning from multiple domains is a primary factor that influences the generalization of a single unified robot system. In this paper, we aim to learn the trajectory prediction model by using broad out-of-domain data to improve its performance and generalization ability. Trajectory model is designed to predict any-point trajectories in the current frame given an instruction and can provide detailed control guidance for robotic policy learning. To handle the diverse out-of-domain data distribution, we propose a sparsely-gated MoE (Top-1 gating strategy) architecture for trajectory model, coined as Tra-MoE. The sparse activation design enables good balance between parameter cooperation and specialization, effectively benefiting from large-scale out-of-domain data while maintaining constant FLOPs per token. In addition, we further introduce an adaptive policy conditioning technique by learning 2D mask representations for predicted trajectories, which is explicitly aligned with image observations to guide action prediction more flexibly. We perform extensive experiments on both simulation and real-world scenarios to verify the effectiveness of Tra-MoE and adaptive policy conditioning technique. We also conduct a comprehensive empirical study to train Tra-MoE, demonstrating that our Tra-MoE consistently exhibits superior performance compared to the dense baseline model, even when the latter is scaled to match Tra-MoE's parameter count.",
+        "bbox": [
+            89,
+            376,
+            485,
+            753
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            89,
+            762,
+            220,
+            777
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In computer vision and natural language processing, significant progress has been made through learning models from multiple domains data. For example, many works use board data to perform self-supervised pre-training, and then adapt to diverse downstream tasks through zero-shot [1, 10, 49] or fine-tuning [13, 15, 23, 58]. Additionally, some research ef",
+        "bbox": [
+            89,
+            785,
+            485,
+            878
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/b20601886bd5d635928fdf1cbebeaae57f139d3d6ba8796761e8f9210cfab061.jpg",
+        "image_caption": [
+            "Figure 1. (a) Training with small-scale, in-domain data. (b) Joint training with in-domain data and large-scale, out-of-domain data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            558,
+            345,
+            870,
+            556
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "forts [43, 51, 71, 72] focus on learning a generalist model capable of unifying various tasks. In contrast, in robot learning, due to data scarcity and homogeneity, some groundbreaking methods [2, 37] resort to training only using in-domain data. Recently, some works [3, 4, 52, 61, 69] have shifted to learn from action-free human video data. A particularly scalable and promising paradigm [5, 20, 59, 63, 65] involves learning a trajectory prediction model from action-free video data, and then using small-scale action-labeled demonstrations to learn trajectory-guided policies, thereby achieving greater sample efficiency and better generalization.",
+        "bbox": [
+            511,
+            608,
+            908,
+            776
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite the potential benefits, it still remains underexplored how to effectively leverage broad out-of-domain video data for jointly training trajectory models. In particular, the impact of incorporating data encompassing diverse environments, objects, skills and embodiments has not fully investigated for in-domain tasks. For instance, ATM [63], a notable work in this field, mainly relies on in-domain data (same tasks and environments) for training both trajectory",
+        "bbox": [
+            511,
+            779,
+            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": "*Corresponding author.",
+        "bbox": [
+            109,
+            887,
+            236,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "6960",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/1116d206222c344b3495caf32192dc759e91b4b7ea155d6cd6e4cf158685afb2.jpg",
+        "image_caption": [
+            "Figure 2. Training trajectory prediction model from multiple domains. (a) The visualization of Dataset  $\\mathcal{D}_{\\text {ood }}$  and  $\\mathcal{D}_{\\text {in }}$ . The former may contain additional environments, objects, skills and embodiments. (b) Our pipeline: first co-training trajectory prediction model and then adapting it for downstream policy learning."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            90,
+            297,
+            257
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/dbc45401311683cf540e6205e3be6a9a5b51e719c61443d62b51f8072d7ecab6.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            308,
+            89,
+            477,
+            258
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "models and policies. We argue that learning from multiple domains is the primary factor to build a powerful trajectory model for policy guidance. Therefore, we propose to train trajectory model under the setup of jointly learning from large-scale out-of-domain and small-scale in-domain data, as shown in Fig. 2(a). However, directly using large-scale out-of-domain video data for hybrid pre-training involves two critical challenges: (i) Joint training on data from different environments, objects, skills, and embodiments will add learning difficulty and might lead to optimization conflicts. It is unclear how to effectively fuse the complementary information of multiple domains data within a single unified model. Naively expanding the out-of-domain training data does not necessarily lead to improved performance on in-domain tasks, as shown in Fig. 1, which results in a 5.6  $(57.6\\to 52.0)$  performance decrease on in-domain tasks. (ii) How to ensure both high performance and computational efficiency during trajectory model scaling up process.",
+        "bbox": [
+            88,
+            340,
+            482,
+            613
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address these challenges, we design a new sparsely-gated Mixture-of-Expert (MoE) architecture [53] to scale up our trajectory model, coined as Tra-MoE. Specifically, we integrate several MoE blocks to replace the original transformer blocks. This design enables efficient joint training of most parameters on the broad out-of-domain data, capturing the complementary patterns for mutual cooperation and the shared knowledge for better generalization. Meanwhile, for different data and different tokens within the same data, our Tra-MoE naturally forms a specialization when activating different experts. To maintain high computational efficiency while scaling up model capacity, we implement a top-1 gating strategy for token-choice, ensuring constant FLOPs per token. In sparse MoE, some auxiliary losses are commonly used to enhance performance, including the router z-loss [74] for improving training stability and the load-balancing loss [36] for balancing expert activations. We further conduct a comprehensive empirical study on scaling up trajectory models and find that the touter z-loss improves",
+        "bbox": [
+            89,
+            614,
+            482,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "performance while the load-balancing loss tends to cause performance degradation. Furthermore, we observe that when expanding pre-training data from small-scale in-domain to large-scale in-domain and out-of-domain data, employing MoE (Tra-MoE) can lead to obvious performance improvement on in-domain tasks, whereas the dense counterpart (Tra-baseline) suffers from performance drop, as illustrated in Fig. 1. Notably, even when expanding the dense Tra-baseline's parameter to match that of Tra-MoE, performance improvements remain elusive. These findings collectively demonstrate the effectiveness of our sparsely-gated MoE architecture, attributing its capability of balancing parameter cooperation and specialization when expanding training with out-of-domain data.",
+        "bbox": [
+            511,
+            90,
+            906,
+            301
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, given a pre-trained trajectory model, how to effectively condition the policy module using the 2D trajectories still remain an unresolved challenge. We further propose an adaptive policy conditioning technique for trajectory-guided policy. This enables better spatial alignment between 2D trajectories and images, while achieving more flexible trajectory representation and guidance for the policy. Finally, we also conduct extensive real-world experiments to further strengthen our conclusions. In summary, our main contributions can be summarized as follows:",
+        "bbox": [
+            511,
+            301,
+            908,
+            452
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose a sparsely-gated MoE architecture to learn trajectory models, called as Tra-MoE, with large-scale, out-of-domain, and action-free video data, coupled with a comprehensive empirical study on its training techniques.",
+            "- Extensive simulation and real-world environments demonstrate that our Tra-MoE (with Top-1 gating) can more effectively benefit from large-scale out-of-domain data (such as cross-embediments data and different physics engine simulated data) compared to the dense baseline.",
+            "- We demonstrate that Tra-MoE still significantly outperform expanded dense counterpart with equivalent parameters, highlighting the effectiveness of sparse MoE architecture in leveraging out-of-domain data through better parameter cooperation and specialization.",
+            "- We propose an adaptive policy conditioning technique for trajectory-guided policy, which ensures adaptive 2D trajectory representation and explicit spatial alignment with the observations, thereby achieving superior performance."
+        ],
+        "bbox": [
+            513,
+            455,
+            906,
+            726
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            513,
+            739,
+            653,
+            753
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Learning from multiple domains in Robot Learning. Scaling laws [30] have fueled substantial advancements in various fields. In robot learning, many recent studies [7, 8, 29, 51] investigate how compute, model size, and training data quantity affect the model performance on robotic manipulation tasks. Based on this, recent works have focused on collecting larger-scale real-world robotic datasets [11, 19, 32, 60, 64], and have successfully trained several robotics foundational models [6, 9, 16, 33, 38, 47, 57,",
+        "bbox": [
+            511,
+            763,
+            906,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "6961",
+        "bbox": [
+            482,
+            945,
+            513,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "[62] that demonstrate superior transfer capabilities. It is worth noting that these works mainly employ pre-training and fine-tuning paradigm rather than joint co-training paradigm, while focusing on evaluating the combination generalization and neglecting the performance impact on in-domain tasks. Additionally, these works directly combine heterogeneous robotic data with different action and observation spaces for joint training, potentially leading to sub-optimal solutions. In contrast, we leverage sparsely-gated MoE to incorporate larger-scale, out-of-domain, and action-less video data. Furthermore, our experiments demonstrate the importance of MoE for expanding out-of-domain data due to its superior capacity for parameter cooperation and specialization.",
+        "bbox": [
+            89,
+            90,
+            480,
+            287
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Mixture-of-Experts. The Mixture-of-Experts (MoE) framework [12] is built upon a simple yet effective concept: dividing a model into specialized components, or 'experts', each focusing on distinct tasks or data aspects. This paradigm allows for selective activation of relevant experts based on the input, thereby maintaining computational efficiency while leveraging a vast reservoir of specialized knowledge through increased model capacity. Recently, in computer vision, natural language processing, and multi-modal, there has been a surge of foundation models trained based on the MoE architecture, including Swin-Moe [24], MoE-LLaVA [39], DeepSeek-V2 [40], Mixtral-8x22B [28], and so on. MoE typically encompasses dense [48] and sparse [14] variants. Dense MoE activates all experts in each iteration, while sparse MoE activates only a subset, generally resulting in lower computational overhead. However, sparse MoE can suffer from expert load imbalance and training instability issues. Techniques such as adding noise, load balancing loss [36], and router z-loss [74] are often employed to mitigate these issues and improve performance. In our work, we conduct a comprehensive empirical study of these techniques when training our sparsely-gated MoE-based trajectory model.",
+        "bbox": [
+            91,
+            292,
+            486,
+            625
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Representations and Techniques for Policy Conditioning. Policy conditioning equips a multi-task robot policy with the ability to modulate its behavior by incorporating task-specific guidance information. This information, which may take various forms, guides the policy's action generation, enabling a single model to effectively learn and perform diverse tasks. Classic policy conditioning representations often encompass task identifiers [50], natural language instructions [8, 22, 45, 54], goal images [44], and demonstration videos [25, 27]. These are typically integrated with visual observations within the policy through concatenation or attention mechanisms. Recent works have explored alternative policy conditioning representations to enhance generalization and sample efficiency. These include leveraging internet-trained models to obtain bounding box [55] and object mask [42, 66, 68, 73], as well as learning plan [4], affordance [3, 46], future observations [17, 18, 34, 67], 2D or 3D trajectory [5, 61, 63, 65, 70] from human video data,",
+        "bbox": [
+            89,
+            628,
+            486,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "and also hand-drawn trajectory [21] or sketch [56]. Different from these approaches that directly utilize raw or manually crafted signals for integration with visual observations, we introduce an adaptive policy conditioning technique that explicitly maps 2D trajectories onto the visual observations in space and encodes them as learnable embeddings. These designs ensure adaptive 2D trajectory representation and explicit spatial alignment with the visual observations.",
+        "bbox": [
+            511,
+            90,
+            908,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Methodology",
+        "text_level": 1,
+        "bbox": [
+            511,
+            224,
+            648,
+            242
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we first describe our problem formulation in Sec. 3.1. Subsequently, we give an overview of our proposed framework, which is composed of a sparsely-gated MoE-based trajectory prediction model and a trajectory-guided policy model using the adaptive policy conditioning technique, detailed in Sec. 3.2. Finally, we describe our training process and loss functions in Sec. 3.3.",
+        "bbox": [
+            511,
+            250,
+            908,
+            354
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Problem Formulation",
+        "text_level": 1,
+        "bbox": [
+            511,
+            364,
+            718,
+            378
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Following the setup of previous works [5, 63, 65], our approach encompasses a trajectory prediction model to conduct language-conditioned any-point trajectory prediction and a trajectory-guided policy model to predict executable robot actions. In particular, as depicted in Fig. 2, we consider two types of datasets for training, which are large-scale, out-of-domain, and action-free videos  $\\mathcal{D}_{\\text{ood}}$ , as well as small-scale, in-domain, and action-labeled demonstrations  $\\mathcal{D}_{\\text{in}}$ . For trajectory prediction model, we employ  $\\mathcal{D}_{\\text{ood}}$  and  $\\mathcal{D}_{\\text{in}}$  for joint pre-training, whereas for trajectory-guided policy model, we only use  $\\mathcal{D}_{\\text{in}}$  for training.",
+        "bbox": [
+            511,
+            386,
+            908,
+            551
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Architecture",
+        "text_level": 1,
+        "bbox": [
+            511,
+            561,
+            648,
+            575
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ATM Revisited. In this paper, we use ATM [63] as the baseline for our method. The overall pipeline of ATM is illustrated on Fig. 2(b). It first learns a trajectory model from video data to perform language-conditioned trajectory prediction. The predicted trajectories is then used to learn a trajectory-guided policy using robot demonstrations. Specifically, as for trajectory model, its objective is to predict the future of any-point trajectories in a frame. Formally, given an image observation  $o_t$  at timestep  $t$ , any set of 2D query points simplified on the image frame  $\\mathbf{p}_t = \\{p_{t,k}\\}_{k=1}^K$ , and a language instruction  $\\ell$ , the trajectory model learns a mapping  $\\mathbf{p}_{t:t+H} = \\tau_\\theta(o_t, \\mathbf{p}_t, \\ell)$  that predicts the coordinates of these query points in the camera frame for the next  $H$  time steps, where  $\\mathbf{p}_{t:t+H} \\in \\mathbb{R}^{H \\times K \\times 2}$ . For the trajectory-guided policy, it leverages the current observation  $o_t$  and the predicted trajectory  $\\mathbf{p}_{t:t+H}$  to predict the subsequent robot actions.",
+        "bbox": [
+            511,
+            583,
+            908,
+            825
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Sparsely-gated MoE-based trajectory model (Tra-MoE). ATM proposes a track transformer to implement the trajectory model. In track transformer, language tokens, image tokens, and track tokens to be decoded involve global interactions within the transformer block. However, directly",
+        "bbox": [
+            511,
+            825,
+            908,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "6962",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/1b360506ebbfb7a53d46babead0df0d6744f34a203836d82d0e49690726e4618.jpg",
+        "image_caption": [
+            "Figure 3. (a) The pipeline of our sparsely-gated MoE-based trajectory model (Tra-MoE). (b) The pipeline of our trajectory-guided policy using the adaptive policy conditioning technique. Mapping means concatenating the trajectory mask with image observations, while learnable refers to setting each point in the trajectory mask as a learnable embedding."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            151,
+            88,
+            557,
+            414
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/d1ccaa97ebefa6d3fb7b4012c22df2c42abd6eb7a4331b344e1bc0685f8f71ae.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            555,
+            89,
+            836,
+            412
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "processing diverse domains and multiple modalities data within a unified transformer may yield suboptimal results. Hence, we propose our Tra-MoE framework. To develop our sparsely-gated MoE-based trajectory model, as depicted in Fig. 3(a), we replace certain layers of the transformer with MoE blocks. In each MoE block, it incorporates multiple feedforward networks (FFNs), each designated as an expert, and utilizes a gating function to activate a selected subset (Top-  $K$ ) of these experts. In this paper, we set  $K$  to a fixed value of 1 to ensure that the FLOPs per token remain constant relative to the baseline. Formally, the gating network  $\\mathcal{G}$ , parameterized by  $\\Theta$  and typically consisting of a linear-softmax network, yields the output  $\\mathcal{G}(\\mathbf{x_s};\\Theta)$ , where  $\\mathbf{x_s}$  represents our token sequences input. We formulate the sparsely-gated token-choice mechanism as follows:",
+        "bbox": [
+            88,
+            477,
+            485,
+            704
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {G} \\left(\\mathbf {x} _ {\\mathrm {s}}; \\boldsymbol {\\Theta}\\right) _ {i} = \\operatorname {s o f t m a x} (\\operatorname {T o p - K} (g \\left(\\mathbf {x} _ {\\mathrm {s}}; \\boldsymbol {\\Theta}\\right), k)) _ {i}, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            109,
+            746,
+            485,
+            763
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {T o p - K} (g \\left(\\mathbf {x} _ {\\mathbf {s}}; \\boldsymbol {\\Theta}\\right), k) _ {i} = \\left\\{ \\begin{array}{l l} g (\\mathbf {x}; \\boldsymbol {\\Theta}) _ {i}, & \\text {c o n d ,} \\\\ - \\infty , & \\text {o t h .} \\end{array} \\right., \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            112,
+            768,
+            483,
+            806
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {c o n d}: \\text {i f} g (\\mathbf {x}; \\boldsymbol {\\Theta}) _ {i} \\text {i s i n t o p - k e l e m e n t s o f} g (\\mathbf {x}; \\boldsymbol {\\Theta}). \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            109,
+            810,
+            483,
+            828
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "More specifically,  $\\operatorname{Top - K}(\\cdot ,k)$  function retains only the top-  $k$  values in a vector, setting the rest to  $-\\infty$ . Consequently, the output of the sparsely-gated MoE layer can be",
+        "bbox": [
+            89,
+            854,
+            485,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "formulated as:",
+        "bbox": [
+            513,
+            477,
+            612,
+            491
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {F} _ {\\text {s p a r s e}} ^ {\\mathrm {M o E}} \\left(\\mathbf {x} _ {\\mathbf {s}}; \\boldsymbol {\\Theta}, \\left\\{\\mathbf {W} _ {i} \\right\\} _ {i = 1} ^ {K}\\right) = \\sum_ {i = 1} ^ {K} \\mathcal {G} \\left(\\mathbf {x} _ {\\mathbf {s}}; \\boldsymbol {\\Theta}\\right) _ {i} f _ {i} \\left(\\mathbf {x} _ {\\mathbf {s}}; \\mathbf {W} _ {i}\\right), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            524,
+            500,
+            906,
+            556
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $f_{i}$  represents the i-th expert, usually a linear-GELU-linear FFN layer, is parameterized by  $\\mathbf{W}_i$ .",
+        "bbox": [
+            511,
+            568,
+            906,
+            598
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Trajectory-guided policy using the adaptive policy conditioning technique. After completing the trajectory model learning, we develop our trajectory-guided policy through the adaptive policy conditioning technique based on the initial ATM, as shown in Fig. 3(b). In our policy model, independently encoded track tokens and image tokens undergo a early fusion within the fusion transformer. Subsequently, the fused features are further integrated with the raw trajectory data through a late fusion mechanism along the channel dimension, ensuring the preservation of original trajectory information for more accurate action prediction.",
+        "bbox": [
+            511,
+            598,
+            908,
+            763
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Fundamentally, the policy model bridges the gap between the predicted 2D any-point trajectories and the executable 3D robot actions, functioning as an inverse dynamic model and eliminating the necessity of other task specification. However, how to effectively condition the policy using the predicted 2D trajectories remains an unresolved challenge. Based on our analysis, we present two key insights: (i) Explicit spatial alignment and fusion of the 2D trajectories with image observations turns out to be highly beneficial.",
+        "bbox": [
+            511,
+            763,
+            910,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "6963",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "It simplifies the mapping of 2D trajectories to image space and highlights motion-relevant regions, enabling the policy model to focus on important areas and reducing the learning difficulty. (ii) Trajectory positions exhibit different characteristics: starting points typically emphasize local motion, while endpoints focus more on global trend. Therefore, an adaptive trajectory representation and conditioning technique is expected to flexibly guide the policy by effectively capturing these variations.",
+        "bbox": [
+            89,
+            90,
+            483,
+            226
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Based on the aforementioned analysis, we propose an adaptive policy conditioning technique to incorporate trajectory information into 2D images. Specifically, we construct an additional mask modality, and populate it with 2D trajectories based on their specific spatial locations, while setting the value of each point in the trajectory mask as a learnable embedding. Before encoding the image, we concatenate this mask modality with the image observations on the channel dimension, forming a tensor of dimensions  $H \\times W \\times 4$ , as shown in Fig. 3(b). To sum up, our adaptive policy conditioning technique yield an adaptive 2D trajectory representation, which could be explicitly aligned with the image observations to guide the policy prediction more flexibly.",
+        "bbox": [
+            89,
+            227,
+            483,
+            422
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3. Training",
+        "text_level": 1,
+        "bbox": [
+            89,
+            430,
+            194,
+            446
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We argue that trajectory models could be shared by multiple domains data. Therefore, we first pre-train a generalizable trajectory model on  $\\mathcal{D}_{\\text{ood}} \\cup \\mathcal{D}_{\\text{in}}$ . The  $\\mathcal{D}_{\\text{ood}}$  typically contains additional environments, objects, skills, and embodiments. Then we train the trajectory-guided policy model on  $\\mathcal{D}_{\\text{in}}$  based on the pre-trained trajectory model. In trajectory-guided policy training phase, we freeze the trajectory model. Pre-training sparsely-gated MoE-based trajectory model. Following ATM, we train the trajectory model using the ground truth tracks generated by CoTracker [31]. We use MSE to optimize the trajectory prediction loss  $\\mathcal{L}_{\\text{tra}}$ . Additionally, we also consider adding an auxiliary load-balancing loss [36] to balance the activation frequency among experts. Given a batch of queries  $\\mathcal{B}$ , it contains B samples and each sample contains S tokens. We define  $\\mathcal{Q}_i$  as the proportion of tokens distributed to expert  $i$ , and  $\\mathcal{P}_i$  as the proportion of the gating probability assigned to expert  $i$ . The  $\\mathcal{L}_{\\text{lo-ba}}$  can be formulated as follows:",
+        "bbox": [
+            89,
+            452,
+            483,
+            724
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {Q} _ {i} = \\frac {1}{S \\cdot B} \\sum_ {x _ {s} ^ {m} \\in \\mathcal {B}} \\mathbb {1} \\left\\{\\operatorname {a r g m a x} \\mathcal {G} \\left(\\mathbf {x} _ {\\mathbf {s}} ^ {\\mathbf {m}}; \\boldsymbol {\\Theta}\\right) = i \\right\\}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            135,
+            742,
+            483,
+            779
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {P} _ {i} = \\frac {1}{S \\cdot B} \\sum_ {x _ {\\mathrm {s}} ^ {m} \\in \\mathcal {B}} \\mathcal {G} \\left(\\mathbf {x} _ {\\mathrm {s}} ^ {\\mathrm {m}}; \\boldsymbol {\\Theta}\\right) _ {i}, \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            133,
+            782,
+            483,
+            819
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {l o - b a}} = N \\sum_ {i = 1} ^ {N} \\mathcal {Q} _ {i} \\mathcal {P} _ {i}. \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            133,
+            824,
+            483,
+            864
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Additionally, ST-MoE [74] proposed using the router  $z$ -loss to penalize large logits entering the gating network,",
+        "bbox": [
+            89,
+            869,
+            483,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "thereby improving training stability. We similarly applies it to the training of our trajectory model, as shown below:",
+        "bbox": [
+            511,
+            90,
+            906,
+            121
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {z} (x) = \\frac {1}{S} \\sum_ {k = 1} ^ {S} \\left(\\log \\sum_ {i = 1} ^ {N} e ^ {\\mathbf {g} _ {i} ^ {(k)}}\\right) ^ {2}, \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            593,
+            132,
+            906,
+            176
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $S$  is the total number of tokens,  $N$  is the total number of experts, and  $\\mathbf{g} \\in \\mathcal{R}^{S \\times N}$  are the logits going into the router. In summary, the trajectory model's training loss is as follow:",
+        "bbox": [
+            511,
+            181,
+            906,
+            229
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {t r a}} \\cdot \\mathcal {L} _ {\\text {t r a}} + \\lambda_ {\\text {l o - b a}} \\cdot \\mathcal {L} _ {\\text {l o - b a}} + \\lambda_ {z} \\cdot \\mathcal {L} _ {z}. \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            553,
+            256,
+            906,
+            272
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training trajectory-guided policy through adaptive policy conditioning technique. For trajectory-guided policy training, we employ the paradigm of behavior cloning. Specifically, we use the MSE between predicted actions and ground truths as the loss function.",
+        "bbox": [
+            511,
+            284,
+            906,
+            359
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            513,
+            375,
+            645,
+            391
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We perform evaluation experiments using the LIBERO benchmark [41] and a real low-cost dual-arm robot [35]. In experiments, we quantitatively find that the success rate of downstream manipulation tasks is generally positively correlated with the performance of the trajectory model. Detailed quantitative experiment results can be found in the appendix. Meanwhile, CoTracker [31] can provide relatively accurate trajectory annotations; therefore, we report the average success rate of downstream manipulation tasks as our metric. Our experiments aim to study the following questions:",
+        "bbox": [
+            511,
+            398,
+            906,
+            550
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "Q1: Does the techniques of router z-loss, load-balancing loss, and adding noise to gating logits enhance the training of sparsely-gated Tra-MoE?",
+            "Q2: Does sparsely-gated Tra-MoE, compared to the dense Tra-baseline, benefit more from large-scale out-of-domain data (such as cross-embediments and different physics engine data) and demonstrate superior scaling up capabilities?",
+            "Q3: When we scale up the dense Tra-baseline to match the parameter count of Tra-MoE, does the sparsely-gated Tra-MoE still maintain its advantage?",
+            "Q4: Is our proposed adaptive policy conditioning technique for trajectory-guided policy more effective than ATM?"
+        ],
+        "bbox": [
+            511,
+            551,
+            906,
+            732
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Simulation Experiments",
+        "text_level": 1,
+        "bbox": [
+            511,
+            742,
+            736,
+            758
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Experiment setup. The LIBERO [41] is divided into five categories: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, LIBERO-Long, and LIBERO-90. LIBERO-90 includes 90 tasks, while each of the other categories contains 10 tasks. Compared to other categories, LIBERO-90 presents a highly varied selection of object categories, layouts, environments, and skill goals. It is typically used for pre-training rather than joint co-training. Therefore, our main ablation experiments select 90 tasks from LIBERO-90 to construct",
+        "bbox": [
+            511,
+            763,
+            906,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "6964",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/9504d9d45d6a4b0a101bfc2bfb7656c82744da79452a9b4ebd7335a1fd79c306.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td>0</td><td>62.0</td><td>73.0</td><td>71.0</td><td>21.5</td><td>56.9</td></tr><tr><td>5·10-4</td><td>69.5</td><td>73.5</td><td>68.0</td><td>22.5</td><td>58.4</td></tr><tr><td>1·10-4</td><td>62.5</td><td>81.0</td><td>73.5</td><td>28.5</td><td>61.4</td></tr><tr><td>1·10-5</td><td>60.0</td><td>65.5</td><td>68.5</td><td>27.0</td><td>55.3</td></tr><tr><td>1·10-6</td><td>60.0</td><td>79.0</td><td>53.0</td><td>31.5</td><td>55.9</td></tr></table>",
+        "bbox": [
+            116,
+            88,
+            472,
+            189
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/0f0712d7bceb9eb4623303c792b6ec47f95513fa514dcad77622b0c57072fcfd.jpg",
+        "table_caption": [
+            "(a) The value of  $\\lambda_z$"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td>w/o. noise</td><td>62.0</td><td>73.0</td><td>71.0</td><td>22.5</td><td>56.9</td></tr><tr><td>w. noise</td><td>71.5</td><td>67.0</td><td>57.5</td><td>27.0</td><td>55.8</td></tr></table>",
+        "bbox": [
+            102,
+            209,
+            478,
+            266
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/03a497c14e81d85247113fd03473c7b66e52786ef05f766549381404274a1fcc.jpg",
+        "table_caption": [
+            "(c) Adding a noise term to the gating logits."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Num.</td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td>1</td><td>49.5</td><td>67.0</td><td>56.5</td><td>35.0</td><td>52.0</td></tr><tr><td>2</td><td>59.0</td><td>72.5</td><td>65.5</td><td>27.0</td><td>56.0</td></tr><tr><td>3</td><td>64.0</td><td>68.0</td><td>71.0</td><td>17.5</td><td>55.1</td></tr><tr><td>4</td><td>62.0</td><td>73.0</td><td>71.0</td><td>21.5</td><td>56.9</td></tr></table>",
+        "bbox": [
+            117,
+            300,
+            460,
+            386
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/750356bdb0dea26187e847fb2fc535848a12c8434edd50e8202d3ed706bd395d.jpg",
+        "table_caption": [
+            "(e) The number of experts."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td>0</td><td>62.0</td><td>73.0</td><td>71.0</td><td>22.5</td><td>56.9</td></tr><tr><td>1·10-3</td><td>58.5</td><td>74.0</td><td>60.0</td><td>18.0</td><td>52.6</td></tr><tr><td>1·10-5</td><td>61.0</td><td>69.5</td><td>56.0</td><td>26.0</td><td>53.1</td></tr><tr><td>1·10-7</td><td>66.5</td><td>67.0</td><td>73.5</td><td>22.0</td><td>57.3</td></tr></table>",
+        "bbox": [
+            526,
+            88,
+            879,
+            174
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/17bdff4583247d83ff735f2dd80fe5f356987ff135fffb08860a3e448251094e.jpg",
+        "table_caption": [
+            "(b) The value of  $\\lambda_{lo - ba}$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td>MoE</td><td>62.5</td><td>81.0</td><td>73.5</td><td>28.5</td><td>61.4</td></tr><tr><td>Width</td><td>66.5</td><td>50.5</td><td>61.5</td><td>15.0</td><td>48.4</td></tr><tr><td>Depth</td><td>61.5</td><td>75.5</td><td>46.5</td><td>26.5</td><td>52.5</td></tr></table>",
+        "bbox": [
+            529,
+            209,
+            875,
+            280
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/6139101233449309a3573c9cefcb724605042ca4863f91c124222887c560fdb4.jpg",
+        "table_caption": [
+            "(d) Scaling Tra-baseline from width and depth."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>MoE</td><td>OOD</td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td></td><td></td><td>67.5</td><td>68.5</td><td>68.0</td><td>26.5</td><td>57.6</td></tr><tr><td>✓</td><td></td><td>63.0</td><td>64.0</td><td>60.5</td><td>19.5</td><td>51.8</td></tr><tr><td></td><td>✓</td><td>49.5</td><td>67.0</td><td>56.5</td><td>35.0</td><td>52.0</td></tr><tr><td>✓</td><td>✓</td><td>62.5</td><td>81.0</td><td>73.5</td><td>28.5</td><td>61.4</td></tr></table>",
+        "bbox": [
+            500,
+            300,
+            895,
+            386
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "dataset  $\\mathcal{D}_{\\text{ood}}$ , utilizing 20 action-free videos per task; and choose 40 tasks from LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long to build dataset  $\\mathcal{D}_{\\text{in}}$ , using 10 action-labeled demonstrations per task. In summary, for training Tra-MoE, our main ablation experiments use a total of 2, 200 action-free videos, with an out-of-domain to in-domain data ratio of  $9:2$ . Additionally, on this basis, we also further select 92 tasks from the RLbench [26] (using CoppeliaSim rather than MuJoCo as physics engine) simulation environment and generate 5 videos for each task, totaling 2, 660 videos in conjunction with the LIBERO data, to train Tra-MoE. The different physics engines and dynamics properties further increase the data distribution discrepancy in  $\\mathcal{D}_{\\text{ood}}$ . Finally, as for the trajectory-guided policy, we train a separate multi-task policy for each category. Each policy undergoes behavioral cloning training with 100 action-labeled demonstrations from 10 tasks.",
+        "bbox": [
+            88,
+            431,
+            485,
+            689
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implement details. Building on ATM [63], we develop our sparsely-gated Tra-MoE and adaptive policy conditioning technique. Specifically, when using only LIBERO data, we replace the 1th, 5th, and 8th layers of the track transformer with MoE blocks. When further incorporating RLbench data, we additionally replace the 2th and 7th layers. Unless otherwise specified, we set the number of experts to 4 by default and use a top-1 gating strategy for token choice to ensure the FLOPs per token remain constant.",
+        "bbox": [
+            88,
+            694,
+            482,
+            830
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Experiment results and analysis. We present empirical simulation results to address the aforementioned research questions and make the following findings:",
+        "bbox": [
+            89,
+            835,
+            482,
+            881
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Result 1: In Tab. 1a, we examine the effect of the router",
+        "bbox": [
+            89,
+            885,
+            482,
+            898
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/694232bc2cfc422c4abc74e342ff3c257c79792d291197b661c0015ed70541ee.jpg",
+        "table_caption": [
+            "(f) Scaling with MoE and OOD data.",
+            "Table 1. The ablation experiments of the sparsely-gated Tra-MoE. All baselines are the track transformer of ATM."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td>ATM</td><td>51.0</td><td>42.0</td><td>55.5</td><td>19.0</td><td>41.9</td></tr><tr><td>Tra-MoE</td><td>54.0</td><td>62.0</td><td>61.0</td><td>41.0</td><td>54.5</td></tr></table>",
+        "bbox": [
+            524,
+            430,
+            893,
+            487
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2. The comparison results when training trajectory model using RLbench video data.",
+        "bbox": [
+            511,
+            497,
+            903,
+            526
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "z-loss weight  $\\lambda_z$ . The results indicate that a higher  $\\lambda_z$  can effectively enhance performance, with an improvement of 4.5 (61.4 vs. 56.9) at  $\\lambda_z = 1 \\cdot 10^{-4}$ . In Tab. 1b, we study the impact of the load-balancing loss weight  $\\lambda_{lo - ba}$ . The results show a significant performance decline with larger values of  $\\lambda_{lo - ba}$ , with only a slight improvement ( $56.7 \\rightarrow 57.3$ ) at  $\\lambda_{lo - ba} = 1 \\cdot 10^{-7}$ . In Tab. 1c, we verify the effectiveness of adding noise to the gating logits, which results in a performance decrease of by 1.1 ( $56.9 \\rightarrow 55.8$ ).",
+        "bbox": [
+            511,
+            534,
+            906,
+            671
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Finding 1: Based on the above results, we find that using router z-loss to penalize large logits entering the gating network can effectively improve training stability and lead to performance gains. Meanwhile, the load-balancing loss used to balance the activation frequency among experts and the addition of noise to enhance expert allocation exploration cannot bring effective performance improvements. An intuitive explanation is that when the data distribution is uneven, the load-balancing loss might force experts to set shared parameters on data with a large domain gap, which could lead to mutual counteraction of learning gradients. This undermines the advantage of parameter specialization in sparsely-gated MoE. Similarly, adding noise is also detrimental to learning parameter specialization among experts.",
+        "bbox": [
+            511,
+            672,
+            908,
+            883
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Result 2: We comprehensively investigate the impact of in",
+        "bbox": [
+            511,
+            885,
+            906,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "6965",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/32262f4e4a8ee1cb12f5562878940f70d642710c03559def97db4bef228462b5.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td>ATM</td><td>62.5</td><td>81.0</td><td>73.5</td><td>28.5</td><td>61.4</td></tr><tr><td>+ DM</td><td>69.0</td><td>58.0</td><td>85.5</td><td>33.5</td><td>61.5</td></tr><tr><td>+ AM</td><td>69.5</td><td>77.0</td><td>88.0</td><td>30.5</td><td>66.3</td></tr></table>",
+        "bbox": [
+            114,
+            88,
+            460,
+            160
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3. The ablation experiments of the adaptive policy conditioning technique. DM (RT-Trajectory style) and AM respectively represent hand-drawn mask and adaptive mask.",
+        "bbox": [
+            89,
+            171,
+            482,
+            213
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "producing sparsely-gated MoE and large-scale out-of-domain data on scaling up trajectory models in Tab. 1f. Firstly, when training only with in-domain data, Tra-baseline achieves an average success rate of 57.6, while Tra-MoE only achieves an average success rate of 51.8. Secondly, when adding out-of-domain data to jointly train Tra-baseline, the average success rate drops to  $52.0$  ( $57.6 \\rightarrow 52.0$ ), with particularly significant performance declines in LIBERO-Spatial and LIBERO-Object. However, when adding out-of-domain data to jointly train Tra-MoE, the average success rate increases to  $61.4$  ( $51.8 \\rightarrow 61.4$ ). Additionally, in Tab. 1e, we also explore the impact of the number of experts in sparsely-gated MoE. The results show that using just two experts can lead to an increase of 4.0 (52.0 vs. 56.0) in the average success rate. With additional experts, performance generally trends upward, albeit with some fluctuations. Finally, we report the comparison results of training trajectory model by further using RLbench data in Tab. 2. The results indicate that although the introduction of RLbench data led to a performance drop in LIBERO, our Tra-MoE still outperform the baseline by 12.6 (54.5 vs. 41.9).",
+        "bbox": [
+            91,
+            220,
+            483,
+            536
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Finding 2: Based on the above results, we have the following findings and analysis: (i) When the data scale is small, naively expanding the model capacity through sparsely-gated MoE does not lead to performance improvements. A explanation is that the larger model may lead to a overfitting problem when training on small-scale data. Moreover, small-scale data often leads to some experts not being sufficiently trained, thereby affecting performance. (ii) Naively introducing out-of-domain data for joint training often leads to performance degradation in the target domain. However, simultaneously incorporating out-of-domain data and expanding models with sparsely-gated MoE may improve performance or alleviate performance degradation. This phenomenon suggests that although out-of-domain data generally enhances model generalization, it could hinder task-specific performance in the target domain due to insufficient in-domain data learning. In contrast, expanding the model with out-of-domain data using sparsely-gated MoE achieves a more optimal balance between parameter cooperation and specialization across diverse data and even different tokens within the same data input. This design allows these large-scale data to jointly train most of parameters, ensuring the capture of general patterns for mutual cooperation. (iii) Even with a few experts, Tra-MoE exhibits significant scaling up",
+        "bbox": [
+            91,
+            537,
+            485,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "capabilities. Additionally, further increasing the number of experts generally leads to performance improvement. However, this improvement is not guaranteed due to factors like data-expert complexity matching and training stability.",
+        "bbox": [
+            511,
+            90,
+            906,
+            151
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Result 3: In Tab. 1d, we also investigate the performance comparison between the dense Tra-baseline and the sparse Tra-MoE with the same number of parameters. Specifically, we expand Tra-baseline to match the model capacity of Tra-MoE by increasing its model width or depth separately. These expanded versions achieve average success rates of 48.4 and 52.5 respectively, which are significantly lower than Tra-MoE's 61.4.",
+        "bbox": [
+            511,
+            154,
+            908,
+            273
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Finding 3: Based on these results, we are surprised to find that simply expanding the dense Tra-baseline in depth and width can not effectively improve performance. This highlights the importance of the sparsely-gated MoE architecture when training with large-scale out-of-domain data, as it ensures dynamic activation of different experts based on different data input and different tokens within the same data input, significantly reducing optimization conflict problems.",
+        "bbox": [
+            511,
+            276,
+            908,
+            398
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Result 4: In Tab. 3, we investigate the effectiveness of the adaptive policy conditioning. Specifically, building upon ATM, we first introduce an additional hand-drawn mask modality before image encoding. In this mask, the first half of the trajectory is set to 128, while the second half is set to 255. We call it 'ATM + hand-drawn mask' and this is similar to RT-Trajectory [21]. Next, we replace the hand-drawn mask with an adaptive mask, where each point on the trajectory is instantiated as a learnable embedding. We call it 'ATM + adaptive mask'. The results show that 'ATM + hand-drawn mask' achieves only a 0.1 improvement compared to the baseline. However, except for a significant performance decrease in LIBERO-Goal, it achieves performance improvements of 6.5, 12.0, and 5.0 in LIBERO-Spatial, LIBERO-Object, and LIBERO-Long respectively. In addition, the 'ATM + adaptive mask' achieves a performance improvement of 4.9 (66.3 vs. 61.4) compared to the baseline. It further improves performance across all suites compared to 'ATM + hand-drawn mask', especially achieving a 19.0 performance gains in LIBERO-Goal (77.0 vs. 58.0).",
+        "bbox": [
+            511,
+            400,
+            908,
+            700
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Finding 4: Based on these results, we find that explicitly mapping 2D trajectories to visual observations significantly improves the performance of LIBERO-Spatial, LIBERO-Object, and LIBERO-Long, but causes a notable decrease in LIBERO-Goal. By instantiating the points in the 2D trajectory as learnable embeddings, we not only further enhances the performance of LIBERO-Spatial, LIBERO-Object, and LIBERO-Long, but also significantly improves LIBERO-Goal. We analyze that the decline in LIBERO-Goal is mainly due to overlapping front and back parts of some trajectories in the 2D images, leading to mapping errors. However, our adaptive trajectory representation and conditioning technique effectively mitigates this issue. In summary, these",
+        "bbox": [
+            511,
+            704,
+            910,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "6966",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/bd6dc923dc11f3b6cfa2edcdc0fc56ce008272cd16430bccf28e46cbcbea5b57.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Human Data</td><td>MoE</td><td>Adaptive Trajectory Mask</td><td>Pour</td><td>Push</td><td>Pick and Pass</td><td>Tissue</td><td>Fold</td><td>Average</td></tr><tr><td></td><td></td><td></td><td>40.0</td><td>45.0</td><td>50.0</td><td>30.0</td><td>25.0</td><td>38.0</td></tr><tr><td>✓</td><td></td><td></td><td>45.0</td><td>35.0</td><td>50.0</td><td>25.0</td><td>35.0</td><td>38.0</td></tr><tr><td>✓</td><td>✓</td><td></td><td>40.0</td><td>50.0</td><td>60.0</td><td>35.0</td><td>45.0</td><td>46.0</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>60.0</td><td>70.0</td><td>65.0</td><td>35.0</td><td>50.0</td><td>56.0</td></tr></table>",
+        "bbox": [
+            125,
+            89,
+            870,
+            175
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4. The ablation of ood human data, MoE architecture, and adaptive policy conditioning technique on our real-world tasks.",
+        "bbox": [
+            119,
+            186,
+            872,
+            199
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/84503aacec77ac5c54356f89d97bcde412c1ed4dba170d2a1d766ebf2aa969d0.jpg",
+        "image_caption": [
+            "Figure 4. Left: The real-world experiments hardware platform setup. Right: The real-world tasks evaluation demonstrations."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            155,
+            204,
+            834,
+            398
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "experiments collectively demonstrate the effectiveness of explicitly aligning 2D trajectories with images in space as well as applying the adaptive conditioning technique.",
+        "bbox": [
+            89,
+            435,
+            483,
+            481
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.2. Real-World Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            494,
+            318,
+            511
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Experiment setup. As shown in Fig 4, we utilize a dual-arm robot [35] to conduct real-world experiments. We evaluate 5 tasks, which include two single-arm tasks: pouring water and pushing a vegetable to the side of the cutting board, as well as three dual-arm tasks: pulling out tissues, folding towels, and picking up and passing a holder. Specifically, for each task, we collect 50 action-labeled robot demonstrations. Concurrently, we also collect 50 videos of human (cross-embediments) performing for each task. Compared to the robot data, the human video encompasses more environment variations and distractor objects. Please see the appendix for more detailed evaluation details and real-world setup.",
+        "bbox": [
+            89,
+            518,
+            483,
+            700
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Implement details. Following our simulation experiments, we utilize two camera views and uniformly resize their resolution to  $128 \\times 128$  for trajectory model and policy training. We train our real-world Tra-MoE on borh human video and robot video data, with a separate policy trained for each task. Therefore, our real-world Tra-MoE integrates data across different embodiments (human, single-arm robot, and dual-arm robot). They have significantly different visual appearances and dynamics properties. Finally, we also present some real-world task evaluation demonstrations in Fig. 4.",
+        "bbox": [
+            89,
+            702,
+            483,
+            853
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Experiment results and analysis. To further strengthen our conclusions that sparsely-gated MoE can better learn trajectory model from out-of-domain data and our adaptive policy",
+        "bbox": [
+            89,
+            854,
+            483,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "condition technique could lead to more flexible guidance for policy learning, we conduct real-world ablation experiments, as shown in Tab. 4. The results indicate that simply introducing human video data cannot lead to performance improvement. However, when expanding the model by MoE architecture, the performance increases from  $38\\%$  to  $46\\%$ . We speculate this may be due to the significant differences in appearance and dynamics between human and robot, with the former typically moving much faster and more flexibly. Finally, when further incorporating an adaptive trajectory mask for policy conditioning, the performance is further increased from  $46\\%$  to  $56\\%$ .",
+        "bbox": [
+            511,
+            435,
+            906,
+            616
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            513,
+            619,
+            633,
+            633
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this work, we have introduced Tra-MoE, a novel framework that utilizes a sparsely-gated MoE architecture with Top-1 gating strategy to learn trajectory model from multiple domains. Tra-MoE achieves superior parameter cooperation and specialization, enabling more effective utilization of large-scale out-of-domain data while maintaining constant FLOPs per token. We train our Tra-MoE and conduct a comprehensive empirical study, demonstrating that TraMoE consistently exhibits superior performance compared to the dense baseline, even when the latter is scaled to match Tra-MoE's parameter count. Additionally, we propose an adaptive policy conditioning technique for trajectory-guided policy, ensuring adaptive 2D trajectory representation and explicit spatial alignment with observations, thereby achieving superior performance. Extensive experiments in both simulated and real-world environments validate our findings and demonstrate the effectiveness of our proposed methods.",
+        "bbox": [
+            511,
+            643,
+            908,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "6967",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            89,
+            90,
+            230,
+            106
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work is supported by the National Key R&D Program of China (No. 2022ZD0160900), Jiangsu Frontier Technology Research and Development Program (No. BF2024076), and the Collaborative Innovation Center of Novel Software Technology and Industrialization.",
+        "bbox": [
+            89,
+            106,
+            485,
+            183
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            91,
+            191,
+            187,
+            208
+        ],
+        "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 preprint arXiv:2303.08774, 2023. 1",
+            "[2] OpenAI: Marcin Andrychowicz, Bowen Baker, Maciek Chogiej, Rafal Jozefowicz, Bob McGrew, Jakub Pachocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, et al. Learning dexterous in-hand manipulation. *IJRR*, 39(1): 3-20, 2020. 1",
+            "[3] Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. Affordances from human videos as a versatile representation for robotics. In CVPR, pages 13778-13790, 2023. 1, 3",
+            "[4] Homanga Bharadhwaj, Abhinav Gupta, Vikash Kumar, and Shubham Tulsiani. Towards generalizable zero-shot manipulation via translating human interaction plans. In ICRA, pages 6904-6911. IEEE, 2024. 1, 3",
+            "[5] Homanga Bharadhwaj, Roozbeh Mottaghi, Abhinav Gupta, and Shubham Tulsiani. Track2act: Predicting point tracks from internet videos enables diverse zero-shot robot manipulation. arXiv preprint arXiv:2405.01527, 2024. 1, 3",
+            "[6] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2",
+            "[7] Konstantinos Bousmalis, Giulia Vezzani, Dushyant Rao, Coline Manon Devin, Alex X Lee, Maria Bauza Villalonga, Todor Davchev, Yuxiang Zhou, Agrim Gupta, Akhil Raju, et al. Robocat: A self-improving generalist agent for robotic manipulation. TMLR, 2023. 2",
+            "[8] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2, 3",
+            "[9] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 2",
+            "[10] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1",
+            "[11] Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xu Huang, Shu Jiang, et al. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems. arXiv preprint arXiv:2503.06669, 2025. 2"
+        ],
+        "bbox": [
+            93,
+            217,
+            483,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[12] Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. A survey on mixture of experts. arXiv preprint arXiv:2407.06204, 2024. 3",
+            "[13] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 1",
+            "[14] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. Mod-squad: Designing mixtures of experts as modular multitask learners. In CVPR, pages 11828-11837, 2023. 3",
+            "[15] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1",
+            "[16] Ria Doshi, Homer Rich Walke, Oier Mees, Sudeep Dasari, and Sergey Levine. Scaling cross-embodied learning: One policy for manipulation, navigation, locomotion and aviation. In CoRL, 2",
+            "[17] Yilun Du, Sherry Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, Leslie Pack Kaelbling, et al. Video language planning. In The Twelfth International Conference on Learning Representations. 3",
+            "[18] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. NeurIPS, 36, 2024. 3",
+            "[19] Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A robotic dataset for learning diverse skills in one-shot. In RSS 2023 Workshop on Learning for Task and Motion Planning, 2023.",
+            "[20] Chongkai Gao, Haozhuo Zhang, Zhixuan Xu, Cai Zhehao, and Lin Shao. Flip: Flow-centric generative planning as general-purpose manipulation world model. In ICLR. 1",
+            "[21] Jiayuan Gu, Sean Kirmani, Paul Wohlhart, Yao Lu, Montserrat Gonzalez Arenas, Kanishka Rao, Wenhao Yu, Chuyuan Fu, Keerthana Gopalakrishnan, Zhuo Xu, et al. Rt-trajectory: Robotic task generalization via hindsight trajectory sketches. In ICLR. 3, 7",
+            "[22] Pierre-Louis Guhur, Shizhe Chen, Ricardo Garcia Pinel, Makarand Tapaswi, Ivan Laptev, and Cordelia Schmid. Instruction-driven history-aware policies for robotic manipulations. In CORL, pages 175-187. PMLR, 2023. 3",
+            "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1",
+            "[24] Changho Hwang, Wei Cui, Yifan Xiong, Ziyue Yang, Ze Liu, Han Hu, Zilong Wang, Rafael Salas, Jithin Jose, Prabhat Ram, et al. Tutel: Adaptive mixture-of-experts at scale. Proceedings of Machine Learning and Systems, 5:269-287, 2023. 3",
+            "[25] Vidhi Jain, Maria Attarian, Nikhil J Joshi, Ayzaan Wahid, Danny Driess, Quan Vuong, Pannag R Sanketi, Pierre Sermanet, Stefan Welker, Christine Chan, et al. Vid2robot: End-to-end video-conditioned policy learning with cross-attention transformers. arXiv preprint arXiv:2403.12943, 2024. 3",
+            "[26] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. RAL, 5(2):3019-3026, 2020. 6"
+        ],
+        "bbox": [
+            516,
+            92,
+            906,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "6968",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[27] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In CoRL, pages 991-1002. PMLR, 2022. 3",
+            "[28] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 3",
+            "[29] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2(3):6, 2022. 2",
+            "[30] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 2",
+            "[31] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 5",
+            "[32] Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset. arXiv preprint arXiv:2403.12945, 2024. 2",
+            "[33] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2",
+            "[34] Po-Chen Ko, Jiayuan Mao, Yilun Du, Shao-Hua Sun, and Joshua B Tenenbaum. Learning to act from actionless videos through dense correspondences. In ICLR. 3",
+            "[35] Alexander Koch. Low-cost robot arm. https://github.com/AlexanderKoch-Koch/low_costrobot, 2024. GitHub repository. 5, 8",
+            "[36] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 2, 3, 5",
+            "[37] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. JMLR, 17(39):1-40, 2016. 1",
+            "[38] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. In ICLR. 2",
+            "[39] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Junwu Zhang, Munan Ning, and Li Yuan. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024. 3",
+            "[40] Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai,"
+        ],
+        "bbox": [
+            91,
+            92,
+            485,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Daya Guo, et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434, 2024. 3",
+            "[41] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. NeurIPS, 36, 2024. 5",
+            "[42] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024. 3",
+            "[43] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR, 2022. 1",
+            "[44] Corey Lynch, Mohi Khansari, Ted Xiao, Vikash Kumar, Jonathan Tompson, Sergey Levine, and Pierre Sermanet. Learning latent plans from play. In CoRL, pages 1113-1132. PMLR, 2020. 3",
+            "[45] Corey Lynch, Ayzaan Wahid, Jonathan Tompson, Tianli Ding, James Betker, Robert Baruch, Travis Armstrong, and Pete Florence. Interactive language: Talking to robots in real time. RAL, 2023. 3",
+            "[46] Soroush Nasiriany, Sean Kirmani, Tianli Ding, Laura Smith, Yuke Zhu, Danny Driess, Dorsa Sadigh, and Ted Xiao. Rt-affordance: Affordances are versatile intermediate representations for robot manipulation. arXiv preprint arXiv:2411.02704, 2024. 3",
+            "[47] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embediment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 2",
+            "[48] Joan Puigcerver, Carlos Riquelme, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. arXiv preprint arXiv:2308.00951, 2023. 3",
+            "[49] 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 ICLR, pages 8748-8763. PMLR, 2021. 1",
+            "[50] Rouhollah Rahmatizadeh, Pooya Abolghasemi, Ladislau Böloni, and Sergey Levine. Vision-based multi-task manipulation for inexpensive robots using end-to-end learning from demonstration. In ICRA, pages 3758-3765. IEEE, 2018. 3",
+            "[51] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022. 1, 2",
+            "[52] Younggyo Seo, Kimin Lee, Stephen L James, and Pieter Abbeel. Reinforcement learning with action-free pre-training from videos. In ICLR, pages 19561-19579. PMLR, 2022. 1",
+            "[53] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2"
+        ],
+        "bbox": [
+            516,
+            92,
+            906,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "6969",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[54] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, pages 894-906. PMLR, 2022. 3",
+            "[55] Austin Stone, Ted Xiao, Yao Lu, Keerthana Gopalakrishnan, Kuang-Huei Lee, Quan Vuong, Paul Wohlhart, Brianna Zitkovich, Fei Xia, Chelsea Finn, et al. Open-world object manipulation using pre-trained vision-language models. In CoRL, 2023. 3",
+            "[56] Priya Sundaresan, Quan Vuong, Jiayuan Gu, Peng Xu, Ted Xiao, Sean Kirmani, Tianhe Yu, Michael Stark, Ajinkya Jain, Karol Hausman, et al. Rt-sketch: Goal-conditioned imitation learning from hand-drawn sketches. 2024. 3",
+            "[57] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 2",
+            "[58] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1",
+            "[59] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. Robotap: Tracking arbitrary points for few-shot visual imitation. In ICRA, pages 5397-5403. IEEE, 2024. 1",
+            "[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In CoRL, pages 1723-1736. PMLR, 2023. 2",
+            "[61] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play. In CoRL. 1, 3",
+            "[62] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. CoRR, abs/2409.20537, 2024. 3",
+            "[63] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 1, 3, 6",
+            "[64] Kun Wu, Chengkai Hou, Jiaming Liu, Zhengping Che, Xiaozhu Ju, Zhuqin Yang, Meng Li, Yinuo Zhao, Zhiyuan Xu, Guang Yang, et al. Robomind: Benchmark on multi-embediment intelligence normative data for robot manipulation. arXiv preprint arXiv:2412.13877, 2024. 2",
+            "[65] Mengda Xu, Zhenjia Xu, Yinghao Xu, Cheng Chi, Gordon Wetzstein, Manuela Veloso, and Shuran Song. Flow as the cross-domain manipulation interface. arXiv preprint arXiv:2407.15208, 2024. 1, 3",
+            "[66] Jiange Yang, Wenhui Tan, Chuhao Jin, Bei Liu, Jianlong Fu, Ruihua Song, and Limin Wang. Pave the way to grasp anything: Transferring foundation models for universal pickplace robots. arXiv preprint arXiv:2306.05716, 1(2), 2023. 3",
+            "[67] Jiange Yang, Bei Liu, Jianlong Fu, Bocheng Pan, Gangshan Wu, and Limin Wang. Spatiotemporal predictive pre-training for robotic motor control. arXiv preprint arXiv:2403.05304, 2024. 3"
+        ],
+        "bbox": [
+            91,
+            90,
+            485,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[68] Jiange Yang, Wenhui Tan, Chuhao Jin, Keling Yao, Bei Liu, Jianlong Fu, Ruihua Song, Gangshan Wu, and Limin Wang. Transferring foundation models for generalizable robotic manipulation. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 1999-2010, 2025. 3",
+            "[69] Seonghyeon Ye, Joel Jang, Byeongguk Jeon, Sejune Joo, Jianwei Yang, Baolin Peng, Ajay Mandlekar, Reuben Tan, Yu-Wei Chao, Bill Yuchen Lin, et al. Latent action pretraining from videos. arXiv preprint arXiv:2410.11758, 2024. 1",
+            "[70] Chengbo Yuan, Chuan Wen, Tong Zhang, and Yang Gao. General flow as foundation affordance for scalable robot learning. arXiv preprint arXiv:2401.11439, 2024. 3",
+            "[71] Jinguo Zhu, Xizhou Zhu, Wenhai Wang, Xiaohua Wang, Hongsheng Li, Xiaogang Wang, and Jifeng Dai. Uniperceiver-moe: Learning sparse generalist models with conditional moes. NeurIPS, 35:2664-2678, 2022. 1",
+            "[72] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pre-training unified architecture for generic perception for zero-shot and few-shot tasks. In CVPR, pages 16804-16815, 2022. 1",
+            "[73] Yifeng Zhu, Zhenyu Jiang, Peter Stone, and Yuke Zhu. Learning generalizable manipulation policies with object-centric 3d representations. In CoRL. 3",
+            "[74] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 2, 3, 5"
+        ],
+        "bbox": [
+            514,
+            90,
+            906,
+            487
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "6970",
+        "bbox": [
+            482,
+            944,
+            516,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/d41e3854-0a29-4aee-85fc-0fcf3b667c89_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df984963d6de2c2be9fab1a77af5ad87c7c8ad512eacc9da9f441d4737f6bada
+size 1784290

2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/full.md

@@ -0,0 +1,323 @@
+# Tra-MoE: Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning
+
+Jiange Yang  $^{1,2}$  Haoyi Zhu  $^{2,3}$  Yating Wang  $^{2,4}$  Gangshan Wu  $^{1}$  Tong He  $^{2}$  Limin Wang  $^{1,2*}$
+
+$^{1}$ State Key Laboratory for Novel Software Technology, Nanjing University, China  $^{2}$ Shanghai AI Laboratory,  $^{3}$ University of Science and Technology of China,  $^{4}$ Tongji University
+
+jiangyang.jgy@gmail.com, {gswu,lmwang}@nju.edu.cn
+
+{zhuhaoyi, wangyating, hetong}@pjlab.org.cn
+
+https://github.com/MCG-NJU/Tra-MoE
+
+# Abstract
+
+Learning from multiple domains is a primary factor that influences the generalization of a single unified robot system. In this paper, we aim to learn the trajectory prediction model by using broad out-of-domain data to improve its performance and generalization ability. Trajectory model is designed to predict any-point trajectories in the current frame given an instruction and can provide detailed control guidance for robotic policy learning. To handle the diverse out-of-domain data distribution, we propose a sparsely-gated MoE (Top-1 gating strategy) architecture for trajectory model, coined as Tra-MoE. The sparse activation design enables good balance between parameter cooperation and specialization, effectively benefiting from large-scale out-of-domain data while maintaining constant FLOPs per token. In addition, we further introduce an adaptive policy conditioning technique by learning 2D mask representations for predicted trajectories, which is explicitly aligned with image observations to guide action prediction more flexibly. We perform extensive experiments on both simulation and real-world scenarios to verify the effectiveness of Tra-MoE and adaptive policy conditioning technique. We also conduct a comprehensive empirical study to train Tra-MoE, demonstrating that our Tra-MoE consistently exhibits superior performance compared to the dense baseline model, even when the latter is scaled to match Tra-MoE's parameter count.
+
+# 1. Introduction
+
+In computer vision and natural language processing, significant progress has been made through learning models from multiple domains data. For example, many works use board data to perform self-supervised pre-training, and then adapt to diverse downstream tasks through zero-shot [1, 10, 49] or fine-tuning [13, 15, 23, 58]. Additionally, some research ef
+
+![](images/b20601886bd5d635928fdf1cbebeaae57f139d3d6ba8796761e8f9210cfab061.jpg)  
+Figure 1. (a) Training with small-scale, in-domain data. (b) Joint training with in-domain data and large-scale, out-of-domain data.
+
+forts [43, 51, 71, 72] focus on learning a generalist model capable of unifying various tasks. In contrast, in robot learning, due to data scarcity and homogeneity, some groundbreaking methods [2, 37] resort to training only using in-domain data. Recently, some works [3, 4, 52, 61, 69] have shifted to learn from action-free human video data. A particularly scalable and promising paradigm [5, 20, 59, 63, 65] involves learning a trajectory prediction model from action-free video data, and then using small-scale action-labeled demonstrations to learn trajectory-guided policies, thereby achieving greater sample efficiency and better generalization.
+
+Despite the potential benefits, it still remains underexplored how to effectively leverage broad out-of-domain video data for jointly training trajectory models. In particular, the impact of incorporating data encompassing diverse environments, objects, skills and embodiments has not fully investigated for in-domain tasks. For instance, ATM [63], a notable work in this field, mainly relies on in-domain data (same tasks and environments) for training both trajectory
+
+![](images/1116d206222c344b3495caf32192dc759e91b4b7ea155d6cd6e4cf158685afb2.jpg)  
+Figure 2. Training trajectory prediction model from multiple domains. (a) The visualization of Dataset  $\mathcal{D}_{\text {ood }}$  and  $\mathcal{D}_{\text {in }}$ . The former may contain additional environments, objects, skills and embodiments. (b) Our pipeline: first co-training trajectory prediction model and then adapting it for downstream policy learning.
+
+![](images/dbc45401311683cf540e6205e3be6a9a5b51e719c61443d62b51f8072d7ecab6.jpg)
+
+models and policies. We argue that learning from multiple domains is the primary factor to build a powerful trajectory model for policy guidance. Therefore, we propose to train trajectory model under the setup of jointly learning from large-scale out-of-domain and small-scale in-domain data, as shown in Fig. 2(a). However, directly using large-scale out-of-domain video data for hybrid pre-training involves two critical challenges: (i) Joint training on data from different environments, objects, skills, and embodiments will add learning difficulty and might lead to optimization conflicts. It is unclear how to effectively fuse the complementary information of multiple domains data within a single unified model. Naively expanding the out-of-domain training data does not necessarily lead to improved performance on in-domain tasks, as shown in Fig. 1, which results in a 5.6  $(57.6\to 52.0)$  performance decrease on in-domain tasks. (ii) How to ensure both high performance and computational efficiency during trajectory model scaling up process.
+
+To address these challenges, we design a new sparsely-gated Mixture-of-Expert (MoE) architecture [53] to scale up our trajectory model, coined as Tra-MoE. Specifically, we integrate several MoE blocks to replace the original transformer blocks. This design enables efficient joint training of most parameters on the broad out-of-domain data, capturing the complementary patterns for mutual cooperation and the shared knowledge for better generalization. Meanwhile, for different data and different tokens within the same data, our Tra-MoE naturally forms a specialization when activating different experts. To maintain high computational efficiency while scaling up model capacity, we implement a top-1 gating strategy for token-choice, ensuring constant FLOPs per token. In sparse MoE, some auxiliary losses are commonly used to enhance performance, including the router z-loss [74] for improving training stability and the load-balancing loss [36] for balancing expert activations. We further conduct a comprehensive empirical study on scaling up trajectory models and find that the touter z-loss improves
+
+performance while the load-balancing loss tends to cause performance degradation. Furthermore, we observe that when expanding pre-training data from small-scale in-domain to large-scale in-domain and out-of-domain data, employing MoE (Tra-MoE) can lead to obvious performance improvement on in-domain tasks, whereas the dense counterpart (Tra-baseline) suffers from performance drop, as illustrated in Fig. 1. Notably, even when expanding the dense Tra-baseline's parameter to match that of Tra-MoE, performance improvements remain elusive. These findings collectively demonstrate the effectiveness of our sparsely-gated MoE architecture, attributing its capability of balancing parameter cooperation and specialization when expanding training with out-of-domain data.
+
+Furthermore, given a pre-trained trajectory model, how to effectively condition the policy module using the 2D trajectories still remain an unresolved challenge. We further propose an adaptive policy conditioning technique for trajectory-guided policy. This enables better spatial alignment between 2D trajectories and images, while achieving more flexible trajectory representation and guidance for the policy. Finally, we also conduct extensive real-world experiments to further strengthen our conclusions. In summary, our main contributions can be summarized as follows:
+
+- We propose a sparsely-gated MoE architecture to learn trajectory models, called as Tra-MoE, with large-scale, out-of-domain, and action-free video data, coupled with a comprehensive empirical study on its training techniques.  
+- Extensive simulation and real-world environments demonstrate that our Tra-MoE (with Top-1 gating) can more effectively benefit from large-scale out-of-domain data (such as cross-embediments data and different physics engine simulated data) compared to the dense baseline.  
+- We demonstrate that Tra-MoE still significantly outperform expanded dense counterpart with equivalent parameters, highlighting the effectiveness of sparse MoE architecture in leveraging out-of-domain data through better parameter cooperation and specialization.  
+- We propose an adaptive policy conditioning technique for trajectory-guided policy, which ensures adaptive 2D trajectory representation and explicit spatial alignment with the observations, thereby achieving superior performance.
+
+# 2. Related Work
+
+Learning from multiple domains in Robot Learning. Scaling laws [30] have fueled substantial advancements in various fields. In robot learning, many recent studies [7, 8, 29, 51] investigate how compute, model size, and training data quantity affect the model performance on robotic manipulation tasks. Based on this, recent works have focused on collecting larger-scale real-world robotic datasets [11, 19, 32, 60, 64], and have successfully trained several robotics foundational models [6, 9, 16, 33, 38, 47, 57,
+
+[62] that demonstrate superior transfer capabilities. It is worth noting that these works mainly employ pre-training and fine-tuning paradigm rather than joint co-training paradigm, while focusing on evaluating the combination generalization and neglecting the performance impact on in-domain tasks. Additionally, these works directly combine heterogeneous robotic data with different action and observation spaces for joint training, potentially leading to sub-optimal solutions. In contrast, we leverage sparsely-gated MoE to incorporate larger-scale, out-of-domain, and action-less video data. Furthermore, our experiments demonstrate the importance of MoE for expanding out-of-domain data due to its superior capacity for parameter cooperation and specialization.
+
+Mixture-of-Experts. The Mixture-of-Experts (MoE) framework [12] is built upon a simple yet effective concept: dividing a model into specialized components, or 'experts', each focusing on distinct tasks or data aspects. This paradigm allows for selective activation of relevant experts based on the input, thereby maintaining computational efficiency while leveraging a vast reservoir of specialized knowledge through increased model capacity. Recently, in computer vision, natural language processing, and multi-modal, there has been a surge of foundation models trained based on the MoE architecture, including Swin-Moe [24], MoE-LLaVA [39], DeepSeek-V2 [40], Mixtral-8x22B [28], and so on. MoE typically encompasses dense [48] and sparse [14] variants. Dense MoE activates all experts in each iteration, while sparse MoE activates only a subset, generally resulting in lower computational overhead. However, sparse MoE can suffer from expert load imbalance and training instability issues. Techniques such as adding noise, load balancing loss [36], and router z-loss [74] are often employed to mitigate these issues and improve performance. In our work, we conduct a comprehensive empirical study of these techniques when training our sparsely-gated MoE-based trajectory model.
+
+Representations and Techniques for Policy Conditioning. Policy conditioning equips a multi-task robot policy with the ability to modulate its behavior by incorporating task-specific guidance information. This information, which may take various forms, guides the policy's action generation, enabling a single model to effectively learn and perform diverse tasks. Classic policy conditioning representations often encompass task identifiers [50], natural language instructions [8, 22, 45, 54], goal images [44], and demonstration videos [25, 27]. These are typically integrated with visual observations within the policy through concatenation or attention mechanisms. Recent works have explored alternative policy conditioning representations to enhance generalization and sample efficiency. These include leveraging internet-trained models to obtain bounding box [55] and object mask [42, 66, 68, 73], as well as learning plan [4], affordance [3, 46], future observations [17, 18, 34, 67], 2D or 3D trajectory [5, 61, 63, 65, 70] from human video data,
+
+and also hand-drawn trajectory [21] or sketch [56]. Different from these approaches that directly utilize raw or manually crafted signals for integration with visual observations, we introduce an adaptive policy conditioning technique that explicitly maps 2D trajectories onto the visual observations in space and encodes them as learnable embeddings. These designs ensure adaptive 2D trajectory representation and explicit spatial alignment with the visual observations.
+
+# 3. Methodology
+
+In this section, we first describe our problem formulation in Sec. 3.1. Subsequently, we give an overview of our proposed framework, which is composed of a sparsely-gated MoE-based trajectory prediction model and a trajectory-guided policy model using the adaptive policy conditioning technique, detailed in Sec. 3.2. Finally, we describe our training process and loss functions in Sec. 3.3.
+
+# 3.1. Problem Formulation
+
+Following the setup of previous works [5, 63, 65], our approach encompasses a trajectory prediction model to conduct language-conditioned any-point trajectory prediction and a trajectory-guided policy model to predict executable robot actions. In particular, as depicted in Fig. 2, we consider two types of datasets for training, which are large-scale, out-of-domain, and action-free videos  $\mathcal{D}_{\text{ood}}$ , as well as small-scale, in-domain, and action-labeled demonstrations  $\mathcal{D}_{\text{in}}$ . For trajectory prediction model, we employ  $\mathcal{D}_{\text{ood}}$  and  $\mathcal{D}_{\text{in}}$  for joint pre-training, whereas for trajectory-guided policy model, we only use  $\mathcal{D}_{\text{in}}$  for training.
+
+# 3.2. Architecture
+
+ATM Revisited. In this paper, we use ATM [63] as the baseline for our method. The overall pipeline of ATM is illustrated on Fig. 2(b). It first learns a trajectory model from video data to perform language-conditioned trajectory prediction. The predicted trajectories is then used to learn a trajectory-guided policy using robot demonstrations. Specifically, as for trajectory model, its objective is to predict the future of any-point trajectories in a frame. Formally, given an image observation  $o_t$  at timestep  $t$ , any set of 2D query points simplified on the image frame  $\mathbf{p}_t = \{p_{t,k}\}_{k=1}^K$ , and a language instruction  $\ell$ , the trajectory model learns a mapping  $\mathbf{p}_{t:t+H} = \tau_\theta(o_t, \mathbf{p}_t, \ell)$  that predicts the coordinates of these query points in the camera frame for the next  $H$  time steps, where  $\mathbf{p}_{t:t+H} \in \mathbb{R}^{H \times K \times 2}$ . For the trajectory-guided policy, it leverages the current observation  $o_t$  and the predicted trajectory  $\mathbf{p}_{t:t+H}$  to predict the subsequent robot actions.
+
+Sparsely-gated MoE-based trajectory model (Tra-MoE). ATM proposes a track transformer to implement the trajectory model. In track transformer, language tokens, image tokens, and track tokens to be decoded involve global interactions within the transformer block. However, directly
+
+![](images/1b360506ebbfb7a53d46babead0df0d6744f34a203836d82d0e49690726e4618.jpg)  
+Figure 3. (a) The pipeline of our sparsely-gated MoE-based trajectory model (Tra-MoE). (b) The pipeline of our trajectory-guided policy using the adaptive policy conditioning technique. Mapping means concatenating the trajectory mask with image observations, while learnable refers to setting each point in the trajectory mask as a learnable embedding.
+
+![](images/d1ccaa97ebefa6d3fb7b4012c22df2c42abd6eb7a4331b344e1bc0685f8f71ae.jpg)
+
+processing diverse domains and multiple modalities data within a unified transformer may yield suboptimal results. Hence, we propose our Tra-MoE framework. To develop our sparsely-gated MoE-based trajectory model, as depicted in Fig. 3(a), we replace certain layers of the transformer with MoE blocks. In each MoE block, it incorporates multiple feedforward networks (FFNs), each designated as an expert, and utilizes a gating function to activate a selected subset (Top-  $K$ ) of these experts. In this paper, we set  $K$  to a fixed value of 1 to ensure that the FLOPs per token remain constant relative to the baseline. Formally, the gating network  $\mathcal{G}$ , parameterized by  $\Theta$  and typically consisting of a linear-softmax network, yields the output  $\mathcal{G}(\mathbf{x_s};\Theta)$ , where  $\mathbf{x_s}$  represents our token sequences input. We formulate the sparsely-gated token-choice mechanism as follows:
+
+$$
+\mathcal {G} \left(\mathbf {x} _ {\mathrm {s}}; \boldsymbol {\Theta}\right) _ {i} = \operatorname {s o f t m a x} (\operatorname {T o p - K} (g \left(\mathbf {x} _ {\mathrm {s}}; \boldsymbol {\Theta}\right), k)) _ {i}, \tag {1}
+$$
+
+$$
+\operatorname {T o p - K} (g \left(\mathbf {x} _ {\mathbf {s}}; \boldsymbol {\Theta}\right), k) _ {i} = \left\{ \begin{array}{l l} g (\mathbf {x}; \boldsymbol {\Theta}) _ {i}, & \text {c o n d ,} \\ - \infty , & \text {o t h .} \end{array} \right., \tag {2}
+$$
+
+$$
+\operatorname {c o n d}: \text {i f} g (\mathbf {x}; \boldsymbol {\Theta}) _ {i} \text {i s i n t o p - k e l e m e n t s o f} g (\mathbf {x}; \boldsymbol {\Theta}). \tag {3}
+$$
+
+More specifically,  $\operatorname{Top - K}(\cdot ,k)$  function retains only the top-  $k$  values in a vector, setting the rest to  $-\infty$ . Consequently, the output of the sparsely-gated MoE layer can be
+
+formulated as:
+
+$$
+\mathcal {F} _ {\text {s p a r s e}} ^ {\mathrm {M o E}} \left(\mathbf {x} _ {\mathbf {s}}; \boldsymbol {\Theta}, \left\{\mathbf {W} _ {i} \right\} _ {i = 1} ^ {K}\right) = \sum_ {i = 1} ^ {K} \mathcal {G} \left(\mathbf {x} _ {\mathbf {s}}; \boldsymbol {\Theta}\right) _ {i} f _ {i} \left(\mathbf {x} _ {\mathbf {s}}; \mathbf {W} _ {i}\right), \tag {4}
+$$
+
+where  $f_{i}$  represents the i-th expert, usually a linear-GELU-linear FFN layer, is parameterized by  $\mathbf{W}_i$ .
+
+Trajectory-guided policy using the adaptive policy conditioning technique. After completing the trajectory model learning, we develop our trajectory-guided policy through the adaptive policy conditioning technique based on the initial ATM, as shown in Fig. 3(b). In our policy model, independently encoded track tokens and image tokens undergo a early fusion within the fusion transformer. Subsequently, the fused features are further integrated with the raw trajectory data through a late fusion mechanism along the channel dimension, ensuring the preservation of original trajectory information for more accurate action prediction.
+
+Fundamentally, the policy model bridges the gap between the predicted 2D any-point trajectories and the executable 3D robot actions, functioning as an inverse dynamic model and eliminating the necessity of other task specification. However, how to effectively condition the policy using the predicted 2D trajectories remains an unresolved challenge. Based on our analysis, we present two key insights: (i) Explicit spatial alignment and fusion of the 2D trajectories with image observations turns out to be highly beneficial.
+
+It simplifies the mapping of 2D trajectories to image space and highlights motion-relevant regions, enabling the policy model to focus on important areas and reducing the learning difficulty. (ii) Trajectory positions exhibit different characteristics: starting points typically emphasize local motion, while endpoints focus more on global trend. Therefore, an adaptive trajectory representation and conditioning technique is expected to flexibly guide the policy by effectively capturing these variations.
+
+Based on the aforementioned analysis, we propose an adaptive policy conditioning technique to incorporate trajectory information into 2D images. Specifically, we construct an additional mask modality, and populate it with 2D trajectories based on their specific spatial locations, while setting the value of each point in the trajectory mask as a learnable embedding. Before encoding the image, we concatenate this mask modality with the image observations on the channel dimension, forming a tensor of dimensions  $H \times W \times 4$ , as shown in Fig. 3(b). To sum up, our adaptive policy conditioning technique yield an adaptive 2D trajectory representation, which could be explicitly aligned with the image observations to guide the policy prediction more flexibly.
+
+# 3.3. Training
+
+We argue that trajectory models could be shared by multiple domains data. Therefore, we first pre-train a generalizable trajectory model on  $\mathcal{D}_{\text{ood}} \cup \mathcal{D}_{\text{in}}$ . The  $\mathcal{D}_{\text{ood}}$  typically contains additional environments, objects, skills, and embodiments. Then we train the trajectory-guided policy model on  $\mathcal{D}_{\text{in}}$  based on the pre-trained trajectory model. In trajectory-guided policy training phase, we freeze the trajectory model. Pre-training sparsely-gated MoE-based trajectory model. Following ATM, we train the trajectory model using the ground truth tracks generated by CoTracker [31]. We use MSE to optimize the trajectory prediction loss  $\mathcal{L}_{\text{tra}}$ . Additionally, we also consider adding an auxiliary load-balancing loss [36] to balance the activation frequency among experts. Given a batch of queries  $\mathcal{B}$ , it contains B samples and each sample contains S tokens. We define  $\mathcal{Q}_i$  as the proportion of tokens distributed to expert  $i$ , and  $\mathcal{P}_i$  as the proportion of the gating probability assigned to expert  $i$ . The  $\mathcal{L}_{\text{lo-ba}}$  can be formulated as follows:
+
+$$
+\mathcal {Q} _ {i} = \frac {1}{S \cdot B} \sum_ {x _ {s} ^ {m} \in \mathcal {B}} \mathbb {1} \left\{\operatorname {a r g m a x} \mathcal {G} \left(\mathbf {x} _ {\mathbf {s}} ^ {\mathbf {m}}; \boldsymbol {\Theta}\right) = i \right\}, \tag {5}
+$$
+
+$$
+\mathcal {P} _ {i} = \frac {1}{S \cdot B} \sum_ {x _ {\mathrm {s}} ^ {m} \in \mathcal {B}} \mathcal {G} \left(\mathbf {x} _ {\mathrm {s}} ^ {\mathrm {m}}; \boldsymbol {\Theta}\right) _ {i}, \tag {6}
+$$
+
+$$
+\mathcal {L} _ {\mathrm {l o - b a}} = N \sum_ {i = 1} ^ {N} \mathcal {Q} _ {i} \mathcal {P} _ {i}. \tag {7}
+$$
+
+Additionally, ST-MoE [74] proposed using the router  $z$ -loss to penalize large logits entering the gating network,
+
+thereby improving training stability. We similarly applies it to the training of our trajectory model, as shown below:
+
+$$
+\mathcal {L} _ {z} (x) = \frac {1}{S} \sum_ {k = 1} ^ {S} \left(\log \sum_ {i = 1} ^ {N} e ^ {\mathbf {g} _ {i} ^ {(k)}}\right) ^ {2}, \tag {8}
+$$
+
+where  $S$  is the total number of tokens,  $N$  is the total number of experts, and  $\mathbf{g} \in \mathcal{R}^{S \times N}$  are the logits going into the router. In summary, the trajectory model's training loss is as follow:
+
+$$
+\mathcal {L} _ {\text {t o t a l}} = \lambda_ {\text {t r a}} \cdot \mathcal {L} _ {\text {t r a}} + \lambda_ {\text {l o - b a}} \cdot \mathcal {L} _ {\text {l o - b a}} + \lambda_ {z} \cdot \mathcal {L} _ {z}. \tag {9}
+$$
+
+Training trajectory-guided policy through adaptive policy conditioning technique. For trajectory-guided policy training, we employ the paradigm of behavior cloning. Specifically, we use the MSE between predicted actions and ground truths as the loss function.
+
+# 4. Experiments
+
+We perform evaluation experiments using the LIBERO benchmark [41] and a real low-cost dual-arm robot [35]. In experiments, we quantitatively find that the success rate of downstream manipulation tasks is generally positively correlated with the performance of the trajectory model. Detailed quantitative experiment results can be found in the appendix. Meanwhile, CoTracker [31] can provide relatively accurate trajectory annotations; therefore, we report the average success rate of downstream manipulation tasks as our metric. Our experiments aim to study the following questions:
+
+Q1: Does the techniques of router z-loss, load-balancing loss, and adding noise to gating logits enhance the training of sparsely-gated Tra-MoE?  
+Q2: Does sparsely-gated Tra-MoE, compared to the dense Tra-baseline, benefit more from large-scale out-of-domain data (such as cross-embediments and different physics engine data) and demonstrate superior scaling up capabilities?  
+Q3: When we scale up the dense Tra-baseline to match the parameter count of Tra-MoE, does the sparsely-gated Tra-MoE still maintain its advantage?  
+Q4: Is our proposed adaptive policy conditioning technique for trajectory-guided policy more effective than ATM?
+
+# 4.1. Simulation Experiments
+
+Experiment setup. The LIBERO [41] is divided into five categories: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, LIBERO-Long, and LIBERO-90. LIBERO-90 includes 90 tasks, while each of the other categories contains 10 tasks. Compared to other categories, LIBERO-90 presents a highly varied selection of object categories, layouts, environments, and skill goals. It is typically used for pre-training rather than joint co-training. Therefore, our main ablation experiments select 90 tasks from LIBERO-90 to construct
+
+<table><tr><td></td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td>0</td><td>62.0</td><td>73.0</td><td>71.0</td><td>21.5</td><td>56.9</td></tr><tr><td>5·10-4</td><td>69.5</td><td>73.5</td><td>68.0</td><td>22.5</td><td>58.4</td></tr><tr><td>1·10-4</td><td>62.5</td><td>81.0</td><td>73.5</td><td>28.5</td><td>61.4</td></tr><tr><td>1·10-5</td><td>60.0</td><td>65.5</td><td>68.5</td><td>27.0</td><td>55.3</td></tr><tr><td>1·10-6</td><td>60.0</td><td>79.0</td><td>53.0</td><td>31.5</td><td>55.9</td></tr></table>
+
+(a) The value of  $\lambda_z$  
+
+<table><tr><td></td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td>w/o. noise</td><td>62.0</td><td>73.0</td><td>71.0</td><td>22.5</td><td>56.9</td></tr><tr><td>w. noise</td><td>71.5</td><td>67.0</td><td>57.5</td><td>27.0</td><td>55.8</td></tr></table>
+
+(c) Adding a noise term to the gating logits.  
+
+<table><tr><td>Num.</td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td>1</td><td>49.5</td><td>67.0</td><td>56.5</td><td>35.0</td><td>52.0</td></tr><tr><td>2</td><td>59.0</td><td>72.5</td><td>65.5</td><td>27.0</td><td>56.0</td></tr><tr><td>3</td><td>64.0</td><td>68.0</td><td>71.0</td><td>17.5</td><td>55.1</td></tr><tr><td>4</td><td>62.0</td><td>73.0</td><td>71.0</td><td>21.5</td><td>56.9</td></tr></table>
+
+(e) The number of experts.  
+
+<table><tr><td></td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td>0</td><td>62.0</td><td>73.0</td><td>71.0</td><td>22.5</td><td>56.9</td></tr><tr><td>1·10-3</td><td>58.5</td><td>74.0</td><td>60.0</td><td>18.0</td><td>52.6</td></tr><tr><td>1·10-5</td><td>61.0</td><td>69.5</td><td>56.0</td><td>26.0</td><td>53.1</td></tr><tr><td>1·10-7</td><td>66.5</td><td>67.0</td><td>73.5</td><td>22.0</td><td>57.3</td></tr></table>
+
+(b) The value of  $\lambda_{lo - ba}$ .  
+
+<table><tr><td></td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td>MoE</td><td>62.5</td><td>81.0</td><td>73.5</td><td>28.5</td><td>61.4</td></tr><tr><td>Width</td><td>66.5</td><td>50.5</td><td>61.5</td><td>15.0</td><td>48.4</td></tr><tr><td>Depth</td><td>61.5</td><td>75.5</td><td>46.5</td><td>26.5</td><td>52.5</td></tr></table>
+
+(d) Scaling Tra-baseline from width and depth.  
+
+<table><tr><td>MoE</td><td>OOD</td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td></td><td></td><td>67.5</td><td>68.5</td><td>68.0</td><td>26.5</td><td>57.6</td></tr><tr><td>✓</td><td></td><td>63.0</td><td>64.0</td><td>60.5</td><td>19.5</td><td>51.8</td></tr><tr><td></td><td>✓</td><td>49.5</td><td>67.0</td><td>56.5</td><td>35.0</td><td>52.0</td></tr><tr><td>✓</td><td>✓</td><td>62.5</td><td>81.0</td><td>73.5</td><td>28.5</td><td>61.4</td></tr></table>
+
+dataset  $\mathcal{D}_{\text{ood}}$ , utilizing 20 action-free videos per task; and choose 40 tasks from LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long to build dataset  $\mathcal{D}_{\text{in}}$ , using 10 action-labeled demonstrations per task. In summary, for training Tra-MoE, our main ablation experiments use a total of 2, 200 action-free videos, with an out-of-domain to in-domain data ratio of  $9:2$ . Additionally, on this basis, we also further select 92 tasks from the RLbench [26] (using CoppeliaSim rather than MuJoCo as physics engine) simulation environment and generate 5 videos for each task, totaling 2, 660 videos in conjunction with the LIBERO data, to train Tra-MoE. The different physics engines and dynamics properties further increase the data distribution discrepancy in  $\mathcal{D}_{\text{ood}}$ . Finally, as for the trajectory-guided policy, we train a separate multi-task policy for each category. Each policy undergoes behavioral cloning training with 100 action-labeled demonstrations from 10 tasks.
+
+Implement details. Building on ATM [63], we develop our sparsely-gated Tra-MoE and adaptive policy conditioning technique. Specifically, when using only LIBERO data, we replace the 1th, 5th, and 8th layers of the track transformer with MoE blocks. When further incorporating RLbench data, we additionally replace the 2th and 7th layers. Unless otherwise specified, we set the number of experts to 4 by default and use a top-1 gating strategy for token choice to ensure the FLOPs per token remain constant.
+
+Experiment results and analysis. We present empirical simulation results to address the aforementioned research questions and make the following findings:
+
+Result 1: In Tab. 1a, we examine the effect of the router
+
+(f) Scaling with MoE and OOD data.  
+Table 1. The ablation experiments of the sparsely-gated Tra-MoE. All baselines are the track transformer of ATM.  
+
+<table><tr><td></td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td>ATM</td><td>51.0</td><td>42.0</td><td>55.5</td><td>19.0</td><td>41.9</td></tr><tr><td>Tra-MoE</td><td>54.0</td><td>62.0</td><td>61.0</td><td>41.0</td><td>54.5</td></tr></table>
+
+Table 2. The comparison results when training trajectory model using RLbench video data.
+
+z-loss weight  $\lambda_z$ . The results indicate that a higher  $\lambda_z$  can effectively enhance performance, with an improvement of 4.5 (61.4 vs. 56.9) at  $\lambda_z = 1 \cdot 10^{-4}$ . In Tab. 1b, we study the impact of the load-balancing loss weight  $\lambda_{lo - ba}$ . The results show a significant performance decline with larger values of  $\lambda_{lo - ba}$ , with only a slight improvement ( $56.7 \rightarrow 57.3$ ) at  $\lambda_{lo - ba} = 1 \cdot 10^{-7}$ . In Tab. 1c, we verify the effectiveness of adding noise to the gating logits, which results in a performance decrease of by 1.1 ( $56.9 \rightarrow 55.8$ ).
+
+Finding 1: Based on the above results, we find that using router z-loss to penalize large logits entering the gating network can effectively improve training stability and lead to performance gains. Meanwhile, the load-balancing loss used to balance the activation frequency among experts and the addition of noise to enhance expert allocation exploration cannot bring effective performance improvements. An intuitive explanation is that when the data distribution is uneven, the load-balancing loss might force experts to set shared parameters on data with a large domain gap, which could lead to mutual counteraction of learning gradients. This undermines the advantage of parameter specialization in sparsely-gated MoE. Similarly, adding noise is also detrimental to learning parameter specialization among experts.
+
+Result 2: We comprehensively investigate the impact of in
+
+<table><tr><td></td><td>Spatial</td><td>Goal</td><td>Object</td><td>Long</td><td>Avg.</td></tr><tr><td>ATM</td><td>62.5</td><td>81.0</td><td>73.5</td><td>28.5</td><td>61.4</td></tr><tr><td>+ DM</td><td>69.0</td><td>58.0</td><td>85.5</td><td>33.5</td><td>61.5</td></tr><tr><td>+ AM</td><td>69.5</td><td>77.0</td><td>88.0</td><td>30.5</td><td>66.3</td></tr></table>
+
+Table 3. The ablation experiments of the adaptive policy conditioning technique. DM (RT-Trajectory style) and AM respectively represent hand-drawn mask and adaptive mask.
+
+producing sparsely-gated MoE and large-scale out-of-domain data on scaling up trajectory models in Tab. 1f. Firstly, when training only with in-domain data, Tra-baseline achieves an average success rate of 57.6, while Tra-MoE only achieves an average success rate of 51.8. Secondly, when adding out-of-domain data to jointly train Tra-baseline, the average success rate drops to  $52.0$  ( $57.6 \rightarrow 52.0$ ), with particularly significant performance declines in LIBERO-Spatial and LIBERO-Object. However, when adding out-of-domain data to jointly train Tra-MoE, the average success rate increases to  $61.4$  ( $51.8 \rightarrow 61.4$ ). Additionally, in Tab. 1e, we also explore the impact of the number of experts in sparsely-gated MoE. The results show that using just two experts can lead to an increase of 4.0 (52.0 vs. 56.0) in the average success rate. With additional experts, performance generally trends upward, albeit with some fluctuations. Finally, we report the comparison results of training trajectory model by further using RLbench data in Tab. 2. The results indicate that although the introduction of RLbench data led to a performance drop in LIBERO, our Tra-MoE still outperform the baseline by 12.6 (54.5 vs. 41.9).
+
+Finding 2: Based on the above results, we have the following findings and analysis: (i) When the data scale is small, naively expanding the model capacity through sparsely-gated MoE does not lead to performance improvements. A explanation is that the larger model may lead to a overfitting problem when training on small-scale data. Moreover, small-scale data often leads to some experts not being sufficiently trained, thereby affecting performance. (ii) Naively introducing out-of-domain data for joint training often leads to performance degradation in the target domain. However, simultaneously incorporating out-of-domain data and expanding models with sparsely-gated MoE may improve performance or alleviate performance degradation. This phenomenon suggests that although out-of-domain data generally enhances model generalization, it could hinder task-specific performance in the target domain due to insufficient in-domain data learning. In contrast, expanding the model with out-of-domain data using sparsely-gated MoE achieves a more optimal balance between parameter cooperation and specialization across diverse data and even different tokens within the same data input. This design allows these large-scale data to jointly train most of parameters, ensuring the capture of general patterns for mutual cooperation. (iii) Even with a few experts, Tra-MoE exhibits significant scaling up
+
+capabilities. Additionally, further increasing the number of experts generally leads to performance improvement. However, this improvement is not guaranteed due to factors like data-expert complexity matching and training stability.
+
+Result 3: In Tab. 1d, we also investigate the performance comparison between the dense Tra-baseline and the sparse Tra-MoE with the same number of parameters. Specifically, we expand Tra-baseline to match the model capacity of Tra-MoE by increasing its model width or depth separately. These expanded versions achieve average success rates of 48.4 and 52.5 respectively, which are significantly lower than Tra-MoE's 61.4.
+
+Finding 3: Based on these results, we are surprised to find that simply expanding the dense Tra-baseline in depth and width can not effectively improve performance. This highlights the importance of the sparsely-gated MoE architecture when training with large-scale out-of-domain data, as it ensures dynamic activation of different experts based on different data input and different tokens within the same data input, significantly reducing optimization conflict problems.
+
+Result 4: In Tab. 3, we investigate the effectiveness of the adaptive policy conditioning. Specifically, building upon ATM, we first introduce an additional hand-drawn mask modality before image encoding. In this mask, the first half of the trajectory is set to 128, while the second half is set to 255. We call it 'ATM + hand-drawn mask' and this is similar to RT-Trajectory [21]. Next, we replace the hand-drawn mask with an adaptive mask, where each point on the trajectory is instantiated as a learnable embedding. We call it 'ATM + adaptive mask'. The results show that 'ATM + hand-drawn mask' achieves only a 0.1 improvement compared to the baseline. However, except for a significant performance decrease in LIBERO-Goal, it achieves performance improvements of 6.5, 12.0, and 5.0 in LIBERO-Spatial, LIBERO-Object, and LIBERO-Long respectively. In addition, the 'ATM + adaptive mask' achieves a performance improvement of 4.9 (66.3 vs. 61.4) compared to the baseline. It further improves performance across all suites compared to 'ATM + hand-drawn mask', especially achieving a 19.0 performance gains in LIBERO-Goal (77.0 vs. 58.0).
+
+Finding 4: Based on these results, we find that explicitly mapping 2D trajectories to visual observations significantly improves the performance of LIBERO-Spatial, LIBERO-Object, and LIBERO-Long, but causes a notable decrease in LIBERO-Goal. By instantiating the points in the 2D trajectory as learnable embeddings, we not only further enhances the performance of LIBERO-Spatial, LIBERO-Object, and LIBERO-Long, but also significantly improves LIBERO-Goal. We analyze that the decline in LIBERO-Goal is mainly due to overlapping front and back parts of some trajectories in the 2D images, leading to mapping errors. However, our adaptive trajectory representation and conditioning technique effectively mitigates this issue. In summary, these
+
+<table><tr><td>Human Data</td><td>MoE</td><td>Adaptive Trajectory Mask</td><td>Pour</td><td>Push</td><td>Pick and Pass</td><td>Tissue</td><td>Fold</td><td>Average</td></tr><tr><td></td><td></td><td></td><td>40.0</td><td>45.0</td><td>50.0</td><td>30.0</td><td>25.0</td><td>38.0</td></tr><tr><td>✓</td><td></td><td></td><td>45.0</td><td>35.0</td><td>50.0</td><td>25.0</td><td>35.0</td><td>38.0</td></tr><tr><td>✓</td><td>✓</td><td></td><td>40.0</td><td>50.0</td><td>60.0</td><td>35.0</td><td>45.0</td><td>46.0</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>60.0</td><td>70.0</td><td>65.0</td><td>35.0</td><td>50.0</td><td>56.0</td></tr></table>
+
+Table 4. The ablation of ood human data, MoE architecture, and adaptive policy conditioning technique on our real-world tasks.
+
+![](images/84503aacec77ac5c54356f89d97bcde412c1ed4dba170d2a1d766ebf2aa969d0.jpg)  
+Figure 4. Left: The real-world experiments hardware platform setup. Right: The real-world tasks evaluation demonstrations.
+
+experiments collectively demonstrate the effectiveness of explicitly aligning 2D trajectories with images in space as well as applying the adaptive conditioning technique.
+
+# 4.2. Real-World Experiments
+
+Experiment setup. As shown in Fig 4, we utilize a dual-arm robot [35] to conduct real-world experiments. We evaluate 5 tasks, which include two single-arm tasks: pouring water and pushing a vegetable to the side of the cutting board, as well as three dual-arm tasks: pulling out tissues, folding towels, and picking up and passing a holder. Specifically, for each task, we collect 50 action-labeled robot demonstrations. Concurrently, we also collect 50 videos of human (cross-embediments) performing for each task. Compared to the robot data, the human video encompasses more environment variations and distractor objects. Please see the appendix for more detailed evaluation details and real-world setup.
+
+Implement details. Following our simulation experiments, we utilize two camera views and uniformly resize their resolution to  $128 \times 128$  for trajectory model and policy training. We train our real-world Tra-MoE on borh human video and robot video data, with a separate policy trained for each task. Therefore, our real-world Tra-MoE integrates data across different embodiments (human, single-arm robot, and dual-arm robot). They have significantly different visual appearances and dynamics properties. Finally, we also present some real-world task evaluation demonstrations in Fig. 4.
+
+Experiment results and analysis. To further strengthen our conclusions that sparsely-gated MoE can better learn trajectory model from out-of-domain data and our adaptive policy
+
+condition technique could lead to more flexible guidance for policy learning, we conduct real-world ablation experiments, as shown in Tab. 4. The results indicate that simply introducing human video data cannot lead to performance improvement. However, when expanding the model by MoE architecture, the performance increases from  $38\%$  to  $46\%$ . We speculate this may be due to the significant differences in appearance and dynamics between human and robot, with the former typically moving much faster and more flexibly. Finally, when further incorporating an adaptive trajectory mask for policy conditioning, the performance is further increased from  $46\%$  to  $56\%$ .
+
+# 5. Conclusion
+
+In this work, we have introduced Tra-MoE, a novel framework that utilizes a sparsely-gated MoE architecture with Top-1 gating strategy to learn trajectory model from multiple domains. Tra-MoE achieves superior parameter cooperation and specialization, enabling more effective utilization of large-scale out-of-domain data while maintaining constant FLOPs per token. We train our Tra-MoE and conduct a comprehensive empirical study, demonstrating that TraMoE consistently exhibits superior performance compared to the dense baseline, even when the latter is scaled to match Tra-MoE's parameter count. Additionally, we propose an adaptive policy conditioning technique for trajectory-guided policy, ensuring adaptive 2D trajectory representation and explicit spatial alignment with observations, thereby achieving superior performance. Extensive experiments in both simulated and real-world environments validate our findings and demonstrate the effectiveness of our proposed methods.
+
+# Acknowledgements
+
+This work is supported by the National Key R&D Program of China (No. 2022ZD0160900), Jiangsu Frontier Technology Research and Development Program (No. BF2024076), and the Collaborative Innovation Center of Novel Software Technology and Industrialization.
+
+# References
+
+[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 preprint arXiv:2303.08774, 2023. 1  
+[2] OpenAI: Marcin Andrychowicz, Bowen Baker, Maciek Chogiej, Rafal Jozefowicz, Bob McGrew, Jakub Pachocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, et al. Learning dexterous in-hand manipulation. *IJRR*, 39(1): 3-20, 2020. 1  
+[3] Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. Affordances from human videos as a versatile representation for robotics. In CVPR, pages 13778-13790, 2023. 1, 3  
+[4] Homanga Bharadhwaj, Abhinav Gupta, Vikash Kumar, and Shubham Tulsiani. Towards generalizable zero-shot manipulation via translating human interaction plans. In ICRA, pages 6904-6911. IEEE, 2024. 1, 3  
+[5] Homanga Bharadhwaj, Roozbeh Mottaghi, Abhinav Gupta, and Shubham Tulsiani. Track2act: Predicting point tracks from internet videos enables diverse zero-shot robot manipulation. arXiv preprint arXiv:2405.01527, 2024. 1, 3  
+[6] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2  
+[7] Konstantinos Bousmalis, Giulia Vezzani, Dushyant Rao, Coline Manon Devin, Alex X Lee, Maria Bauza Villalonga, Todor Davchev, Yuxiang Zhou, Agrim Gupta, Akhil Raju, et al. Robocat: A self-improving generalist agent for robotic manipulation. TMLR, 2023. 2  
+[8] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2, 3  
+[9] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 2  
+[10] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1  
+[11] Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xu Huang, Shu Jiang, et al. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems. arXiv preprint arXiv:2503.06669, 2025. 2
+
+[12] Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. A survey on mixture of experts. arXiv preprint arXiv:2407.06204, 2024. 3  
+[13] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 1  
+[14] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. Mod-squad: Designing mixtures of experts as modular multitask learners. In CVPR, pages 11828-11837, 2023. 3  
+[15] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1  
+[16] Ria Doshi, Homer Rich Walke, Oier Mees, Sudeep Dasari, and Sergey Levine. Scaling cross-embodied learning: One policy for manipulation, navigation, locomotion and aviation. In CoRL, 2  
+[17] Yilun Du, Sherry Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, Leslie Pack Kaelbling, et al. Video language planning. In The Twelfth International Conference on Learning Representations. 3  
+[18] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. NeurIPS, 36, 2024. 3  
+[19] Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A robotic dataset for learning diverse skills in one-shot. In RSS 2023 Workshop on Learning for Task and Motion Planning, 2023.  
+[20] Chongkai Gao, Haozhuo Zhang, Zhixuan Xu, Cai Zhehao, and Lin Shao. Flip: Flow-centric generative planning as general-purpose manipulation world model. In ICLR. 1  
+[21] Jiayuan Gu, Sean Kirmani, Paul Wohlhart, Yao Lu, Montserrat Gonzalez Arenas, Kanishka Rao, Wenhao Yu, Chuyuan Fu, Keerthana Gopalakrishnan, Zhuo Xu, et al. Rt-trajectory: Robotic task generalization via hindsight trajectory sketches. In ICLR. 3, 7  
+[22] Pierre-Louis Guhur, Shizhe Chen, Ricardo Garcia Pinel, Makarand Tapaswi, Ivan Laptev, and Cordelia Schmid. Instruction-driven history-aware policies for robotic manipulations. In CORL, pages 175-187. PMLR, 2023. 3  
+[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1  
+[24] Changho Hwang, Wei Cui, Yifan Xiong, Ziyue Yang, Ze Liu, Han Hu, Zilong Wang, Rafael Salas, Jithin Jose, Prabhat Ram, et al. Tutel: Adaptive mixture-of-experts at scale. Proceedings of Machine Learning and Systems, 5:269-287, 2023. 3  
+[25] Vidhi Jain, Maria Attarian, Nikhil J Joshi, Ayzaan Wahid, Danny Driess, Quan Vuong, Pannag R Sanketi, Pierre Sermanet, Stefan Welker, Christine Chan, et al. Vid2robot: End-to-end video-conditioned policy learning with cross-attention transformers. arXiv preprint arXiv:2403.12943, 2024. 3  
+[26] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. RAL, 5(2):3019-3026, 2020. 6
+
+[27] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In CoRL, pages 991-1002. PMLR, 2022. 3  
+[28] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 3  
+[29] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2(3):6, 2022. 2  
+[30] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 2  
+[31] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 5  
+[32] Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset. arXiv preprint arXiv:2403.12945, 2024. 2  
+[33] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2  
+[34] Po-Chen Ko, Jiayuan Mao, Yilun Du, Shao-Hua Sun, and Joshua B Tenenbaum. Learning to act from actionless videos through dense correspondences. In ICLR. 3  
+[35] Alexander Koch. Low-cost robot arm. https://github.com/AlexanderKoch-Koch/low_costrobot, 2024. GitHub repository. 5, 8  
+[36] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 2, 3, 5  
+[37] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. JMLR, 17(39):1-40, 2016. 1  
+[38] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. In ICLR. 2  
+[39] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Junwu Zhang, Munan Ning, and Li Yuan. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024. 3  
+[40] Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai,
+
+Daya Guo, et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434, 2024. 3  
+[41] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. NeurIPS, 36, 2024. 5  
+[42] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024. 3  
+[43] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR, 2022. 1  
+[44] Corey Lynch, Mohi Khansari, Ted Xiao, Vikash Kumar, Jonathan Tompson, Sergey Levine, and Pierre Sermanet. Learning latent plans from play. In CoRL, pages 1113-1132. PMLR, 2020. 3  
+[45] Corey Lynch, Ayzaan Wahid, Jonathan Tompson, Tianli Ding, James Betker, Robert Baruch, Travis Armstrong, and Pete Florence. Interactive language: Talking to robots in real time. RAL, 2023. 3  
+[46] Soroush Nasiriany, Sean Kirmani, Tianli Ding, Laura Smith, Yuke Zhu, Danny Driess, Dorsa Sadigh, and Ted Xiao. Rt-affordance: Affordances are versatile intermediate representations for robot manipulation. arXiv preprint arXiv:2411.02704, 2024. 3  
+[47] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embediment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 2  
+[48] Joan Puigcerver, Carlos Riquelme, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. arXiv preprint arXiv:2308.00951, 2023. 3  
+[49] 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 ICLR, pages 8748-8763. PMLR, 2021. 1  
+[50] Rouhollah Rahmatizadeh, Pooya Abolghasemi, Ladislau Böloni, and Sergey Levine. Vision-based multi-task manipulation for inexpensive robots using end-to-end learning from demonstration. In ICRA, pages 3758-3765. IEEE, 2018. 3  
+[51] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022. 1, 2  
+[52] Younggyo Seo, Kimin Lee, Stephen L James, and Pieter Abbeel. Reinforcement learning with action-free pre-training from videos. In ICLR, pages 19561-19579. PMLR, 2022. 1  
+[53] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2
+
+[54] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, pages 894-906. PMLR, 2022. 3  
+[55] Austin Stone, Ted Xiao, Yao Lu, Keerthana Gopalakrishnan, Kuang-Huei Lee, Quan Vuong, Paul Wohlhart, Brianna Zitkovich, Fei Xia, Chelsea Finn, et al. Open-world object manipulation using pre-trained vision-language models. In CoRL, 2023. 3  
+[56] Priya Sundaresan, Quan Vuong, Jiayuan Gu, Peng Xu, Ted Xiao, Sean Kirmani, Tianhe Yu, Michael Stark, Ajinkya Jain, Karol Hausman, et al. Rt-sketch: Goal-conditioned imitation learning from hand-drawn sketches. 2024. 3  
+[57] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 2  
+[58] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1  
+[59] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. Robotap: Tracking arbitrary points for few-shot visual imitation. In ICRA, pages 5397-5403. IEEE, 2024. 1  
+[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In CoRL, pages 1723-1736. PMLR, 2023. 2  
+[61] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play. In CoRL. 1, 3  
+[62] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. CoRR, abs/2409.20537, 2024. 3  
+[63] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 1, 3, 6  
+[64] Kun Wu, Chengkai Hou, Jiaming Liu, Zhengping Che, Xiaozhu Ju, Zhuqin Yang, Meng Li, Yinuo Zhao, Zhiyuan Xu, Guang Yang, et al. Robomind: Benchmark on multi-embediment intelligence normative data for robot manipulation. arXiv preprint arXiv:2412.13877, 2024. 2  
+[65] Mengda Xu, Zhenjia Xu, Yinghao Xu, Cheng Chi, Gordon Wetzstein, Manuela Veloso, and Shuran Song. Flow as the cross-domain manipulation interface. arXiv preprint arXiv:2407.15208, 2024. 1, 3  
+[66] Jiange Yang, Wenhui Tan, Chuhao Jin, Bei Liu, Jianlong Fu, Ruihua Song, and Limin Wang. Pave the way to grasp anything: Transferring foundation models for universal pickplace robots. arXiv preprint arXiv:2306.05716, 1(2), 2023. 3  
+[67] Jiange Yang, Bei Liu, Jianlong Fu, Bocheng Pan, Gangshan Wu, and Limin Wang. Spatiotemporal predictive pre-training for robotic motor control. arXiv preprint arXiv:2403.05304, 2024. 3
+
+[68] Jiange Yang, Wenhui Tan, Chuhao Jin, Keling Yao, Bei Liu, Jianlong Fu, Ruihua Song, Gangshan Wu, and Limin Wang. Transferring foundation models for generalizable robotic manipulation. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 1999-2010, 2025. 3  
+[69] Seonghyeon Ye, Joel Jang, Byeongguk Jeon, Sejune Joo, Jianwei Yang, Baolin Peng, Ajay Mandlekar, Reuben Tan, Yu-Wei Chao, Bill Yuchen Lin, et al. Latent action pretraining from videos. arXiv preprint arXiv:2410.11758, 2024. 1  
+[70] Chengbo Yuan, Chuan Wen, Tong Zhang, and Yang Gao. General flow as foundation affordance for scalable robot learning. arXiv preprint arXiv:2401.11439, 2024. 3  
+[71] Jinguo Zhu, Xizhou Zhu, Wenhai Wang, Xiaohua Wang, Hongsheng Li, Xiaogang Wang, and Jifeng Dai. Uniperceiver-moe: Learning sparse generalist models with conditional moes. NeurIPS, 35:2664-2678, 2022. 1  
+[72] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pre-training unified architecture for generic perception for zero-shot and few-shot tasks. In CVPR, pages 16804-16815, 2022. 1  
+[73] Yifeng Zhu, Zhenyu Jiang, Peter Stone, and Yuke Zhu. Learning generalizable manipulation policies with object-centric 3d representations. In CoRL. 3  
+[74] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 2, 3, 5

2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ca36ea14986ad168fc5cfe0fc3fbb7316bf23bcd149f97ce0081ecf4a754d56
+size 460906

2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/0f12efc0-153e-4cda-95dd-b16491feb48e_content_list.json

@@ -0,0 +1,1335 @@
+[
+    {
+        "type": "text",
+        "text": "TraF-Align: Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception",
+        "text_level": 1,
+        "bbox": [
+            156,
+            130,
+            843,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zhiying Song $^{1}$ , Lei Yang $^{1}$ , Fuxi Wen $^{1,2,\\dagger}$ , Jun Li $^{1,2}$ $^{1}$ School of Vehicle and Mobility, Tsinghua University †Corresponding author  \n $^{2}$ State Key Lab of Intelligent Green Vehicle and Mobility  \n{song-zy24,yanglei20}@ mails.tsinghua.edu.cn, {wenfuxi, lijun1958}@tsinghua.edu.cn",
+        "bbox": [
+            143,
+            202,
+            849,
+            276
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            308,
+            326,
+            325
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Cooperative perception presents significant potential for enhancing the sensing capabilities of individual vehicles, however, inter-agent latency remains a critical challenge. Latencies cause misalignments in both spatial and semantic features, complicating the fusion of real-time observations from the ego vehicle with delayed data from others. To address these issues, we propose TraF-Align, a novel framework that learns the flow path of features by predicting the feature-level trajectory of objects from past observations up to the ego vehicle's current time. By generating temporally ordered sampling points along these paths, TraF-Align directs attention from the current-time query to relevant historical features along each trajectory, supporting the reconstruction of current-time features and promoting semantic interaction across multiple frames. This approach corrects spatial misalignment and ensures semantic consistency across agents, effectively compensating for motion and achieving coherent feature fusion. Experiments on two real-world datasets, V2V4Real and DAIR-V2X-Seq, show that TraF-Align sets a new benchmark for asynchronous cooperative perception. The code is available at https://github.com/zhyingS/TraF-Align.",
+        "bbox": [
+            88,
+            342,
+            483,
+            675
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            89,
+            705,
+            220,
+            722
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Intelligent vehicles rely on advanced onboard sensors, such as LiDAR and cameras, to perceive and understand the surrounding world. Nevertheless, the limited field of view and detection range of onboard sensors restrict their ability to perceive objects in occluded areas and at greater distances [2, 23, 36, 47]. Vehicle-to-everything (V2X) communication-enabled cooperative perception systems bridge this gap by exchanging perception information between vehicles and other agents, thereby extending the perception capability of a single agent [3, 4, 20, 41, 45].",
+        "bbox": [
+            89,
+            732,
+            482,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In real-world scenarios, cooperative perception systems",
+        "bbox": [
+            109,
+            885,
+            482,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/f3d4dfeee0aaaa6a9eee9851008ffb564a0a3cc14d5fed9de5113269dbcf1eaf.jpg",
+        "image_caption": [
+            "Figure 1. Asynchronous observations of ego vehicle and other agents result in spatial and semantic misalignments."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            529,
+            309,
+            901,
+            506
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "are cursed with inter-agent time delays [33, 35]. Consider a network with an ego agent and agent  $i$ , where their observations of the same object are shown in Figure 1. Latency in agent  $i$ 's message, captured before ego's observation time, can result in either a complete loss of features (1-I) or feature misalignment relative to ego's observation (1-II). Corresponding to Figure 1-I and 1-II, there are two main challenges: i) Spatial misalignment, where observations from other agents fail to align with the ground truth in space, lacking any reference from the ego vehicle's observations. ii) Semantic misalignment, where both the ego vehicle and another agent observe the same object but fail to recognize it as the same entity. These misalignments lead to misplaced objects and increased false positives [29, 30].",
+        "bbox": [
+            511,
+            579,
+            906,
+            792
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address the first challenge, some studies have proposed potential solutions, by predicting the current-frame features using historical frames [15], or predicting the pixel-level flow of features and then extrapolating them [44] or directly warping them to their future positions [34]. For instance, FFNet [44] uses a self-supervised approach to learn the first-order derivatives of features between consecutive",
+        "bbox": [
+            511,
+            795,
+            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": "12048",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "frames, and CoBEVFlow [34] generates an instance-level region of interest (ROI) and calculates feature motion vectors by matching and tracking these regions. However, predicting the movement of high-dimensional neural features is non-trivial. Despite their efforts, these methods struggle with larger delays and lack end-to-end capability. For the second challenge, the significance of semantic misalignment has not been well investigated in previous work.",
+        "bbox": [
+            89,
+            90,
+            480,
+            212
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address these misalignments, our core idea is to reconstruct current-time features and achieve semantic consensus across agents on multi-frame LiDAR features within a unified framework. We guide the flow and interaction of high-dimensional features by learning the motion of low-dimensional trajectory features. A trajectory field is defined as a spatiotemporal map of each object's anticipated path, guiding the generation of temporally ordered sampling points along each trajectory. Cross-attention between each query and its aligned sampling points allows the query to access relevant past features along the trajectory, enabling effective feature reconstruction and semantic interaction for robust spatial alignment and inter-agent consensus. Consider the spatial misalignment scenario shown in Figure 1-I. By aggregating multi-frame point cloud features from agent  $i$ , we learn the object's past trajectory distribution and predict its motion direction up to the ego vehicle's current time, forming a trajectory field. This field guides attention from the predicted position to historical features along the same trajectory, enabling the reconstruction of the object's current features near the ground truth and compensating for motion. For semantic misalignment cases (Figure 1-II), while delays may misalign agent  $i$ 's observations with the ego vehicle's, the continuity of object motion implies overlap with point clouds from past ego frames. By learning the trajectory distribution from the ego's multi-frame point clouds and applying attention to compute feature correlations along the trajectory, the model recognizes that all features along the trajectory correspond to the same object, thus achieving cross-agent semantic alignment.",
+        "bbox": [
+            91,
+            217,
+            486,
+            670
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Building on the abovementioned idea, we propose TraFAlign, a novel framework for asynchronous multi-agent cooperative perception (as shown in Figure 2). The proposed method includes three key parts: i) a field predictor that generates synchronized trajectory fields aligned with the ego vehicle's time, ii) an offset generator that produces sampling points for each query position and distributes them along the corresponding trajectory, and iii) attention layers where each query position is treated as a token, with attention focused only on positions specified by the offset generator. Extensive experiments are conducted on two widely used large-scale real-world datasets, V2V4Real [38] and DAIR-V2X-Seq [43], one for vehicle-to-vehicle (V2V) cooperative perception, another for vehicle-to-infrastructure (V2I) applications. Experimental results show that state-of",
+        "bbox": [
+            89,
+            674,
+            482,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "the-art (SOTA) performance is achieved for TraF-Align in both object detection precision and robustness to inter-agent latency. The contributions are summarized as follows:",
+        "bbox": [
+            511,
+            90,
+            903,
+            136
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose TraF-Align, a novel framework for asynchronous cooperative perception, addressing spatial and semantic misalignments caused by time delays.",
+            "- We leverage a field predictor and an offset generator to align trajectory fields and distribute attention points along object trajectories, enabling robust feature reconstruction and interaction along spatiotemporally aligned paths.",
+            "- We introduce a field loss and an offset loss to supervise trajectory alignment and the generation of attention points, providing an end-to-end trainable solution."
+        ],
+        "bbox": [
+            511,
+            136,
+            903,
+            287
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related work",
+        "text_level": 1,
+        "bbox": [
+            513,
+            301,
+            650,
+            316
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Asynchronous V2X cooperative perception. V2X cooperative perception extends the sensing range of single vehicles by exchanging information with other agents via wireless communication [13]. Previous studies have conducted extensive research under time-synchronized conditions, such as F-Cooper [3], V2VNet [33], V2X-ViT [35, 39], ERMVP [45], etc. In the presence of latency, previous research emphasizes using historical data to estimate and compensate for object motion. For instance, at the output level, the object state filters are designed to predict and update the dynamic state of the received objects [1]. Using raw data, Zhang et al. [46] and Dao et al. [8] estimated the motion of point clouds and predicted compensatory flows for late-arriving data. However, these levels may encounter challenges, such as short prediction time horizons or high communication data volumes. At the feature level, which has a better balance between accuracy and transmission bandwidth [18], SyncNet [15] employs a pyramidal LSTM to iteratively process and learn from delayed features, enabling the prediction of current-frame features. FFNet [44] predicts inter-frame feature changes in a self-supervised feature manner to synchronize features across agents. Nevertheless, predicting high-dimensional features is challenging and prone to uncontrollable variations. Additionally, CoBEVFlow [34] extracts and warps object-specific features to anticipated locations by predicting instance-wise motion vectors. However, this approach requires generating ROIs and tracking their motion, resulting in a non-end-to-end two-stage method. This paper uses multi-frame LiDAR sequences to predict low-dimensional trajectory features and guide attention to fuse features on past trajectories. By operating directly at the feature level, our approach enables end-to-end training and avoids the complexities of high-dimensional feature prediction.",
+        "bbox": [
+            511,
+            325,
+            906,
+            840
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Learning from LiDAR sequences across agents. Multiframe point cloud fusion is widely applied in single-vehicle autonomous systems to amalgamate semantic data across different perspectives [5, 21, 22, 24, 32, 42]. However, its",
+        "bbox": [
+            511,
+            840,
+            903,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "12049",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/2cdf8216f524cb86daeb1fdb5422271a921dc7b3ef7ca977ca25426faa07647e.jpg",
+        "image_caption": [
+            "Figure 2. Proposed Architecture. With asynchronous LiDAR inputs from the ego agent (time  $t$ ) and agent  $i$  (time  $s = t - \\tau_t^i$ ), BEV features are extracted using an onboard sparse encoder. Agent  $i$ 's features are transmitted to the ego agent and stored in memory. TraFAlign predicts a trajectory field up to time  $t$ , generates target attention positions, and uses features at these positions as keys/values in attention layers. Finally, the multi-agent features are fused and processed by the head to generate cooperative predictions at time  $t$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            101,
+            85,
+            897,
+            361
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "application in V2X cooperative perception remains under-explored. Although studies like SCOPE [40] and MRC-Net [11] have leveraged recurrent neural networks to fuse sequences from ego vehicles, these methods are resource-intensive and limited by sequence length. This has restricted their use to ego vehicles rather than a broader agent network, necessitating a more straightforward multi-frame fusion network. In cooperative perception scenarios, multiframe point clouds offer crucial motion cues that enable agents to synchronize their understanding of object positions despite time delays, thereby correlating trajectories across agents. This aspect has yet to be fully explored at the feature level. This paper proposes a novel approach using trajectory fields and customized attention with offsets for direct multi-frame fusion, thus establishing a consistent understanding of object motion across the network.",
+        "bbox": [
+            88,
+            454,
+            485,
+            698
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            89,
+            714,
+            183,
+            729
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Problem formulation",
+        "text_level": 1,
+        "bbox": [
+            89,
+            741,
+            290,
+            755
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Consider a group of networked agents indexed by  $\\mathcal{I} = \\{e, 1, 2, \\dots, N - 1\\}$ , each is equipped with LiDAR and communication units. Index  $e$  indicates the ego agent, which integrates information from other agents to extend its sensing range and reduce the blind spots. The LiDAR observation of the  $i$ th agent at its local timestamp  $s$  is denoted as  $L_s^i$ , where  $i \\in \\mathcal{I}$ . A shared neural network  $f(\\cdot)$  processes this observation into deep features. At each timestamp  $s$ , agent  $i$  transmits  $F_s^i = f(L_s^i)$  to the ego agent,",
+        "bbox": [
+            89,
+            763,
+            483,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "who then accumulates these features in the cache, forming a historical feature set for agent  $i$  at time  $s$ , represented as  $\\mathcal{F}_s^i = \\{F_{s - k}^i\\}_{k\\in K}$ , where  $K = \\{k\\in \\mathbb{Z}|1\\leqslant k\\leqslant m_i\\}$ ,  $m_{i}$  is the maximal sequence length for agent  $i$ . Accounting for delays in the collaborative process, messages from various agents arrive at ego at different times. Let  $\\tau_t^i$  be the delay for agent  $i$ 's message, and this means the information sent by agent  $i$  at time  $t - \\tau_t^i$  is received by the ego at time  $t$ . If ignoring feature extraction delay, then  $s = t - \\tau_t^i$ . The task is to fuse these multi-agent features and estimate the state of objects, i.e., the bounding boxes, at ego timestamp  $t$  within the detection range. The information available for fusion includes: i) the ego's historical features  $\\mathcal{F}_t^e$ , and ii) the historical features from other agents, denoted as  $\\{\\mathcal{F}_{t - \\tau_t^i}^i\\}_{i\\in \\mathcal{T}\\backslash \\{e\\}}$ . The proposed method is shown in Figure 2.",
+        "bbox": [
+            511,
+            454,
+            908,
+            684
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Feature extraction",
+        "text_level": 1,
+        "bbox": [
+            511,
+            696,
+            692,
+            710
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The encoder converts single-frame raw point cloud data into BEV features. Similar to PointPillars [14], the point cloud is first discretized into a regular grid on the  $x - y$  plane. Within each pillar, the point features are encapsulated as the vector  $[x,y,z,x_c,y_c,z_c,x_p,y_p,z_p]\\in \\mathbb{R}^9$ . Here,  $[x,y,z]$  is the point's position, while  $[x_c,y_c,z_c]$  indicates the center of the pillar. The vector  $[x_p,y_p,z_p]$  represents the point's offset from the pillar center. The pillar feature extraction and backbone network architectures are adopted from PillarNext [16]. Multi-Layer Perceptron (MLP) serves to encode point-wise features, while a sparse variant of ResNet [10] is utilized as the backbone to distill deeper and more",
+        "bbox": [
+            511,
+            719,
+            908,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "12050",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "complex features with size  $(C, H, W)$ . The sparse network is implemented using Spatially Sparse Convolution (SpConv) library [6]. The extracted single-frame features are transmitted and stored in the ego's feature bank  $(\\mathcal{F}_t^e$  and  $\\mathcal{F}_s^i)$ , which maintains a fixed capacity by admitting new features and discarding the oldest ones, ensuring an up-to-date repository for feature reuse. The extracted features of each agent are then processed using TraF-Align separately.",
+        "bbox": [
+            89,
+            90,
+            480,
+            210
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Temporal Encoding. Temporal information is decoupled from the point features to facilitate the reuse of historical features across frames. This allows for the immediate retrieval of historical features at any timestamp. Inspired by the positional embedding in Transformer [31], a temporal embedding  $TE \\in \\mathbb{R}^{C \\times H \\times W}$  is defined as",
+        "bbox": [
+            89,
+            212,
+            480,
+            301
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nT E _ {(\\tau , 2 j)} = \\sin (\\xi), T E _ {(\\tau , 2 j + 1)} = \\cos (\\xi), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            143,
+            313,
+            480,
+            330
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\xi = \\tau / \\varepsilon^{2j/C}$ ,  $j$  indexes the dimension along  $C$  channels,  $\\tau$  represents the number of frames delayed from the current frame, and  $\\varepsilon$  is set to eight in this study. Each feature within the memory cache,  $\\mathcal{F}_t^e$  for the ego agent and  $\\mathcal{F}_s^i$  for agent  $i$ , associated with a specific temporal delay, is concatenated with a corresponding temporal embedding. These concatenated features are refined through a convolutional layer with a  $1 \\times 1$  kernel and then warped to time  $t$  by ego-motion compensation.",
+        "bbox": [
+            89,
+            340,
+            482,
+            477
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Trajectory field",
+        "text_level": 1,
+        "bbox": [
+            89,
+            486,
+            246,
+            502
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Each agent processes its aggregated multi-frame feature map via a field predictor, which generates a trajectory field consisting of position and orientation fields. The position field creates heatmap peaks along the object's trajectory, while the orientation field captures the inverse tangent direction of the trajectory. Due to delays, agents' local observations may be outdated. To achieve temporal alignment, the field predictor predicts the trajectory features of all objects up to the ego vehicle's latest time (with temporal information encoded in the features via Equation 1). This ensures that, for the same object, trajectory fields from different viewpoints spatially overlap and are aligned temporally to the ego vehicle's current time. Our model uses a variant of UNet [25] as the field predictor, whose architecture is detailed in the supplementary material, along with a supervision signal to guide the learning process.",
+        "bbox": [
+            88,
+            508,
+            482,
+            750
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 3, when generating the ground truth field, we first extract bounding box annotations of each object from the local perspective of each agent within the maximum temporal window, starting from the object's initial appearance up to the ego's current time. A trajectory is then constructed by connecting the centers of these bounding boxes. Specifically, the length of the object's trajectory is given by  $\\lceil \\tau_i \\cdot \\omega_i \\rceil + m_i$ ,  $\\forall i \\in \\mathcal{I}$ , where  $\\lceil \\cdot \\rceil$  denotes the ceiling function,  $\\omega_i$  is the LiDAR sampling frequency  $(\\mathrm{s}^{-1})$  of agent  $i$ , and  $\\tau_i$  represents its cooperative latency (s). The",
+        "bbox": [
+            89,
+            750,
+            482,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/0af345f65626421913b5a14c7b1c3fb65639dfffa01dae3308eeaa52744f6a10.jpg",
+        "image_caption": [
+            "Figure 3. Illustration of trajectory field, which includes a position field indicating the occupancy grid of the trajectory and a direction field depicting the trajectory's flow orientation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            93,
+            898,
+            340
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "trajectory field is then generated by interpolating along the trajectory and projecting it onto the BEV feature space. If two object trajectories overlap, the trajectory with the more recent timestamp will overwrite the older one.",
+        "bbox": [
+            511,
+            422,
+            903,
+            482
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For cases in Figure 1-I, such a design compels the network to predict the trajectory field up to the present moment, thereby supporting the learning of motion information. Failing to do so would result in the network detecting only the delayed bounding box. Compared to the direct forecasting of high-dimensional features from the backbone network, predicting the trajectory field provides a more straightforward and more reliable method with predictable outcomes.",
+        "bbox": [
+            511,
+            484,
+            905,
+            619
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4. Trajectory-aware attention",
+        "text_level": 1,
+        "bbox": [
+            511,
+            635,
+            756,
+            650
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We flatten the input feature map  $F^i (\\forall i \\in \\mathcal{I})$  of size  $C \\times H \\times W$  into  $H \\times W$  embeddings, each serving as a query  $q \\in \\mathbb{R}^{1 \\times C}$ . For each query, the offset generator creates a response set  $R \\in \\mathbb{R}^{n \\times C}$ , capturing features at designated attention positions that are learned based on predefined ground truth.  $n$  is the number of attention positions for each query, and it's set to 18 in this paper. To generate the ground truth offsets, we select  $n$  pixels on the feature map that satisfy three conditions: i) The position field has a response (indicating the presence of an object), ii) The pixels belong to the same trajectory as the query (indicating the same object), and iii) The pixels correspond to earlier timestamps than the query (indicating historical positions of the object). As shown in Figure 4b, this ground truth aligns each query with preceding trajectory positions, enabling it to draw in relevant historical information. The",
+        "bbox": [
+            511,
+            657,
+            906,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "12051",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "trajectory field has effectively modeled the motion and flow of the historical features, offering valuable guidance to the offset generator. With this rich information, the generator can rely on simple convolution layers to produce good attention offsets, and we use PReLU activation [9] to provide flexibility in adjusting the offsets' directional orientation. Given the response set  $R$ , the feature of query  $q$  is updated or reconstructed by the attention process:",
+        "bbox": [
+            89,
+            90,
+            483,
+            212
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\text {A t t e n t i o n} (q, R) = \\text {S o f t m a x} \\left[ \\frac {q W ^ {q} \\left(R W ^ {k}\\right) ^ {T}}{\\sqrt {d}} \\right] R W ^ {v}, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            99,
+            219,
+            483,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $W^{q}, W^{k}, W^{v} \\in \\mathbb{R}^{C \\times d}$  are the query, key and value projection matrices. Similar to the original Transformer [31], we also incorporate feed-forward networks, addition and normalization layers, and a multi-head attention design, the details are shown in Figure 4a. Using attention along the trajectory transfers features to the current time step and enables feature interaction along the trajectory. This allows the model to recognize which features belong to the same object, aligning the features across agents for effective fusion. As a result, each agent aligns its feature map to the ego vehicle's latest time step.",
+        "bbox": [
+            89,
+            268,
+            483,
+            436
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.5. Inter-agent fusion, detection head and loss",
+        "text_level": 1,
+        "bbox": [
+            89,
+            444,
+            452,
+            460
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "While inter-agent fusion has been extensively studied [3, 36, 45], this paper adopts a simplified approach by concatenating features from all agents and utilizing a basic convolution layer for fusion. We use an anchor-based detection head and employ three losses: a detection loss, a field loss, and an offset loss:",
+        "bbox": [
+            89,
+            465,
+            483,
+            555
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {d e t e c t i o n}} + \\alpha \\mathcal {L} _ {\\text {f i e l d}} + \\beta \\mathcal {L} _ {\\text {o f f s e t}}, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            173,
+            566,
+            480,
+            580
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{L}_{\\mathrm{detection}}$  is consistent with the loss used in [14],  $\\mathcal{L}_{\\mathrm{field}}$  includes a focal loss [17] for the position field and an  $L1$  loss for the orientation field. The constants  $\\alpha$  and  $\\beta$  balance the contributions of each loss term. Here, we set  $\\alpha = \\beta = 0.05$ . Considering the offset loss, because the generated attention positions are unordered, we first apply the Sinkhorn algorithm [7, 26, 27] to compute the matching probabilities  $P_{jk} \\in \\mathbb{R}^{n \\times n}$  between the generated attention positions and the ground truth. This is done using the L1 distance matrix  $C_{jk} \\in \\mathbb{R}^{n \\times n}$  as the cost for the pixel located at  $(j, k)$  on the feature map. The offset loss is then defined as:",
+        "bbox": [
+            89,
+            590,
+            483,
+            771
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {o f f s e t}} = \\sum_ {j, k} P _ {j k} \\cdot C _ {j k}, \\forall j \\in [ 1, H ], k \\in [ 1, W ]. \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            114,
+            777,
+            483,
+            811
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            830,
+            223,
+            847
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To evaluate robustness to cross-agent latency, we conduct 3D object detection experiments on two large-scale real-world datasets, V2V4Real [38] and DAIR-V2X-Seq [43].",
+        "bbox": [
+            89,
+            854,
+            483,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/55564513160de392774c28db019896855f74f3937e31077014e26ee013505a5b.jpg",
+        "image_caption": [
+            "(a) Trajectory-aware attention layers."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            527,
+            87,
+            903,
+            233
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/e47348146adec98c21b18cdd1b1ad9c97eedf19a6115575a1a39490bceb3a2f8.jpg",
+        "image_caption": [
+            "(b) Query and ground truth attention positions.",
+            "Figure 4. Attention positions and trajectory-aware attention layers."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            529,
+            253,
+            898,
+            393
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "These datasets encompass both V2V and V2I scenarios, as well as sparse and dense LiDAR conditions. The detection results are evaluated using Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 (AP50) and 0.7 (AP70). To enable a fair comparison, we omit feature compression operations from the methods under evaluation.",
+        "bbox": [
+            511,
+            460,
+            905,
+            551
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Implementation details",
+        "text_level": 1,
+        "bbox": [
+            511,
+            561,
+            728,
+            577
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Datasets and Benchmarks. V2V4Real [38] is specialized for V2V cooperative detection in urban settings, and DAIR-V2X-Seq [43] captures sequential data in V2I scenarios at intersections. The benchmarks include single-agent perception relying on the ego vehicle's data for detection, alongside eight SOTA intermediate fusion models: F-Cooper [3], V2VNet [33], Where2comm [12], AttFuse [36], V2X-ViT [35], CoBEVT [37], MRCNet [11], and ERMVP [45].",
+        "bbox": [
+            511,
+            583,
+            905,
+            703
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training details. All the models are trained within the same codebase for 60 epochs on nine RTX 3090 GPUs, employing the one-cycle learning rate strategy [28] with AdamW [19] as the optimizer. The point clouds are segmented at a grid size of  $0.4 \\times 0.4\\mathrm{m}$ , and the backbone's feature map size output is one-fourth the original size. During training on V2V4Real, all models shuffle the vehicle list and randomly select the ego vehicle at the start of each epoch. Standard LiDAR data augmentations are applied, including random flip, scaling, rotation, and translation. Multi-frame methods are trained with random latency augmentation, following a uniform distribution from 0 to  $400~\\mathrm{ms}$ . TraF-Align leverages two LiDAR frames for the ego vehicle and four",
+        "bbox": [
+            511,
+            704,
+            906,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "12052",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/e322f68fde4a5aa4df1725cd69255fcc4ffaf9c03f3b71f7714021cea3620972.jpg",
+        "table_caption": [
+            "Table 1. Performance comparison of vehicle class on V2V4Real validation and DAIR-V2X-Seq validation dataset with different latencies. The results are reported in AP 50/70 (%)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"5\">V2V4Real</td><td colspan=\"5\">DAIR-V2X-Seq</td></tr><tr><td>0ms</td><td>100 ms</td><td>200 ms</td><td>300 ms</td><td>400 ms</td><td>0ms</td><td>100 ms</td><td>200 ms</td><td>300 ms</td><td>400 ms</td></tr><tr><td>Single Agent</td><td>48.75/30.53</td><td>48.75/30.53</td><td>48.75/30.53</td><td>48.75/30.53</td><td>48.75/30.53</td><td>51.31/40.45</td><td>51.31/40.45</td><td>51.31/40.45</td><td>51.31/40.45</td><td>51.31/40.45</td></tr><tr><td>AttFuse [36]</td><td>63.84/33.30</td><td>59.16/27.36</td><td>55.57/26.03</td><td>53.25/25.44</td><td>51.69/25.09</td><td>68.94/49.68</td><td>65.83/45.76</td><td>63.02/44.24</td><td>61.61/43.28</td><td>60.19/42.53</td></tr><tr><td>F-Cooper [3]</td><td>66.86/33.39</td><td>62.15/27.61</td><td>58.03/26.30</td><td>55.31/25.46</td><td>53.92/24.95</td><td>69.47/50.01</td><td>65.85/44.96</td><td>62.74/43.21</td><td>61.12/42.28</td><td>60.01/41.70</td></tr><tr><td>V2VNet [33]</td><td>67.11/35.47</td><td>61.97/29.14</td><td>58.09/27.85</td><td>56.26/27.43</td><td>54.62/26.75</td><td>71.44/51.91</td><td>67.26/46.63</td><td>63.84/44.59</td><td>61.68/43.24</td><td>59.78/42.08</td></tr><tr><td>Where2comm [12]</td><td>67.03/37.55</td><td>61.86/31.52</td><td>58.22/30.20</td><td>56.22/29.71</td><td>54.47/29.07</td><td>68.37/48.50</td><td>64.81/44.32</td><td>61.51/42.80</td><td>59.66/41.63</td><td>58.25/40.80</td></tr><tr><td>V2X-ViT [35]</td><td>68.51/35.88</td><td>63.45/29.57</td><td>59.42/28.04</td><td>57.35/27.65</td><td>55.60/27.20</td><td>70.21/50.92</td><td>66.42/46.44</td><td>62.58/44.29</td><td>61.00/43.56</td><td>59.63/42.66</td></tr><tr><td>MRCNet [11]</td><td>67.83/37.37</td><td>62.65/31.59</td><td>58.98/30.39</td><td>56.70/29.53</td><td>55.03/28.94</td><td>69.27/50.67</td><td>65.90/45.60</td><td>62.65/44.15</td><td>60.09/42.39</td><td>58.19/41.35</td></tr><tr><td>CoBEVT [37]</td><td>70.59/38.82</td><td>65.50/32.11</td><td>61.55/30.40</td><td>58.91/29.54</td><td>57.05/29.26</td><td>71.51/51.31</td><td>67.60/46.55</td><td>64.38/45.03</td><td>62.39/44.11</td><td>60.74/43.08</td></tr><tr><td>ERMVP [45]</td><td>70.74/38.45</td><td>65.53/31.91</td><td>61.63/30.53</td><td>59.24/30.18</td><td>57.26/29.31</td><td>71.19/53.85</td><td>67.04/48.09</td><td>63.67/46.37</td><td>61.23/44.76</td><td>59.55/43.93</td></tr><tr><td>TraF-Align (Ours)</td><td>74.28/44.14</td><td>73.04/39.99</td><td>71.51/39.54</td><td>70.07/38.99</td><td>69.41/38.49</td><td>76.90/58.31</td><td>73.95/53.68</td><td>72.69/52.83</td><td>71.66/52.33</td><td>71.22/52.01</td></tr></table>",
+        "bbox": [
+            91,
+            128,
+            898,
+            296
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "for the cooperative vehicle on V2V4Real, while using four LiDAR frames for the vehicle and infrastructure on DAIR-V2X-Seq. The number of attention layers is set to two, each containing four attention heads.",
+        "bbox": [
+            88,
+            321,
+            480,
+            382
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Latency settings. To comprehensively evaluate robustness to time delays, we introduce delays of 0, 100, 200, 300, and  $400\\mathrm{ms}$  to one of the agents separately, extensively testing the model's performance under each delay condition. Unless specified otherwise, these delays are applied to the point cloud data of the non-ego agent.",
+        "bbox": [
+            88,
+            383,
+            482,
+            474
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Comparison with state-of-the-art",
+        "text_level": 1,
+        "bbox": [
+            89,
+            484,
+            382,
+            500
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Detection performance. Table 1 presents a comparison of detection performance on two real-world datasets for vehicle class, which includes cars, vans, trucks, buses, etc. It's observed that the proposed method i) achieves state-of-the-art performance across all benchmarks under various latency conditions on both datasets. ii) In synchronous scenarios, i.e., when the delay is  $0\\mathrm{ms}$ , our model improves AP 50/70 by  $25.53\\% / 13.61\\%$  (on V2V4Real) and  $25.59\\% / 18.26\\%$  (on DAIR-V2X-Seq) compared to single-agent perception, it also surpasses ERMVP, the SOTA performance of feature-level methods, by  $3.54\\% / 5.69\\%$  and  $5.71\\% / 4.46\\%$ , respectively.",
+        "bbox": [
+            88,
+            507,
+            482,
+            686
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Robustness to inter-agent latency. As shown in Table 1, our model demonstrates stronger robustness in the presence of time delays. Even under a  $400~\\mathrm{ms}$  delay, our AP decreases by only  $4.87\\% /5.65\\%$  (on V2V4Real) and  $5.68\\% /6.30\\%$  (on DAIR-V2X-Seq) compared to the synchronous scenario. This performance significantly outperforms the ERMVP, which shows declines of  $13.48\\% /9.14\\%$  and  $11.64\\% /9.92\\%$  , respectively.",
+        "bbox": [
+            88,
+            688,
+            482,
+            809
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To gain a deeper understanding of the impact of latency on collaboration, we plot the Precision-Recall curves of the models under various latency conditions, as shown in Figure 5. It is obvious that cooperation is effective in enhancing the recall limit of single-vehicle perception, helping the ego vehicle to detect a more significant number of objects.",
+        "bbox": [
+            88,
+            810,
+            483,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d7d5e17242f79c59992fe9cc8f2b2a1a2073234a6b1571b037c6d9470d18cc6c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            526,
+            321,
+            704,
+            479
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/33902ffe0d7955c3ac3b544de806c2fa363eab9a2ef4149b0a5e5d9f1ded6ee2.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            714,
+            320,
+            895,
+            479
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/1a9ad1243c0ef0fbc8106297e8e9cc52ee1392f15c426150cc3e9b800a8dee69.jpg",
+        "image_caption": [
+            "Figure 5. Precision-Recall curves for models on the V2V4Real validation dataset under various latency conditions."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            484,
+            704,
+            672
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/dd396af796c57a7a09b42fec6a53f420d55d851cace7b4f4551f0fcc0ef62558.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            715,
+            484,
+            897,
+            672
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "However, at a  $200\\mathrm{ms}$  latency, state-of-the-art methods exhibit lower precision in the low-recall range compared to single-vehicle perception. This suggests that the delayed collaborative information interferes with the objects that the single vehicle could detect, corresponding to the semantic misalignment issues depicted in Figure 1-II. In areas not visible to the ego vehicle, specifically at high recall levels, the precision of SOTA methods significantly declines with the introduction of latency. This indicates that spatial misalignment issues (Figure 1-I) become dominant, lead",
+        "bbox": [
+            511,
+            750,
+            906,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "12053",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "ing to decreased performance. These findings reinforce our analysis of latency's impact on cooperative perception. Furthermore, using ERMVP as an example, the differences between the PR curves diminish as latency increases, indicating that existing models progressively lose their effectiveness in managing delays. In contrast, TraF-Align retains the original performance of single-agent perception in lowrecall regions and shows no significant decline in performance in high-recall areas due to latency. This highlights our model's ability to establish strong semantic consistency between the ego vehicle's data and the received information, demonstrating robust reconstruction capability and effectively addressing the scenarios in Figures 1-I and 1-II.",
+        "bbox": [
+            89,
+            90,
+            483,
+            287
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3. Ablation studies",
+        "text_level": 1,
+        "bbox": [
+            89,
+            300,
+            251,
+            315
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this section, we conducted ablation studies on the effectiveness of significant components and loss and multiframe point clouds using average precision as the evaluation metric. Other ablation experiments and analyses (using PR curves) are presented in the supplementary material.",
+        "bbox": [
+            89,
+            323,
+            483,
+            398
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Effectiveness of Major Components. To evaluate the contribution of modules, we progressively remove i) the field predictor, ii) the offset generator, and iii) the attention layers. The results are shown in Table 2. The experiments are conducted on the DAIR-V2X-Seq dataset. To ensure a comprehensive assessment of each component's contributions, we train the model for 120 epochs, double the duration of the experiments presented in Table 1. Additionally, we test the effects of adding latency to both the ego and non-ego agents to fully evaluate the impact of each module. The consistent rise in detection precision over different latency settings demonstrates the effectiveness of each introduced component. Notably, applying latency to vehicles results in lower AP than when latency is applied to infrastructure. This discrepancy arises from the different distributions of object visibility between the vehicle agent and roadside in",
+        "bbox": [
+            89,
+            400,
+            483,
+            642
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/0f35fae71d44b655b919c9be8faebeede34fcc70d6af33dc6bf11f5af92e8fed.jpg",
+        "table_caption": [
+            "Table 2. Results of the ablation study of the core components on the DAIR-V2X-Seq dataset."
+        ],
+        "table_footnote": [
+            "FP: Field predictor. OG: Offset generator. AL: Attention layers."
+        ],
+        "table_body": "<table><tr><td rowspan=\"2\">FP OG AL</td><td colspan=\"5\">Latency of Infrastructure</td></tr><tr><td>0ms</td><td>100ms</td><td>200ms</td><td>300ms</td><td>400ms</td></tr><tr><td></td><td>75.35/54.72</td><td>73.01/50.81</td><td>71.88/50.17</td><td>71.04/49.77</td><td>70.09/49.45</td></tr><tr><td></td><td>✓</td><td>76.31/57.02</td><td>74.60/53.53</td><td>73.90/52.34</td><td>72.91/51.99</td></tr><tr><td></td><td>✓</td><td>77.11/58.63</td><td>75.15/54.77</td><td>74.18/53.77</td><td>73.24/52.99</td></tr><tr><td></td><td>✓</td><td>✓</td><td>77.83/59.72</td><td>75.74/55.85</td><td>74.99/54.82</td></tr><tr><td rowspan=\"2\">FP OG AL</td><td colspan=\"5\">Latency of Vehicle</td></tr><tr><td>0ms</td><td>100ms</td><td>200ms</td><td>300ms</td><td>400ms</td></tr><tr><td></td><td>75.35/54.72</td><td>74.31/49.61</td><td>72.17/43.32</td><td>67.19/34.82</td><td>56.93/24.51</td></tr><tr><td></td><td>✓</td><td>76.31/57.02</td><td>75.10/51.93</td><td>72.91/46.22</td><td>69.22/39.15</td></tr><tr><td></td><td>✓</td><td>77.11/58.63</td><td>75.76/53.48</td><td>73.44/47.55</td><td>69.92/40.52</td></tr><tr><td></td><td>✓</td><td>✓</td><td>77.83/59.72</td><td>76.41/54.88</td><td>74.24/49.46</td></tr></table>",
+        "bbox": [
+            91,
+            707,
+            480,
+            882
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/b03905b131ae06e2b47176f978f2208e3eb522d897484cb47d1bead45e31a7bc.jpg",
+        "table_caption": [
+            "Table 3. Results of the ablation study of the loss on the DAIR-V2X-Seq dataset."
+        ],
+        "table_footnote": [
+            "FL:Field loss. OL: Offset loss."
+        ],
+        "table_body": "<table><tr><td rowspan=\"2\">FL OL</td><td colspan=\"5\">Latency of Infrastructure</td></tr><tr><td>0ms</td><td>100ms</td><td>200ms</td><td>300ms</td><td>400ms</td></tr><tr><td rowspan=\"2\">✓</td><td>76.43/57.74</td><td>74.44/53.91</td><td>73.65/52.97</td><td>72.71/52.10</td><td>71.83/51.87</td></tr><tr><td>77.12/58.36</td><td>74.96/54.51</td><td>74.02/53.68</td><td>72.98/52.91</td><td>72.01/52.27</td></tr><tr><td>✓</td><td>77.15/58.70</td><td>75.19/54.58</td><td>74.31/53.66</td><td>73.68/53.07</td><td>72.52/52.28</td></tr><tr><td>✓</td><td>77.83/59.72</td><td>75.74/55.85</td><td>74.99/54.82</td><td>74.10/54.07</td><td>72.89/53.21</td></tr><tr><td rowspan=\"2\">FL OL</td><td colspan=\"5\">Latency of Vehicle</td></tr><tr><td>0ms</td><td>100ms</td><td>200ms</td><td>300ms</td><td>400ms</td></tr><tr><td rowspan=\"2\">✓</td><td>76.43/57.74</td><td>75.09/52.81</td><td>72.84/46.54</td><td>69.14/39.12</td><td>61.53/29.15</td></tr><tr><td>77.12/58.36</td><td>75.69/53.23</td><td>72.76/46.85</td><td>69.25/39.78</td><td>62.13/30.41</td></tr><tr><td>✓</td><td>77.15/58.70</td><td>75.80/53.90</td><td>73.34/48.49</td><td>70.12/41.80</td><td>63.21/32.02</td></tr><tr><td>✓</td><td>77.83/59.72</td><td>76.41/54.88</td><td>74.24/49.46</td><td>70.75/42.62</td><td>63.59/32.31</td></tr></table>",
+        "bbox": [
+            516,
+            128,
+            903,
+            303
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/730d661a4453700fb5be3e7a5cb5543a368244299c1e2f39368ab5567b3902f5.jpg",
+        "image_caption": [
+            "(a) LiDAR points."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            335,
+            633,
+            541
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4f8815cb30ce6da13f277e6a32c79dbc91b6687d022eb8dc87d423dcd62c34f6.jpg",
+        "image_caption": [
+            "(b) Trajectory field.",
+            "Figure 6. Visualization of TraF-Align in real-world scenarios from the DAIR-V2X-Seq dataset with latency of  $400~\\mathrm{ms}$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            638,
+            335,
+            746,
+            545
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/a3b0ddecd3e40ab65b9a8bfcd3225a97fcdd0821f18f7e2c632b84b4ac408f3b.jpg",
+        "image_caption": [
+            "(c) Deep features."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            764,
+            335,
+            880,
+            544
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "frastructure agent in the dataset.",
+        "bbox": [
+            511,
+            628,
+            727,
+            642
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Effectiveness of loss. Table 3 presents the ablation study of field loss and offset loss, both designed to guide the attention biases, under the same settings as in Table 2. The results show that introducing the field loss provides better initial values for the offset generator, enhancing the model's performance. Optimal results are achieved only when both losses are incorporated.",
+        "bbox": [
+            511,
+            643,
+            906,
+            747
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Impact of LiDAR frames. Table 4 presents the ablation study conducted on V2V4Real, investigating the impact of different historical frame numbers for both the ego and cooperative vehicle on detection performance and robustness to time delays. It's observed that increasing the ego frames primarily enhances AP70, indicating improved accuracy in object localization, while AP50 shows no significant increase. Conversely, increasing the CV frames enhances both AP50 and the model's robustness to time delays, as most latency originates from the cooperative vehicle's Li",
+        "bbox": [
+            511,
+            750,
+            908,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "12054",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/695ed3c096d1b7b5b1e7404d998ec06bb86258471cf4611290aa867e0e8d4986.jpg",
+        "image_caption": [
+            "Figure 7. Qualitative results in real-world scenarios from the DAIR-V2X-Seq dataset with latency of  $400\\mathrm{ms}$ . Green and red boxes denote the ground truths and detection results, respectively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            106,
+            85,
+            903,
+            281
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/d5d1248b7c8bcaa2d1c427ad7851ba727862a3b3bb32a101d5528f036df64954.jpg",
+        "table_caption": [
+            "Table 4. Results of the ablation study of the number of LiDAR frames on the V2V4Real dataset."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\" colspan=\"2\">Ego CV</td><td colspan=\"5\">Latency of Cooperative Vehicle (CV)</td></tr><tr><td>0ms</td><td>100ms</td><td>200ms</td><td>300ms</td><td>400ms</td></tr><tr><td>1</td><td>1</td><td>73.35/40.54</td><td>69.76/34.03</td><td>66.51/32.54</td><td>65.17/32.13</td><td>64.28/32.28</td></tr><tr><td>4</td><td>1</td><td>72.85/44.81</td><td>69.45/38.32</td><td>66.73/37.16</td><td>66.00/37.28</td><td>65.67/37.32</td></tr><tr><td>4</td><td>2</td><td>73.02/45.00</td><td>71.74/39.88</td><td>69.71/38.25</td><td>67.88/37.49</td><td>66.98/37.39</td></tr><tr><td>1</td><td>4</td><td>71.28/42.16</td><td>70.28/39.18</td><td>69.31/38.15</td><td>67.68/37.52</td><td>66.47/36.90</td></tr><tr><td>2</td><td>4</td><td>74.28/44.14</td><td>73.04/39.99</td><td>71.51/39.54</td><td>70.07/38.99</td><td>69.41/38.49</td></tr><tr><td>4</td><td>4</td><td>74.76/44.03</td><td>73.28/39.58</td><td>71.64/38.58</td><td>70.10/37.70</td><td>69.08/37.08</td></tr></table>",
+        "bbox": [
+            96,
+            382,
+            478,
+            500
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "DAR. Furthermore, having too many point cloud frames for the ego vehicle may not yield optimal results, as it could introduce additional motion blur issues, which have been proved in the context of single-vehicle perception [24].",
+        "bbox": [
+            89,
+            522,
+            483,
+            583
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.4. Qualitative results",
+        "text_level": 1,
+        "bbox": [
+            89,
+            590,
+            269,
+            607
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Visualization of detection results. Figure 7a-7d compare TraF-Align with three previous methods on a real-world scenario from DAIR-V2X-Seq dataset under a latency of  $400~\\mathrm{ms}$ . It is observed that in region  $A_{1}$ , due to the latency in roadside (Agent 2), V2VNet, COBEVT, and V2X-ViT all exhibit a false positive detection and fail to capture the correct bounding box; a similar issue occurs in region  $A_{2}$ . The reasons for this phenomenon have been discussed in section 1. In contrast, TraF-Align not only identifies the object but also avoids false positives. This improvement is due to our trajectory-guided attention, which enables both agents to achieve a consensus on the same object.",
+        "bbox": [
+            89,
+            613,
+            483,
+            794
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Visualization of trajectory field. Figure 6b illustrates the predicted trajectory fields (specifically, the position field) from both the vehicle (Agent 1) and roadside (Agent 2) perspectives. Despite the delays in the roadside point clouds, the predicted trajectory field is aligned to the latest frame.",
+        "bbox": [
+            89,
+            795,
+            483,
+            869
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Visualization of attention. Figure 7e shows the response within the  $A_{1}$  box of Figure 7d from the roadside perspec",
+        "bbox": [
+            89,
+            869,
+            483,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "tive, with six different query positions. Since the ego data is undelayed, the offset distribution centers around the query points, resembling a convolution kernel. In contrast, the roadside agent extends attention along the trajectory. After fusion by TraF-Align, as shown in Figure 6c, the roadside features align toward the vehicles, demonstrating spatial alignment. The feature intensity near the trajectory is uniform, indicating semantic alignment. More results are available in the supplementary material.",
+        "bbox": [
+            511,
+            344,
+            906,
+            482
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusions",
+        "text_level": 1,
+        "bbox": [
+            513,
+            496,
+            640,
+            512
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We propose TraF-Align, a novel cooperative perception framework robust to inter-agent latencies. Our method handles delays by predicting the flow of object features and aligning them across agents through attention within the expected trajectory paths. Extensive experiments on real-world datasets demonstrate its effectiveness.",
+        "bbox": [
+            511,
+            522,
+            905,
+            611
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitations and future work. TraF-Align is primarily trained on data with regular latencies (multiples of 100 ms). In practical applications, more realistic latency settings should be considered. Additionally, while TraF-Align is robust to latency, detection precision under delay can still be improved. The main challenges are: i) Designing more accurate trajectory field predictors and attention position generators to enhance feature reconstruction capabilities; ii) Improving object localization precision. Although multi-frame information improves motion understanding, it also introduces motion blur, making precise object localization challenging and iii) Addressing pose estimation errors. Multi-frame processing relies on accurate inter-frame and inter-agent pose alignment, errors in these estimates can degrade performance. Future work will explore methods to reduce the impact of these factors.",
+        "bbox": [
+            511,
+            613,
+            906,
+            853
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments. This research is supported by the National Key Research and Development Program of China under Grant 2024YFB2505803.",
+        "bbox": [
+            511,
+            854,
+            908,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "12055",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "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] Kunyang Cai, Ting Qu, Bingzhao Gao, and Hong Chen. Consensus-based distributed cooperative perception for connected and automated vehicles. IEEE Transactions on Intelligent Transportation Systems, 24(8):8188-8208, 2023. 2",
+            "[2] Antoine Caillot, Safa Ouerghi, Pascal Vasseur, Rémi Boutteau, and Yohan Dupuis. Survey on cooperative perception in an automotive context. IEEE Transactions on Intelligent Transportation Systems, 23(9):14204-14223, 2022. 1",
+            "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 2, 5, 6",
+            "[4] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524. IEEE, 2019. 1",
+            "[5] Xuesong Chen, Shaoshuai Shi, Benjamin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In European Conference on Computer Vision, pages 680-697. Springer, 2022. 2",
+            "[6] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.4",
+            "[7] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 5",
+            "[8] Minh-Quan Dao, Julie Stephany Berrio, Vincent Frémont, Mao Shan, Elwan Héry, and Stewart Worrall. Practical collaborative perception: A framework for asynchronous and multi-agent 3d object detection. IEEE Transactions on Intelligent Transportation Systems, 2024. 2",
+            "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 5",
+            "[10] 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. 3",
+            "[11] Shixin Hong, Yu Liu, Zhi Li, Shaohui Li, and You He. Multiagent collaborative perception via motion-aware robust communication network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15301-15310, 2024. 3, 5, 6",
+            "[12] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 5, 6",
+            "[13] Tao Huang, Jianan Liu, Xi Zhou, Dinh C Nguyen, Mostafa Rahimi Azghadi, Yuxuan Xia, Qing-Long Han, and Sumei Sun. V2x cooperative perception for autonomous"
+        ],
+        "bbox": [
+            93,
+            114,
+            482,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "driving: Recent advances and challenges. arXiv preprint arXiv:2310.03525, 2023. 2",
+            "[14] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 5",
+            "[15] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In Computer Vision - ECCV 2022, pages 316-332, Cham, 2022. Springer Nature Switzerland. 1, 2",
+            "[16] Jinyu Li, Chenxu Luo, and Xiaodong Yang. Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17567-17576, 2023. 3",
+            "[17] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5",
+            "[18] Si Liu, Chen Gao, Yuan Chen, Xingyu Peng, Xianghao Kong, Kun Wang, Runsheng Xu, Wentao Jiang, Hao Xiang, Jiaqi Ma, et al. Towards vehicle-to-everything autonomous driving: A survey on collaborative perception. arXiv preprint arXiv:2308.16714, 2023. 2",
+            "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5",
+            "[20] Yifan Lu, Yue Hu, Yiqi Zhong, Dequan Wang, Yanfeng Wang, and Siheng Chen. An extensible framework for open heterogeneous collaborative perception. In The Twelfth International Conference on Learning Representations, 2024. 1",
+            "[21] Tao Ma, Xuemeng Yang, Hongbin Zhou, Xin Li, Botian Shi, Junjie Liu, Yuchen Yang, Zhizheng Liu, Liang He, Yu Qiao, et al. Detzero: Rethinking offboard 3d object detection with long-term sequential point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6736-6747, 2023. 2",
+            "[22] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6134-6144, 2021. 2",
+            "[23] Hang Qiu, Po-Han Huang, Namo Asavisanu, Xiaochen Liu, Konstantinos Psounis, and Ramesh Govindan. Autocast: scalable infrastructure-less cooperative perception for distributed collaborative driving. In Proceedings of the 20th Annual International Conference on Mobile Systems, Applications and Services, pages 128-141, 2022. 1",
+            "[24] Yao Rong, Xiangyu Wei, Tianwei Lin, Yueyu Wang, and Enkelejda Kasneci. Dynstatf: An efficient feature fusion strategy for lidar 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3237-3246, 2023. 2, 8",
+            "[25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted"
+        ],
+        "bbox": [
+            516,
+            92,
+            906,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "12056",
+        "bbox": [
+            480,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4",
+            "[26] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 5",
+            "[27] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967. 5",
+            "[28] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, pages 369-386. SPIE, 2019. 5",
+            "[29] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. A cooperative perception system robust to localization errors. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-6. IEEE, 2023. 1",
+            "[30] Zhiying Song, Tenghui Xie, Hailiang Zhang, Jiaxin Liu, Fuxi Wen, and Jun Li. A spatial calibration method for robust cooperative perception. IEEE Robotics and Automation Letters, 9(5):4011-4018, 2024. 1",
+            "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5",
+            "[32] Haiyan Wang and Yingli Tian. Sequential point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2",
+            "[33] Tsun-Hsuan Wang, Sivaban Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 605–621. Springer, 2020. 1, 2, 5, 6",
+            "[34] Sizhe Wei, Yuxi Wei, Yue Hu, Yifan Lu, Yiqi Zhong, Siheng Chen, and Ya Zhang. Asynchrony-robust collaborative perception via bird's eye view flow. In Advances in Neural Information Processing Systems, pages 28462-28477. Curran Associates, Inc., 2023. 1, 2",
+            "[35] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 1, 2, 5, 6",
+            "[36] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 1, 5, 6",
+            "[37] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view"
+        ],
+        "bbox": [
+            91,
+            90,
+            483,
+            883
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "semantic segmentation with sparse transformers. In Conference on Robot Learning, pages 989-1000. PMLR, 2023. 5, 6",
+            "[38] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 5",
+            "[39] Runsheng Xu, Chia-Ju Chen, Zhengzhong Tu, and Ming-Hsuan Yang. V2x-vitv2: Improved vision transformers for vehicle-to-everything cooperative perception. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2",
+            "[40] Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23383-23392, 2023. 3",
+            "[41] Lei Yang, Kaicheng Yu, Tao Tang, Jun Li, Kun Yuan, Li Wang, Xinyu Zhang, and Peng Chen. Bevheight: A robust framework for vision-based roadside 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21611-21620, 2023. 1",
+            "[42] Zetong Yang, Yin Zhou, Zhifeng Chen, and Jiquan Ngiam. 3d-man: 3d multi-frame attention network for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1863-1872, 2021. 2",
+            "[43] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5486-5495, 2023. 2, 5",
+            "[44] Haibao Yu, Yingjuan Tang, Enze Xie, Jilei Mao, Ping Luo, and Zaiqing Nie. Flow-based feature fusion for vehicleinfrastructure cooperative 3d object detection. Advances in Neural Information Processing Systems, 36, 2024. 1, 2",
+            "[45] Jingyu Zhang, Kun Yang, Yilei Wang, Hanqi Wang, Peng Sun, and Liang Song. ERMVP: Communication-efficient and collaboration-robust multi-vehicle perception in challenging environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12575–12584, 2024. 1, 2, 5, 6",
+            "[46] Qingzhao Zhang, Xumiao Zhang, Ruiyang Zhu, Fan Bai, Mohammad Naserian, and Z Morley Mao. Robust real-time multi-vehicle collaboration on asynchronous sensors. In Proceedings of the 29th Annual International Conference on Mobile Computing and Networking, pages 1–15, 2023. 2",
+            "[47] Xumiao Zhang, Anlan Zhang, Jiachen Sun, Xiao Zhu, Y Ethan Guo, Feng Qian, and Z Morley Mao. Emp: Edge-assisted multi-vehicle perception. In Proceedings of the 27th Annual International Conference on Mobile Computing and Networking, pages 545–558, 2021. 1"
+        ],
+        "bbox": [
+            516,
+            92,
+            905,
+            883
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "12057",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    }
+]