Add Batch de207d3f-4356-4c2f-9c32-ffe3a45e9154 data

.gitattributes

@@ -2588,3 +2588,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/V2PE_[[:space:]]Improving[[:space:]]Multimodal[[:space:]]Long-Context[[:space:]]Capability[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models[[:space:]]with[[:space:]]Variable[[:space:]]Visual[[:space:]]Position[[:space:]]Encoding/75b04339-69b5-48e5-a71f-b17b99c6243a_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/V2XPnP_[[:space:]]Vehicle-to-Everything[[:space:]]Spatio-Temporal[[:space:]]Fusion[[:space:]]for[[:space:]]Multi-Agent[[:space:]]Perception[[:space:]]and[[:space:]]Prediction/36b3719d-faf2-4104-a9f7-9253a3958218_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/V2XScenes_[[:space:]]A[[:space:]]Multiple[[:space:]]Challenging[[:space:]]Traffic[[:space:]]Conditions[[:space:]]Dataset[[:space:]]for[[:space:]]Large-Range[[:space:]]Vehicle-Infrastructure[[:space:]]Collaborative[[:space:]]Perception/e74c8f66-2f14-41e8-a64e-91a69d371436_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VA-MoE_[[:space:]]Variables-Adaptive[[:space:]]Mixture[[:space:]]of[[:space:]]Experts[[:space:]]for[[:space:]]Incremental[[:space:]]Weather[[:space:]]Forecasting/122157a9-e33d-4cdd-b0f9-d8b983f2a104_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VACE_[[:space:]]All-in-One[[:space:]]Video[[:space:]]Creation[[:space:]]and[[:space:]]Editing/bb0c9cf5-4f56-4fe0-bdd0-bf9454fc4819_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VAFlow_[[:space:]]Video-to-Audio[[:space:]]Generation[[:space:]]with[[:space:]]Cross-Modality[[:space:]]Flow[[:space:]]Matching/7812feed-641c-4918-a582-f934ca29c2a8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VAGUE_[[:space:]]Visual[[:space:]]Contexts[[:space:]]Clarify[[:space:]]Ambiguous[[:space:]]Expressions/321d7adc-1768-447e-aad1-8d77c13702f1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VALLR_[[:space:]]Visual[[:space:]]ASR[[:space:]]Language[[:space:]]Model[[:space:]]for[[:space:]]Lip[[:space:]]Reading/88644d63-e953-49ce-b2d3-f0418686524b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VCA_[[:space:]]Video[[:space:]]Curious[[:space:]]Agent[[:space:]]for[[:space:]]Long[[:space:]]Video[[:space:]]Understanding/a73c253b-5afa-42a6-b79a-31559f118997_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VEGGIE_[[:space:]]Instructional[[:space:]]Editing[[:space:]]and[[:space:]]Reasoning[[:space:]]Video[[:space:]]Concepts[[:space:]]with[[:space:]]Grounded[[:space:]]Generation/d727aeb7-1f89-498a-ac85-9bcc05bddf8b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VFlowOpt_[[:space:]]A[[:space:]]Token[[:space:]]Pruning[[:space:]]Framework[[:space:]]for[[:space:]]LMMs[[:space:]]with[[:space:]]Visual[[:space:]]Information[[:space:]]Flow-Guided[[:space:]]Optimization/b08d8e36-4b30-4260-aad4-6cb0fd4d67ca_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VGGSounder_[[:space:]]Audio-Visual[[:space:]]Evaluations[[:space:]]for[[:space:]]Foundation[[:space:]]Models/a6042e74-fb7f-4488-a331-f3b644cd7772_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VGMamba_[[:space:]]Attribute-to-Location[[:space:]]Clue[[:space:]]Reasoning[[:space:]]for[[:space:]]Quantity-Agnostic[[:space:]]3D[[:space:]]Visual[[:space:]]Grounding/db8ed93c-5bcb-43cd-a62c-cc51ec338aab_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VIGFace_[[:space:]]Virtual[[:space:]]Identity[[:space:]]Generation[[:space:]]for[[:space:]]Privacy-Free[[:space:]]Face[[:space:]]Recognition[[:space:]]Dataset/570e97bc-508f-4feb-946d-7b35db443845_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VIPerson_[[:space:]]Flexibly[[:space:]]Generating[[:space:]]Virtual[[:space:]]Identity[[:space:]]for[[:space:]]Person[[:space:]]Re-Identification/9c97a79e-c7c9-4628-ab6c-7d4d09ccbb66_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VISION-XL_[[:space:]]High[[:space:]]Definition[[:space:]]Video[[:space:]]Inverse[[:space:]]Problem[[:space:]]Solver[[:space:]]using[[:space:]]Latent[[:space:]]Image[[:space:]]Diffusion[[:space:]]Models/339912e1-b9b8-4ac6-a699-6e214154b6e0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VISO_[[:space:]]Accelerating[[:space:]]In-orbit[[:space:]]Object[[:space:]]Detection[[:space:]]with[[:space:]]Language-Guided[[:space:]]Mask[[:space:]]Learning[[:space:]]and[[:space:]]Sparse[[:space:]]Inference/bca87654-a43f-4442-9368-eee092911b00_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VITAL_[[:space:]]More[[:space:]]Understandable[[:space:]]Feature[[:space:]]Visualization[[:space:]]through[[:space:]]Distribution[[:space:]]Alignment[[:space:]]and[[:space:]]Relevant[[:space:]]Information[[:space:]]Flow/65791404-56e7-4d3c-a033-7abfb8aab04e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VLABench_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Benchmark[[:space:]]for[[:space:]]Language-Conditioned[[:space:]]Robotics[[:space:]]Manipulation[[:space:]]with[[:space:]]Long-Horizon[[:space:]]Reasoning[[:space:]]Tasks/bbac0723-47a6-4c9f-bec1-1f18e03a36ac_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VLDrive_[[:space:]]Vision-Augmented[[:space:]]Lightweight[[:space:]]MLLMs[[:space:]]for[[:space:]]Efficient[[:space:]]Language-grounded[[:space:]]Autonomous[[:space:]]Driving/03e7ad20-412e-4720-b543-ad870d80aec2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VLIPP_[[:space:]]Towards[[:space:]]Physically[[:space:]]Plausible[[:space:]]Video[[:space:]]Generation[[:space:]]with[[:space:]]Vision[[:space:]]and[[:space:]]Language[[:space:]]Informed[[:space:]]Physical[[:space:]]Prior/eba6c702-728f-4007-b7d6-27e6fa6404b9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VLM4D_[[:space:]]Towards[[:space:]]Spatiotemporal[[:space:]]Awareness[[:space:]]in[[:space:]]Vision[[:space:]]Language[[:space:]]Models/7ad15710-76ed-462f-ba27-85159b1f2146_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VLR-Driver_[[:space:]]Large[[:space:]]Vision-Language-Reasoning[[:space:]]Models[[:space:]]for[[:space:]]Embodied[[:space:]]Autonomous[[:space:]]Driving/a365b909-1be9-4f17-b2ec-25f898d74a15_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VLRMBench_[[:space:]]A[[:space:]]Comprehensive[[:space:]]and[[:space:]]Challenging[[:space:]]Benchmark[[:space:]]for[[:space:]]Vision-Language[[:space:]]Reward[[:space:]]Models/8e730ad2-3996-4a2e-9875-e28304c3e917_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VMBench_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Perception-Aligned[[:space:]]Video[[:space:]]Motion[[:space:]]Generation/1a1d91b4-eb97-44d0-a07d-41746c53ad96_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VMem_[[:space:]]Consistent[[:space:]]Interactive[[:space:]]Video[[:space:]]Scene[[:space:]]Generation[[:space:]]with[[:space:]]Surfel-Indexed[[:space:]]View[[:space:]]Memory/1781a2ca-916a-47a8-9f71-bf548311fbcb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VOVTrack_[[:space:]]Exploring[[:space:]]the[[:space:]]Potentiality[[:space:]]in[[:space:]]Raw[[:space:]]Videos[[:space:]]for[[:space:]]Open-Vocabulary[[:space:]]Multi-Object[[:space:]]Tracking/cb6ae00b-bc8e-44a2-b37e-6f565b52690f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VOccl3D_[[:space:]]A[[:space:]]Video[[:space:]]Benchmark[[:space:]]Dataset[[:space:]]for[[:space:]]3D[[:space:]]Human[[:space:]]Pose[[:space:]]and[[:space:]]Shape[[:space:]]Estimation[[:space:]]under[[:space:]]real[[:space:]]Occlusions/bee260a8-0889-4cd7-b8a5-2c0fb163c69d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VPO_[[:space:]]Aligning[[:space:]]Text-to-Video[[:space:]]Generation[[:space:]]Models[[:space:]]with[[:space:]]Prompt[[:space:]]Optimization/1912f989-0c41-45f7-ba83-c030d1f76cc0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VPR-Cloak_[[:space:]]A[[:space:]]First[[:space:]]Look[[:space:]]at[[:space:]]Privacy[[:space:]]Cloak[[:space:]]Against[[:space:]]Visual[[:space:]]Place[[:space:]]Recognition/ecc5979e-3031-4d35-8396-33fead75fa7b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VQ-SGen_[[:space:]]A[[:space:]]Vector[[:space:]]Quantized[[:space:]]Stroke[[:space:]]Representation[[:space:]]for[[:space:]]Creative[[:space:]]Sketch[[:space:]]Generation/27ec5cff-30ea-4eda-ad45-f2004607bb53_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VQ-VLA_[[:space:]]Improving[[:space:]]Vision-Language-Action[[:space:]]Models[[:space:]]via[[:space:]]Scaling[[:space:]]Vector-Quantized[[:space:]]Action[[:space:]]Tokenizers/a05d8d10-adb9-4812-8715-77d71aa82af2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VRBench_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Multi-Step[[:space:]]Reasoning[[:space:]]in[[:space:]]Long[[:space:]]Narrative[[:space:]]Videos/13199d03-d2f0-4f02-9446-3b78b6cbc2ed_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VRM_[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]via[[:space:]]Virtual[[:space:]]Relation[[:space:]]Matching/4752d372-1ed6-459b-bee6-6b202a028bae_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VSC_[[:space:]]Visual[[:space:]]Search[[:space:]]Compositional[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Model/9f76eb31-5853-4235-b96e-6306079810f4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VSP_[[:space:]]Diagnosing[[:space:]]the[[:space:]]Dual[[:space:]]Challenges[[:space:]]of[[:space:]]Perception[[:space:]]and[[:space:]]Reasoning[[:space:]]in[[:space:]]Spatial[[:space:]]Planning[[:space:]]Tasks[[:space:]]for[[:space:]]MLLMs/262f63b9-5c82-43f4-b645-d16941a910c1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VSRM_[[:space:]]A[[:space:]]Robust[[:space:]]Mamba-Based[[:space:]]Framework[[:space:]]for[[:space:]]Video[[:space:]]Super-Resolution/6e341a1e-2a3f-4224-8262-4e1ca75e0c71_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VSSD_[[:space:]]Vision[[:space:]]Mamba[[:space:]]with[[:space:]]Non-Causal[[:space:]]State[[:space:]]Space[[:space:]]Duality/4e862f57-d9f1-4a7e-9559-4423a19c05c9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VTimeCoT_[[:space:]]Thinking[[:space:]]by[[:space:]]Drawing[[:space:]]for[[:space:]]Video[[:space:]]Temporal[[:space:]]Grounding[[:space:]]and[[:space:]]Reasoning/340dea8a-103d-45f5-8ee9-2f2409e86042_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vamba_[[:space:]]Understanding[[:space:]]Hour-Long[[:space:]]Videos[[:space:]]with[[:space:]]Hybrid[[:space:]]Mamba-Transformers/41032422-fba8-4ef6-8117-de06ce1069a8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Variance-Based[[:space:]]Pruning[[:space:]]for[[:space:]]Accelerating[[:space:]]and[[:space:]]Compressing[[:space:]]Trained[[:space:]]Networks/5aed5135-0f01-4c2d-b7ee-d3f1987cb1ed_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vector[[:space:]]Contrastive[[:space:]]Learning[[:space:]]For[[:space:]]Pixel-Wise[[:space:]]Pretraining[[:space:]]In[[:space:]]Medical[[:space:]]Vision/b863182b-c21f-4dc0-8b6a-73de7593a982_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VehicleMAE_[[:space:]]View-asymmetry[[:space:]]Mutual[[:space:]]Learning[[:space:]]for[[:space:]]Vehicle[[:space:]]Re-identification[[:space:]]Pre-training[[:space:]]via[[:space:]]Masked[[:space:]]AutoEncoders/ec3290f4-caa9-4573-afb5-3440940874ee_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Verbalized[[:space:]]Representation[[:space:]]Learning[[:space:]]for[[:space:]]Interpretable[[:space:]]Few-Shot[[:space:]]Generalization/e86880d5-4a52-49e7-83a5-bbf2d31e9dd1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Versatile[[:space:]]Transition[[:space:]]Generation[[:space:]]with[[:space:]]Image-to-Video[[:space:]]Diffusion/f3fc723c-8a6e-4204-9849-5843b598e850_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VertexRegen_[[:space:]]Mesh[[:space:]]Generation[[:space:]]with[[:space:]]Continuous[[:space:]]Level[[:space:]]of[[:space:]]Detail/342c3a2f-65c0-4b35-b717-21b0c8a1a0ef_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViCTr_[[:space:]]Vital[[:space:]]Consistency[[:space:]]Transfer[[:space:]]for[[:space:]]Pathology[[:space:]]Aware[[:space:]]Image[[:space:]]Synthesis/08391b0d-1c48-48a7-aee4-a1aed1fb2b8c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViLLa_[[:space:]]Video[[:space:]]Reasoning[[:space:]]Segmentation[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Model/9374d1a1-c3e0-498b-83e7-5023233fa5ab_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViLU_[[:space:]]Learning[[:space:]]Vision-Language[[:space:]]Uncertainties[[:space:]]for[[:space:]]Failure[[:space:]]Prediction/9693ec7c-e63e-4549-b95a-8e863d6bee58_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViM-VQ_[[:space:]]Efficient[[:space:]]Post-Training[[:space:]]Vector[[:space:]]Quantization[[:space:]]for[[:space:]]Visual[[:space:]]Mamba/bb5b37a9-8d33-488a-b6b3-ef3b094ba666_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViSpeak_[[:space:]]Visual[[:space:]]Instruction[[:space:]]Feedback[[:space:]]in[[:space:]]Streaming[[:space:]]Videos/4c2298b6-a1cd-4d0c-bdd9-742253ff8150_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViT-EnsembleAttack_[[:space:]]Augmenting[[:space:]]Ensemble[[:space:]]Models[[:space:]]for[[:space:]]Stronger[[:space:]]Adversarial[[:space:]]Transferability[[:space:]]in[[:space:]]Vision[[:space:]]Transformers/9e6fb5c3-a13e-4a97-9aed-2462169214a0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViT-Linearizer_[[:space:]]Distilling[[:space:]]Quadratic[[:space:]]Knowledge[[:space:]]into[[:space:]]Linear-Time[[:space:]]Vision[[:space:]]Models/4d67ff7b-60ce-453d-9e8b-8fe983e07bca_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ViT-Split_[[:space:]]Unleashing[[:space:]]the[[:space:]]Power[[:space:]]of[[:space:]]Vision[[:space:]]Foundation[[:space:]]Models[[:space:]]via[[:space:]]Efficient[[:space:]]Splitting[[:space:]]Heads/62f8e79c-d973-4405-a368-6eb7d50f7246_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vid-Group_[[:space:]]Temporal[[:space:]]Video[[:space:]]Grounding[[:space:]]Pretraining[[:space:]]from[[:space:]]Unlabeled[[:space:]]Videos[[:space:]]in[[:space:]]the[[:space:]]Wild/b7f7fdbe-e75b-4823-a9b7-19405b49fb65_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video[[:space:]]Color[[:space:]]Grading[[:space:]]via[[:space:]]Look-Up[[:space:]]Table[[:space:]]Generation/b0b9c231-2082-4b4c-ad77-337ee0cc9045_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video[[:space:]]Individual[[:space:]]Counting[[:space:]]for[[:space:]]Moving[[:space:]]Drones/1cf4d87f-9c8a-44cd-a5ee-135b6f1daa18_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video[[:space:]]Motion[[:space:]]Graphs/78363dc2-8dfc-4b9b-80e6-1323ef2ac37e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video-T1_[[:space:]]Test-time[[:space:]]Scaling[[:space:]]for[[:space:]]Video[[:space:]]Generation/1019b1f8-19d0-4e70-9f39-8bbdb082c721_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Video2BEV_[[:space:]]Transforming[[:space:]]Drone[[:space:]]Videos[[:space:]]to[[:space:]]BEVs[[:space:]]for[[:space:]]Video-based[[:space:]]Geo-localization/1c533f62-65f2-4dff-a54f-3aa8e0de6fdc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoAds[[:space:]]for[[:space:]]Fast-Paced[[:space:]]Video[[:space:]]Understanding/8a639327-709f-4460-aabc-c77b069d813b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoAuteur_[[:space:]]Towards[[:space:]]Long[[:space:]]Narrative[[:space:]]Video[[:space:]]Generation/31b80179-4ba4-49ae-a3df-b51fb25c8cd6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoLLaMB_[[:space:]]Long[[:space:]]Streaming[[:space:]]Video[[:space:]]Understanding[[:space:]]with[[:space:]]Recurrent[[:space:]]Memory[[:space:]]Bridges/ef7f527a-2c19-4223-8b21-ac615c5af575_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoMiner_[[:space:]]Iteratively[[:space:]]Grounding[[:space:]]Key[[:space:]]Frames[[:space:]]of[[:space:]]Hour-Long[[:space:]]Videos[[:space:]]via[[:space:]]Tree-based[[:space:]]Group[[:space:]]Relative[[:space:]]Policy[[:space:]]Optimization/d7b4639d-d9a0-4b6c-986f-11dd15c86ebf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoOrion_[[:space:]]Tokenizing[[:space:]]Object[[:space:]]Dynamics[[:space:]]in[[:space:]]Videos/872b582e-2d26-46e1-93e2-773aad7bd594_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoRFSplat_[[:space:]]Direct[[:space:]]Scene-Level[[:space:]]Text-to-3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]Generation[[:space:]]with[[:space:]]Flexible[[:space:]]Pose[[:space:]]and[[:space:]]Multi-View[[:space:]]Joint[[:space:]]Modeling/0f70facc-7c84-4b86-b0dc-d3d1bdbeaad1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoSetDiff_[[:space:]]Identifying[[:space:]]and[[:space:]]Reasoning[[:space:]]Similarities[[:space:]]and[[:space:]]Differences[[:space:]]in[[:space:]]Similar[[:space:]]Videos/3d626d6f-d212-46b3-bc48-4c5f49732e8e_origin.pdf filter=lfs diff=lfs merge=lfs -text

2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/122157a9-e33d-4cdd-b0f9-d8b983f2a104_content_list.json

@@ -0,0 +1,1781 @@
+[
+    {
+        "type": "text",
+        "text": "VA-MoE: Variables-Adaptive Mixture of Experts for",
+        "text_level": 1,
+        "bbox": [
+            233,
+            130,
+            764,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Incremental Weather Forecasting",
+        "text_level": 1,
+        "bbox": [
+            308,
+            162,
+            651,
+            184
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/408a686bb8235a3b61e408da6aca9f7a13881c231183d5f5fbc8364f783e3092.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            653,
+            151,
+            687,
+            179
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Hao Chen $^{1\\dagger}$  Han Tao $^{1\\dagger}$  Guo Song $^{1*}$  Jie Zhang $^{1}$  Yonghan Dong $^{2}$  Yunlong Yu $^{3}$  Lei Bai $^{4}$ $^{1}$ Hong Kong University of Science and Technology (HKUST)  $^{2}$ Huawei Technologies Ltd.  \n $^{3}$ Zhejiang University  $^{4}$ Shanghai AI Laboratory",
+        "bbox": [
+            137,
+            210,
+            893,
+            265
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Equal contribution *Corresponding author {hchener, thanad}@connect.ust.hk",
+        "bbox": [
+            181,
+            268,
+            841,
+            282
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{songguo, csejzhang}@ust.hk dongyonghan@huawei.com yuyunlong@zju.edu.cn bailei@pjlab.org.cn",
+        "bbox": [
+            101,
+            285,
+            918,
+            301
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            335,
+            326,
+            351
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "This paper presents Variables-Adaptive Mixture of Experts (VA-MoE), a novel framework for incremental weather forecasting that dynamically adapts to evolving spatiotemporal patterns in real-time data. Traditional weather prediction models often struggle with exorbitant computational expenditure and the need to continuously update forecasts as new observations arrive. VA-MoE addresses these challenges by leveraging a hybrid architecture of experts, where each expert specializes in capturing distinct sub-patterns of atmospheric variables (e.g., temperature, humidity, wind speed). Moreover, the proposed method employs a variable-adaptive gating mechanism to dynamically select and combine relevant experts based on the input context, enabling efficient knowledge distillation and parameter sharing. This design significantly reduces computational overhead while maintaining high forecast accuracy. Experiments on ERA5 dataset demonstrate that VA-MoE performs comparable against state-of-the-art models in both short-term (e.g., 1-3 days) and long-term (e.g., 5 days) forecasting tasks, with only about  $25\\%$  of trainable parameters and  $50\\%$  of the initial training data. Code: https://github.com/chenhao-zju/VAMoE",
+        "bbox": [
+            89,
+            369,
+            483,
+            704
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            91,
+            737,
+            220,
+            752
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Weather forecasting remains one of the most critical scientific challenges with profound socio-economic impacts. While traditional Numerical Weather Prediction (NWP) [2] relies on solving complex partial different equations to simulate atmospheric dynamics, recent advances in data-driven artificial intelligence [3, 4, 9, 22] have revolutionized earth system modeling. Deep learning models, in particular, have demonstrated remarkable potential in capturing diverse spatiotemporal patterns across weather and climate phenomena.",
+        "bbox": [
+            89,
+            763,
+            483,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/a597efb3a93059332d6e7ca7100ecc4ea802f35eb2a605ebaae8713017f4c997.jpg",
+        "image_caption": [
+            "Figure 1. Illustration of two different training paradigms of weather forecasting when new variables arrive: (c) Full-Train the whole model (d) Incremental-Train only the module corresponding new variables. (b) Performances of these two paradigms."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            333,
+            903,
+            488
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ena, giving rise to the emerging field of AI for Weather.",
+        "bbox": [
+            511,
+            566,
+            880,
+            583
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, a fundamental limitation of existing AI4Weather approaches lies in their assumption that all variables are available synchronously during training and inference. In reality, meteorological variables are often heterogeneous in terms of data sources, collection frequencies, and spatial-temporal distributions. For instance, the upper-air variables (e.g., temperature profiles) are sparse and sampled via radiosondes/satellites, while the surface variable (e.g., precipitation, wind) are dense but updated in near-real-time. This asynchrony poses significant challenges: when introducing new variables (e.g., satellite-derived aerosol data), existing models must be entirely retrained from scratch, incurring prohibitive computational costs. For example, Pangu [3] required 64 days and 192 V100 GPUs for a full retraining cycle.",
+        "bbox": [
+            509,
+            583,
+            906,
+            809
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address this challenge, we propose Incremental Weather Forecasting (IWF), a novel paradigm designed to dynamically expand meteorological forecast variables in response to increasing observational data. By enabling parameter-efficient variable scaling, IWF revolutionizes operational meteorology by offering a streamlined alternative",
+        "bbox": [
+            511,
+            810,
+            908,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV 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": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "7915",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "to conventional models. This innovation tackles a critical limitation in the field: the need for scalable modeling solutions as datasets grow gradually, ensuring adaptability without compromising computational efficiency.",
+        "bbox": [
+            89,
+            90,
+            480,
+            151
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "A pivotal limitation of existing incremental approaches is catastrophic forgetting, i.e., when introducing new variables, pre-trained parameters drift towards new distributions, causing significant performance degradation on original variables. To overcome this, we present the Variables-Adaptive Mixture of Experts (VA-MoE) model, the first expert-based system explicitly designed for hierarchical atmospheric variables.",
+        "bbox": [
+            89,
+            154,
+            480,
+            273
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Mixture of Experts (MoE) models inherently enable scalable representation learning for multi-variable atmospheric datasets by dynamically expanding expert modules as input dimensions grow. While traditional MoE architectures excel in representational capacity, their parallelized design which distributes uniform weights across experts, often leads to homogeneous outputs, limiting specialized knowledge acquisition. To address this limitation, we introduce VA-MoE, an auxiliary-loss-free framework that decouples expert specialization from computational efficiency. VA-MoE employs a phased training strategy: Initial experts are pretrained on base variables, then incrementally expanded to accommodate new variables. Crucially, the pretrained experts are frozen during expansion phases to prevent catastrophic forgetting, ensuring robust retention of prior expertise while adapting to novel inputs. To foster expert diversity without auxiliary losses, VA-MoE incorporates variable index embeddings, i.e., positional metadata encodings that guide experts to develop domain-specific specialization. These embeddings dynamically activate contextually relevant computational pathways for each variable, optimizing resource allocation at inference time. By integrating variable index embeddings that encode spatial-temporal metadata, VA-MoE directs experts to specialize in specific atmospheric domains. This context-aware routing reduces inference latency compared to static MoE models while maintaining expert diversity.",
+        "bbox": [
+            91,
+            276,
+            483,
+            684
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Besides, we propose a gradient-scaled variable-adaptive loss function that dynamically aligns variable optimization rates with their inherent spatiotemporal characteristics. By quantifying the distinct temporal evolution patterns of atmospheric variables, our method allocates differential loss weights, i.e., larger gradient magnitudes prioritize fast-transient fields like temperature to capture immediate atmospheric dynamics, while gradual weight adjustments stabilize slow-changing fields such as geopotential height to preserve long-term system behavior.",
+        "bbox": [
+            89,
+            686,
+            480,
+            837
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As shown in Fig. 1, our paradigm employs a two-phase training paradigm: an initial phase optimizing base atmospheric variables, followed by an incremental phase integrating new variables while freezing previously trained ex",
+        "bbox": [
+            89,
+            839,
+            480,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "perts. Experiments with the same 40-year dataset reveal that our paradigm Fig. 1(a)+(d) achieves superior z500 predictions performance compared to full retraining Fig. 1(a)+(c).",
+        "bbox": [
+            511,
+            90,
+            903,
+            136
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In conclusion, the contributions of this work include:",
+        "bbox": [
+            532,
+            137,
+            883,
+            151
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- This work initiates systematic research on incremental learning paradigms for weather forecasting. We propose the quantitative benchmark to evaluate the trade-offs between model scalability and generalization in incremental weather modeling.",
+            "- We present Variables-Adaptive Mixture of Experts (VA-MoE), the first framework tailored for incremental atmospheric modeling. VA-MoE achieves expert specialization through contextual variable activation driven by variable index embeddings, enabling dynamic assignment of experts to variables during both training and inference.",
+            "- Extensive experiments on the ERA5 dataset demonstrate that VA-MoE achieves comparable performance for surface variables, while delivering superior accuracy in upper-air variables against the existing competitors under  $50\\%$  reduced dataset size and  $25\\%$  fewer parameters."
+        ],
+        "bbox": [
+            511,
+            152,
+            903,
+            393
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related work",
+        "text_level": 1,
+        "bbox": [
+            513,
+            409,
+            650,
+            425
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Data-Driven Weather Forecasting",
+        "text_level": 1,
+        "bbox": [
+            513,
+            435,
+            808,
+            450
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In recent decades, traditional NWP [2] methods have dominated the weather forecasting field due to the robust predictions and rigorous mathematical validation. However, NWP methods [26, 33] require training from scratch for new prediction, resulting in slow computation and high costs.",
+        "bbox": [
+            511,
+            457,
+            903,
+            532
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recent advances in deep learning have catalyzed a paradigm shift toward data-driven models for medium-range weather forecasting. Pioneering works like FourCastNet [22] leverage Fourier Neural Operators (FNOs) [24] to learn spatiotemporal weather patterns, while PanguWeather [3] employs a 3D vision transformer [13, 25] to model atmospheric dynamics. Subsequent innovations, including Neural Operators [5, 24, 38] and specialized architectures [9, 10, 15, 23, 31, 43, 44, 46], have achieved accuracy rivaling NWP at a fraction of computation.",
+        "bbox": [
+            511,
+            534,
+            905,
+            684
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "While existing data-driven models excel at fixed-variable forecasting after costly initial training, their inflexibility poses a critical barrier: retraining from scratch is required to incorporate new variables, rendering cross-domain adaptation impractical. In contrast, our work introduces incremental weather forecasting, a novel paradigm evolving spatiotemporal patterns in real-time data. By decoupling variable-specific expertise from shared dynamics, our approach supports dynamic expansion to new atmospheric parameters with minimal retraining overhead.",
+        "bbox": [
+            511,
+            685,
+            905,
+            837
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Mixture of Experts for Incremental Learning",
+        "text_level": 1,
+        "bbox": [
+            511,
+            847,
+            897,
+            864
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Advances in deep learning[6, 17] have spurred interest in parameter-efficient transfer strategies, where models adapt",
+        "bbox": [
+            511,
+            869,
+            903,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "7916",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/0f6c306f6dc259a2ccf6fca6af91be6c6aae1b11d49c857416c88197d46d07c4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            116,
+            92,
+            459,
+            180
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/124c88fa5df353a7a3ccd95e9e7a9a285c004e52ab5edbcdfdd0876865d9dbfd.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            462,
+            93,
+            880,
+            180
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/d77f605a7ffeb8dd32d6de34d598d180acb7152790dd370ed404951bb38af79d.jpg",
+        "image_caption": [
+            "Figure 2. Illustration of the incremental weather forecasting paradigm. (a), (b) are the initial and incremental stages, respectively. (c) presents the detailed structure in both initial and incremental stages. During the initial stage, the model is trained with upper-air variables(UV). In the incremental stage, the fire components are trained while the snow components remain frozen. The index embedding is trained by the Index Encoder and divided into six parts based on the upper-air variables (UV: Z, Q, U, V, T) and surface variables (SV)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            116,
+            181,
+            880,
+            401
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "to new domains by fine-tuning only a subset of parameters. A key paradigm, incremental learning, addresses the stability-plasticity dilemma: retaining knowledge of prior tasks while integrating new concepts. However, real-world data often involves domain shifts, such as variations in variable distributions or task objectives, leading to two dominant research threads. Class Incremental Learning (CIL) [16, 21, 35, 47], focuses on expanding classification tasks with new classes while Task Incremental Learning (TIL) [29, 34, 40, 42] adapts models to distinct task sequences.",
+        "bbox": [
+            88,
+            492,
+            480,
+            643
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "One of the most popular structures for addressing incremental learning tasks is the Mixture of Experts (MoE) model [19], which learns representations of new concepts through different experts. Due to its sparse architecture, many works have adopted the MoE structure to reduce inference costs and enhance model capacity [37, 41]. An early work introducing MoE to incremental learning is Expert Gate [1], which trains multiple backbones as different experts and assigns new domains to the relevant expert. Lifelong-MoE [11] leverages pretrained experts and gates to retain the representational knowledge of the training domain. In addition to these approaches, MoE is used as an adapter in MoE-Adapter [45], where adapters are attached to the main structure. These works have demonstrated the promising performance of MoE in visual and natural language incremental learning.",
+        "bbox": [
+            89,
+            645,
+            482,
+            885
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In contrast to MoE designed for fixed-variable systems,",
+        "bbox": [
+            109,
+            886,
+            482,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "our propose incremental learning of variable distributions by assigning distinct pressure-level groups to specialized experts. Unlike prior incremental learning literature that primarily focus on task or dataset-specific adaptation, our incremental paradigm targets variable-centric expansion, enabling models to adapt to evolving observational data without retraining from scratch.",
+        "bbox": [
+            511,
+            492,
+            906,
+            599
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Methodology",
+        "text_level": 1,
+        "bbox": [
+            511,
+            613,
+            648,
+            631
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our work focuses on a novel paradigm of incremental weather forecasting with VA-MoE. In this section, we first formally define the problem in Sec. 3.1. Then, we present the structure of VA-MoE, consisting with the Channel-Adaptive Expert (CAE) and the Shared Expert in Sec. 3.2. The model during incremental stage is introduced in Sec. 3.3. Finally, loss function is presented in Sec. 3.4.",
+        "bbox": [
+            511,
+            638,
+            906,
+            747
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Overview",
+        "text_level": 1,
+        "bbox": [
+            511,
+            756,
+            624,
+            771
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For the weather forecasting task, the AI-based model  $\\Phi$  aims to predict the future weather variables  $\\mathbf{X}^{t + 1}$  with the historical weather variables  $\\mathbf{X}^t$ , i.e.,  $\\mathbf{X}^{t + 1} = \\Phi (\\mathbf{X}^t)$ .",
+        "bbox": [
+            511,
+            777,
+            905,
+            824
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In the incremental weather forecasting task, the weather variables  $\\mathbf{X}^t$  are divided into two sets: (i) the initial training variables at  $t$  time  $\\mathbf{X}_h^t \\in \\mathbb{R}^{H \\times W \\times N}$ , and, (ii) the incremental variables at  $t$  time  $\\mathbf{X}_{sv}^t \\in \\mathbb{R}^{H \\times W \\times M}$ . Unlike traditional weather forecasting tasks, where models are trained",
+        "bbox": [
+            511,
+            825,
+            906,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "7917",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/ec231ab293262e1e38372cb270935bd9ebecfb3ad7ef96b9c46580860695194a.jpg",
+        "image_caption": [
+            "Figure 3. Illustration of the VA-MoE. In the left subgraph, the input features comprise both upper-air and surface variables. During the initial training stage, five distinct CAE modules process the upper-air variables. In the incremental stage, surface variables are handled by a dedicated module,  $\\mathrm{CAESV}$ , while the original five CAEs remain frozen to preserve learned representations. The right subgraph details the CAE module. The Index Embedding is selected through a multi-layer process and applied to weight the Input Embedding via multiplication. Initially discrete, the index embedding is transformed into a continuous representation after passing through the GateEmbed layer, enabling it to dynamically reflect the relevance of specific channels to the input variables."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            93,
+            88,
+            480,
+            282
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "with all variables available simultaneously, the incremental weather forecasting task sequentially incorporates dynamic variables as they become available over time. The incremental weather forecasting process is structured into a sequence of learning stages: an initial training stage followed by subsequent training stages.",
+        "bbox": [
+            89,
+            470,
+            483,
+            561
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In the initial training stage, the available variables of 40 years are used to train a weather model, i.e.,  $\\mathrm{X}_h^{t + 1} = \\Phi_h(\\mathrm{X}_h^t)$ , as shown in Fig. 2(a). In the incremental training stage, the half variables used at the initial stage and the surface variables at the incremental stage are incorporated together, to train the newly added modules for the surface variables while keeping the model parameters from the initial stage frozen, i.e.,  $\\mathrm{X}^{t + 1} = \\Phi (\\mathrm{X}^t)$ , where  $\\Phi$  consists of both the fixed parameters of the pre-trained model  $\\Phi_h$  and the trainable parameters of  $\\Phi_{sv}$ ,  $\\mathrm{X}^t$  consists of all the available variables at time  $t$ ,  $\\mathrm{X}^{t + 1}$  consists of the predicted variables at time  $t + 1$ , as shown in Fig. 2(b). In this work, the upper-air variables are utilized during the initial training stage while the surface variables become available and are incorporated during the incremental training stage. The upper-air variables include five different types, Z, Q, U, V, and T, each defined across 13 different levels.",
+        "bbox": [
+            89,
+            563,
+            483,
+            819
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. VA-MoE",
+        "text_level": 1,
+        "bbox": [
+            89,
+            830,
+            194,
+            845
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We introduce index embedding within transformer blocks, a novel mechanism to dynamically guide experts in learning hierarchical relationships between meteorological vari",
+        "bbox": [
+            89,
+            854,
+            483,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ables. As new variables are incrementally integrated, corresponding experts are added to the transformer architecture and optimized via index-based affinity assignments. Building on this framework, we adapt the Mixture-of-Experts (MoE) paradigm which enables task-specific specialization and flexible integration of new experts—into weather forecasting transformers.",
+        "bbox": [
+            511,
+            90,
+            903,
+            196
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The structure of the proposed VA-MoE is illustrated in Fig. 2(c). The input features, extracted by the encoder, first pass through a normalization layer and a self-attention layer. The output of the self-attention layer is then combined with the residual connection. Then, another normalization layer is applied, followed by a VA-MoE module, which is also equipped with a residual connection. This process can be formulated as:",
+        "bbox": [
+            511,
+            196,
+            906,
+            316
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathbf {X} _ {m i d} = \\mathbf {X} _ {i n} + \\operatorname {S A} (\\text {L a y e r N o r m} (\\mathbf {X} _ {\\text {i n}})), \\\\ \\mathbf {X} _ {i n} = \\mathbf {X} _ {i n - 1} + \\operatorname {V A M E} (\\mathbf {I} _ {i n - 1}, \\mathbf {N} _ {i n - 1} (\\mathbf {X} _ {i n})) \\end{array} \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            529,
+            328,
+            903,
+            353
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {X} _ {\\text {o u t}} = \\mathbf {X} _ {\\text {m i d}} + \\operatorname {V A - M o E} (\\operatorname {L a y e r N o r m} (\\mathbf {X} _ {\\text {m i d}})),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            534,
+            349,
+            854,
+            363
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{X}_{in}$  and  $\\mathbf{X}_{out}$  are the input and the output of each Transformer Block, respectively.",
+        "bbox": [
+            511,
+            375,
+            906,
+            406
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.1. Variables and Index Embedding",
+        "text_level": 1,
+        "bbox": [
+            511,
+            416,
+            784,
+            431
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As illustrated in Fig. 3, during the training stage, the input feature  $\\mathbf{X}_h^t\\in \\mathbb{R}^{H\\times W\\times N}$  consists of five different types of variables:  $\\mathbf{X}_Z^t,\\mathbf{X}_Q^t,\\mathbf{X}_U^t,\\mathbf{X}_V^t,\\mathbf{X}_T^t\\in \\mathbb{R}^{H\\times W\\times (N / 5)}$ , where,  $H\\times W$  denotes the spatial resolution. These five types of variables are concatenated and processed by a shared expert and five distinct Channel-Adaptive Experts (CAEs).",
+        "bbox": [
+            511,
+            435,
+            906,
+            525
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In addition to these variables, the framework introduces an extra one-hot index embedding  $\\mathbf{I}_h\\in \\mathbb{R}^{5\\times N}$  to guide experts in learning variable affinity. The number  $N$  corresponds to the number of variables, and 5 corresponds to the number of variable types. Similar to the input weather features, the index embedding is composed of five separate embeddings:  $\\mathbf{I}_Z,\\mathbf{I}_Q,\\mathbf{I}_U,\\mathbf{I}_V,\\mathbf{I}_T\\in \\mathbb{R}^{1\\times 1\\times N}$ . This initial index embedding is then encoded into a latent space using a Linear layer, expressed as  $\\mathbf{I}_h = \\mathrm{Linear}(\\mathbf{I}_h)$ .",
+        "bbox": [
+            511,
+            527,
+            905,
+            662
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.2. Channel-Adaptive Expert (CAE)",
+        "text_level": 1,
+        "bbox": [
+            511,
+            671,
+            790,
+            686
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The key design of VA-MoE is the channel-adaptive expert, as shown in Fig. 3. For each variable  $Z$ , the corresponding expert module processes both the feature embeddings of the input variables  $\\mathbf{X}_h^t$  and the index embedding  $\\mathbf{I}_Z$ , which is formulated as:",
+        "bbox": [
+            511,
+            691,
+            905,
+            765
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {X} _ {Z} ^ {t, C A E} = \\operatorname {C A E} _ {Z} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {Z}\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            614,
+            777,
+            903,
+            797
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathrm{CAE}_Z$  is the abbreviation of channel-adaptive expert module for variable  $Z$ .",
+        "bbox": [
+            511,
+            809,
+            903,
+            838
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In the CAE module for variable  $Z$ , the index embedding  $\\mathbf{I}_Z$  is encoded through a linear layer for feature projection and then multiplied channel-wise with the input embedding  $\\mathbf{X}_h^t$ . The resulting fused feature is processed",
+        "bbox": [
+            511,
+            839,
+            905,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "7918",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "by the GateEmbed layer, which produces a gate embedding to enhance the fusion of the index and variable embeddings. This enhanced feature is then normalized using a SoftMax layer and filtered through a TopK operation, which selects the top-K high-rank channels. This process yields the GateIndex  $\\mathrm{GI}_Z\\in \\mathbb{R}^{H\\times W\\times K}$  and GateWeight  $\\mathrm{GW}_Z\\in \\mathbb{R}^{H\\times W\\times K}$ . The process is formulated as:",
+        "bbox": [
+            89,
+            90,
+            483,
+            196
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {I} _ {Z} ^ {t o p k}, \\mathbf {W} _ {Z} ^ {t o p k} = \\mathrm {T O P} _ {k} (\\operatorname {S o f t M a x} (\\operatorname {M L P} _ {Z} (\\mathbf {X} _ {h} ^ {t} \\odot \\mathbf {I} _ {Z}))), (3)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            99,
+            208,
+            483,
+            229
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{I}_Z^{topk}$  and  $\\mathbf{W}_Z^{topk}$  denote the index and weight of the Top-k features. Once the GateIndex and GateWeight are computed, the input embedding is selectively processed based on the GateIndex, while the variable embedding is weighted channel-wise according to the GateWeight. The GateEmbed layer uses variable embeddings to guide the index embedding, enabling the creation of a variable-specific confidence matrix. Within the CAE module, this embedding enhances the confidence matrix by incorporating both semantic and positional information, rather than relying solely on positional information.",
+        "bbox": [
+            89,
+            239,
+            483,
+            407
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The selected features for variable  $Z$  are obtained with:",
+        "bbox": [
+            109,
+            409,
+            470,
+            422
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {X} _ {Z} ^ {t, s e l e c t e d} = \\mathbf {W} _ {Z} ^ {t o p k} * \\operatorname {S e l e c t} _ {Z} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {Z} ^ {t o p k}\\right), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            140,
+            435,
+            482,
+            455
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathrm{Select}_Z$  denotes selecting Top-K embedding from the input feature  $\\mathbf{X}_h^t$  within the  $\\mathrm{CAE}_Z$  module.",
+        "bbox": [
+            89,
+            467,
+            482,
+            497
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The selected and weighted embedding captures the key features associated with the specific type of variables. In the  $\\mathrm{CAE}_Z$  module, these features serve as the training set to model the distribution of the variable  $Z$  with  $\\mathrm{Expert}_Z$ . The expert module adopts a parallel structure, splitting the channels into multiple parts and enhancing each part with a sparse MoE. This process is formulated as:",
+        "bbox": [
+            89,
+            498,
+            483,
+            603
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {X} _ {Z} ^ {t, C A E} = \\operatorname {E x p e r t} _ {Z} (\\mathbf {X} _ {Z} ^ {t, s e l e c t e d}), \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            174,
+            614,
+            482,
+            636
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{Expert}_Z$  is implemented with a multiple layer.",
+        "bbox": [
+            89,
+            647,
+            452,
+            662
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2.3. Shared Expert.",
+        "text_level": 1,
+        "bbox": [
+            89,
+            672,
+            243,
+            686
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The shared expert processes the overall features of all variables and operates in parallel with the variable-specific modules. Each variable is handled by its corresponding CAE module, and their outputs are summarized to produce the fused features  $\\mathbf{X}_h^{t,fused}$  for all variables. Since the channel count of the selected variable embeddings is lower than that of the original embeddings, an additional up-channel linear layer,  $\\mathrm{Linear}_{\\mathrm{up}}$ , is introduced. The fused features and the output from the shared expert are then combined via pixel-wise addition. This process is expressed as:",
+        "bbox": [
+            89,
+            691,
+            483,
+            843
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {X} _ {h} ^ {t, f u s e d} = \\operatorname {C A E} _ {Z} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {Z}\\right) + \\dots + \\operatorname {C A E} _ {T} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {T}\\right), \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            101,
+            853,
+            482,
+            873
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left(\\mathbf {X} ^ {\\prime}\\right) _ {h} ^ {t} = \\operatorname {E x p e r t} _ {s h a r e d} \\left(\\mathbf {X} _ {h} ^ {t}\\right) + \\operatorname {L i n e a r} _ {u p} \\left(\\mathbf {X} _ {h} ^ {t, f u s e d}\\right). \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            107,
+            882,
+            482,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3. Incremental Learning Stage",
+        "text_level": 1,
+        "bbox": [
+            513,
+            90,
+            767,
+            107
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "When new variables are incrementally introduced to the model, specialized experts are dynamically integrated into the transformer blocks to process these variables. Concurrently, both the encoder and decoder architectures undergo synchronized updates to maintain compatibility with the expanded variable set.",
+        "bbox": [
+            511,
+            112,
+            906,
+            203
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3.1. Encoder and Decoder Modules.",
+        "text_level": 1,
+        "bbox": [
+            511,
+            210,
+            779,
+            224
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "When  $K$  new variables are introduced incrementally, the encoder's input dimension expands to  $H \\times W \\times (N + M)$ , necessitating an adjustment to the convolutional layer's kernel from  $(3, 3, N, C)$  to  $(3, 3, N + M, C)$ . To accommodate this, the layer's parameters are partitioned into two components: The  $(3, 3, N, C)$  segment retains pretrained weights from earlier stages, ensuring continuity. The  $(3, 3, M, C)$  segment, corresponding to the new variables, is initialized randomly to learn emergent patterns.",
+        "bbox": [
+            511,
+            229,
+            906,
+            366
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The same strategy is applied to the position embedding and the Decoder module.",
+        "bbox": [
+            511,
+            367,
+            905,
+            396
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3.2. VA-MoE Module.",
+        "text_level": 1,
+        "bbox": [
+            513,
+            405,
+            681,
+            419
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "During the incremental learning phase,  $M$  new experts are dynamically integrated into the MoE layer of each transformer block to accommodate  $M$  newly introduced variables. In this phase, only the new experts and the shared expert are actively trained, while all preexisting experts remain frozen to preserve previously learned knowledge.",
+        "bbox": [
+            511,
+            424,
+            905,
+            515
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3.3. Index Embedding Module.",
+        "text_level": 1,
+        "bbox": [
+            513,
+            523,
+            745,
+            537
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "When  $M$  new variables become available during the incremental stage, the index embedding changes from  $\\mathbf{I}_h\\in$ $\\mathbb{R}^{(N\\times l)\\times N}$  to  $\\mathbf{I}_h\\in \\mathbb{R}^{(N\\times l + M\\times r)\\times (N + M)}$ , where each  $M$  new variables has  $r$  levels. Additionally, the index encoder needs to be retrained.",
+        "bbox": [
+            511,
+            542,
+            906,
+            617
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4. Objective Function",
+        "text_level": 1,
+        "bbox": [
+            513,
+            628,
+            696,
+            643
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To train the model, we introduce a novel dynamic prediction loss combined with a reconstruction loss.",
+        "bbox": [
+            511,
+            650,
+            905,
+            680
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4.1. Dynamic Prediction Loss.",
+        "text_level": 1,
+        "bbox": [
+            511,
+            689,
+            736,
+            704
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The prediction loss aims to minimize the difference between the model's predicted results and the actual future weather variables. Unlike existing weather forecasting methods that treat all the variables equally, we argue that different variables, such as temperature and surface pressure, follow distinct distributions. Thus, we introduce a novel dynamic prediction loss that dynamically assigns a weight to each channel of the input data. This loss is formulated as:",
+        "bbox": [
+            511,
+            708,
+            905,
+            829
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nO b j _ {p r e d} = \\left(\\hat {\\mathbf {X}} ^ {t + 1} - \\mathbf {X} ^ {t + 1}\\right) \\odot \\left(\\hat {\\mathbf {X}} ^ {t + 1} - \\mathbf {X} ^ {t + 1}\\right) / e ^ {\\mathbf {w}} + \\mathbf {w}, (8)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            519,
+            839,
+            906,
+            859
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\odot$  denotes Hadamard product,  $e$  denotes the base of the natural logarithm,  $\\mathbf{X}^{t + 1}\\in \\mathbb{R}^{H\\times W\\times C}$ , and  $\\mathbf{w}\\in$",
+        "bbox": [
+            511,
+            869,
+            905,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "7919",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/041e988ddd89d40a80ac5822062df3d6dd06335024de3d68c9678d208845afb6.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Name</td><td>Description</td><td>Levels</td></tr><tr><td colspan=\"3\">Upper-Air</td></tr><tr><td>Z</td><td>Geopotential</td><td>13</td></tr><tr><td>Q</td><td>Specific humidity</td><td>13</td></tr><tr><td>U</td><td>x-direction wind</td><td>13</td></tr><tr><td>V</td><td>y-direction wind</td><td>13</td></tr><tr><td>T</td><td>Temperature</td><td>13</td></tr><tr><td colspan=\"3\">Surface-Incremental Learning</td></tr><tr><td>u10</td><td>x-direction wind at 10m height</td><td>Single</td></tr><tr><td>v10</td><td>y-direction wind at 10m height</td><td>Single</td></tr><tr><td>t2m</td><td>Temperature at 2m height</td><td>Single</td></tr><tr><td>msl</td><td>Mean sea-level pressure</td><td>Single</td></tr><tr><td>sp</td><td>Surface pressure</td><td>Single</td></tr></table>",
+        "bbox": [
+            125,
+            88,
+            455,
+            313
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "$\\mathbb{R}^{1\\times 1\\times C}$  is a learnable vector. With this loss function, the model could dynamically prioritize variables based on their distributional characteristics.",
+        "bbox": [
+            89,
+            373,
+            483,
+            417
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.4.2. Reconstruction Loss.",
+        "text_level": 1,
+        "bbox": [
+            89,
+            428,
+            282,
+            441
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We introduce an additional reconstruction loss that connects the encoder and decoder directly, formulated as:",
+        "bbox": [
+            89,
+            446,
+            483,
+            477
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nO b j _ {r e c o n} = \\left(\\hat {\\mathbf {X}} ^ {t} - \\mathbf {X} ^ {t}\\right) \\odot \\left(\\hat {\\mathbf {X}} ^ {t} - \\mathbf {X} ^ {t}\\right), \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            156,
+            487,
+            482,
+            506
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\hat{\\mathbf{X}}^t = \\mathrm{Dec}(\\mathrm{Enc}(\\mathbf{X}^t))$  represents the reconstructed variables. With the loss function, the encoder and decoder modules specialize in encoding and decoding features, while the intermediate transformer blocks focus on learning the data distribution. This ensures that the decoding process is entirely handled by decoder, allowing the intermediate blocks to focus on their task and not participate in decoding, thereby optimizing network efficiency.",
+        "bbox": [
+            89,
+            516,
+            483,
+            638
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To this end, the model is trained with the final objective function, a combination of the dynamic prediction loss and reconstruction loss. The final objective is formulated as:",
+        "bbox": [
+            89,
+            638,
+            483,
+            684
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nO b j _ {f i n a l} = O b j _ {p r e d} + \\lambda O b j _ {r e c o n}, \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            166,
+            696,
+            482,
+            714
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda$  denotes a hyper-parameter.",
+        "bbox": [
+            89,
+            724,
+            328,
+            739
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            753,
+            223,
+            771
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Dataset. In this work, we conduct experiments on a popular weather dataset, i.e., ERA5[18], provided by the ECMWF. ERA5 dataset is a reanalysis atmospheric dataset, consisting of the atmospheric variables from 1979 to the present day with a  $0.25^{\\circ}$  spatial resolution with  $721 \\times 1440$ . We train the model on 40-year dataset in the initial stage and continuously train the model on 20-year dataset in the incremental stage, which contains the weather variables from",
+        "bbox": [
+            89,
+            780,
+            483,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5c32d76ef3271d634424358f221973a4c9a674afb759538b61aacedf3e343fc8.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            90,
+            705,
+            218
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/a0d1aa500d574a7e95efd8c0c9b28ef1e85a3a448b254b80980622a66a0bc2a2.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            714,
+            90,
+            903,
+            218
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/052ad5c9ca6a4a26d4335087addf28eac50c9b1211888b4204be97c3469dbdba.jpg",
+        "image_caption": [
+            "Figure 4. Comparative analysis of RMSE  $\\downarrow$  across 10 data-driven models for four variables, including Z500 and T850 (upper-air variables) in the initial stage, as well as T2M and U10 (incremental surface variables) in the incremental stage."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            219,
+            705,
+            344
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/bd3379cdf299482f4eddfea62f542bb052f9854ee8afec711d077e1cd24bf117.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            714,
+            219,
+            903,
+            344
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/5c27927caadf954f2ab0a6c58768eeb365f2cd7f4cb19f217f1e1d4f1c9d5cd4.jpg",
+        "table_caption": [
+            "Table 1. A summary of atmospheric variables. The 13 levels are 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000 hPa. 'Single' denotes the variables under earth's surface."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>——ClimaX</td><td>——EWMoE</td><td>——WeatherGFT</td><td>——Stormer</td></tr><tr><td>——FourCastNet</td><td>——ClimODE</td><td>——GraphCast</td><td>——Ours(IL)</td></tr><tr><td>——Panguo-Weather</td><td>——Keisler</td><td></td><td></td></tr></table>",
+        "bbox": [
+            540,
+            345,
+            887,
+            375
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "1979 to 2020 and 2000 to 2020 year, respectively. And we test the model on one-year dataset of weather variables in 2021. In this work, the model processes 5 upper-air variables and 5 surface variables as Tab. 1, where the upper-air variables are exploited in the training stage, and the surface variables are exploited in the incremental stage.",
+        "bbox": [
+            511,
+            455,
+            905,
+            546
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. The main structure of this work follows the novel backbones [7, 8, 12]. We apply the AdamW optimizer with 0.0002 and 0.00005 learning rates to initial and incremental stages, respectively. In two stages, we train 100 epochs and set batch size to 16. Our model is trained with PyTorch using 16 A100 GPUs.",
+        "bbox": [
+            511,
+            547,
+            905,
+            638
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Main Results",
+        "text_level": 1,
+        "bbox": [
+            511,
+            650,
+            651,
+            664
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Fig. 4 presents a quantitative comparison of our structure with existing state-of-the-art approaches, including ClimaX [27], Pangu-Weather [3], ClimODE [36], WeatherGFT [39], FourCastNet [22], EWMoE [14], Keisler [20], GraphCast [23], and Stormer [28], for the 5-day weather prediction task. All the competitors are trained with both upper-air and surface variables on a  $0.25^{\\circ}$  weather dataset with a resolution of  $721 \\times 1440$ .",
+        "bbox": [
+            511,
+            672,
+            906,
+            792
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For the surface variables, including T2M and U10, our VA-MoE demonstrates performance similar to the leading methods Stormer [28] and Graphcast [23]. Compared to other competitors, our structure shows significant advantages in both short-term and long-term predictions. For the upper-air variable Z500, our structure delivers one of the best performances among all 10 methods, outperforming",
+        "bbox": [
+            511,
+            795,
+            905,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "7920",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/e1e47cf80a602d88639422c583130b60dbcfc1b6d6cbe0dc96828b9436f47f10.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td rowspan=\"2\">Dataset(years)</td><td rowspan=\"2\">Iteration\\( (×10^4) \\)</td><td colspan=\"3\">T2M(K)↓</td><td colspan=\"3\">U10(m/s)↓</td><td colspan=\"3\">V10(m/s)↓</td><td colspan=\"3\">MSL(Pa)↓</td><td colspan=\"3\">SP(Pa)↓</td></tr><tr><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td></tr><tr><td colspan=\"18\">Plain Training</td></tr><tr><td>ViT* [13]</td><td>1979-2020</td><td>40</td><td>0.72</td><td>1.35</td><td>1.86</td><td>0.66</td><td>1.98</td><td>3.01</td><td>0.68</td><td>2.02</td><td>3.11</td><td>40.2</td><td>208.5</td><td>393.9</td><td>63.3</td><td>222.1</td><td>397.0</td></tr><tr><td>IFS [32]</td><td>1979-2020</td><td>40</td><td>1.09</td><td>1.38</td><td>1.74</td><td>0.96</td><td>1.87</td><td>2.78</td><td>0.99</td><td>1.93</td><td>2.87</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Pangu-Weather [3]</td><td>1979-2020</td><td>40</td><td>0.82</td><td>1.09</td><td>1.53</td><td>0.77</td><td>1.63</td><td>2.54</td><td>0.79</td><td>1.68</td><td>2.65</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>FourCastNet [22]</td><td>1979-2020</td><td>40</td><td>0.82</td><td>1.02</td><td>1.77</td><td>0.82</td><td>2.08</td><td>3.34</td><td>0.84</td><td>2.11</td><td>3.41</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>ClimaX [27]</td><td>1979-2020</td><td>40</td><td>1.11</td><td>1.47</td><td>1.83</td><td>1.04</td><td>2.02</td><td>2.79</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Graphcast [23]</td><td>1979-2020</td><td>40</td><td>0.51</td><td>0.94</td><td>1.37</td><td>0.38</td><td>1.51</td><td>2.37</td><td>-</td><td>-</td><td>-</td><td>23.4</td><td>135.2</td><td>278.2</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Fengwu [9]</td><td>1979-2020</td><td>40</td><td>0.58</td><td>1.03</td><td>1.41</td><td>0.42</td><td>1.53</td><td>2.32</td><td>-</td><td>-</td><td>-</td><td>23.2</td><td>137.1</td><td>276.9</td><td>-</td><td>-</td><td>-</td></tr><tr><td>FuXi [10]</td><td>1979-2020</td><td>40</td><td>0.55</td><td>0.99</td><td>1.41</td><td>0.42</td><td>1.50</td><td>2.36</td><td>0.43</td><td>1.54</td><td>2.44</td><td>27.2</td><td>136.7</td><td>282.9</td><td>-</td><td>-</td><td>-</td></tr><tr><td>VA-MoE</td><td>1979-2020</td><td>40</td><td>0.57</td><td>1.03</td><td>1.42</td><td>0.43</td><td>1.41</td><td>2.25</td><td>0.44</td><td>1.46</td><td>2.34</td><td>27.5</td><td>131.1</td><td>275.9</td><td>57.1</td><td>168.9</td><td>302.4</td></tr><tr><td colspan=\"18\">Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (79-20)</td></tr><tr><td>VA-MoE (IL)</td><td>1979-2020</td><td>20</td><td>0.58</td><td>1.05</td><td>1.45</td><td>0.48</td><td>1.47</td><td>2.33</td><td>0.47</td><td>1.54</td><td>2.41</td><td>27.9</td><td>137.3</td><td>281.6</td><td>59.3</td><td>173.4</td><td>312.4</td></tr><tr><td colspan=\"18\">Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (00-20)</td></tr><tr><td>VA-MoE (IL)</td><td>2000-2020</td><td>10</td><td>0.73</td><td>1.17</td><td>1.57</td><td>0.54</td><td>1.58</td><td>2.49</td><td>0.55</td><td>1.63</td><td>2.57</td><td>30.0</td><td>148.8</td><td>304.7</td><td>60.6</td><td>171.4</td><td>314.8</td></tr></table>",
+        "bbox": [
+            96,
+            88,
+            906,
+            345
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/6c48125c0977a9f970ee7a1f1510428c67d7c7c44508d3eb463f98acbe402368.jpg",
+        "table_caption": [
+            "Table 2. Prediction performances on Incremental Training with 5 surface variables, i.e., T2M, U10, V10, MSL, and SP. * denotes running by ourselves. The best results are marked in bold. All experiments are in  ${0.25}^{ \\circ  }$  with  ${721} \\times  {1440}$  resolutions."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td rowspan=\"2\">Para. (M)</td><td colspan=\"3\">Z500(m2/s2) ↓</td><td colspan=\"3\">Q500(×e-3, g/kg) ↓</td><td colspan=\"3\">U500(m/s) ↓</td><td colspan=\"3\">V500(m/s) ↓</td><td colspan=\"3\">T500(K) ↓</td></tr><tr><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td></tr><tr><td>IFS [32]</td><td>-</td><td>28.31</td><td>154.08</td><td>333.96</td><td>0.31</td><td>0.61</td><td>0.75</td><td>1.43</td><td>3.23</td><td>5.12</td><td>1.40</td><td>3.58</td><td>5.64</td><td>0.36</td><td>0.98</td><td>1.70</td></tr><tr><td>Pangu-Weather [3]</td><td>-</td><td>24.88</td><td>167.90</td><td>391.26</td><td>0.25</td><td>0.55</td><td>0.69</td><td>0.96</td><td>3.13</td><td>4.73</td><td>0.91</td><td>3.52</td><td>5.15</td><td>0.27</td><td>0.94</td><td>1.56</td></tr><tr><td>Graphcast [23]</td><td>-</td><td>15.23</td><td>125.42</td><td>275.35</td><td>-</td><td>-</td><td>-</td><td>0.77</td><td>2.86</td><td>4.49</td><td>0.74</td><td>2.92</td><td>4.67</td><td>0.23</td><td>0.87</td><td>1.48</td></tr><tr><td>VA-MoE</td><td>665</td><td>19.28</td><td>134.63</td><td>295.52</td><td>0.17</td><td>0.49</td><td>0.62</td><td>0.84</td><td>2.99</td><td>4.71</td><td>0.84</td><td>3.04</td><td>4.89</td><td>0.25</td><td>0.76</td><td>1.36</td></tr><tr><td colspan=\"17\">Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (00-20)</td></tr><tr><td>VA-MoE (IL)</td><td>137</td><td>18.23</td><td>133.14</td><td>292.63</td><td>0.17</td><td>0.49</td><td>0.61</td><td>0.84</td><td>3.01</td><td>4.74</td><td>0.84</td><td>3.51</td><td>4.93</td><td>0.25</td><td>0.76</td><td>1.37</td></tr></table>",
+        "bbox": [
+            96,
+            397,
+            906,
+            532
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Prediction performances on Initial Training with 5 upper-air variables. All experiments are in  ${0.25}^{ \\circ  }$  with  ${721} \\times  {1440}$  resolutions.",
+        "bbox": [
+            91,
+            542,
+            903,
+            556
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Pangu-Weather [3] and Stormer [28], and significantly surpassing ClimaX [27] and FourCastNet [22], while slightly inferior to GraphCast [23]. For T850, our structure is slightly inferior to EWMoE [14], GraphCast [23], and Stormer [28] in the short-term predictions within 48 hours. After that period, our structure achieves the best performance in the long-term predictions.",
+        "bbox": [
+            88,
+            583,
+            482,
+            686
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Results of Surface Variables Prediction. This study evaluates model performance on five surface variables: T2M, U10, V10, MSL, and SP. While competing structures are trained with both upper-air and surface variables, VA-MoE (IL) employs a two-stage training approach: initial training with upper-air variables followed by incremental training with all five surface variables.",
+        "bbox": [
+            88,
+            688,
+            482,
+            792
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Analysis of the results from Tab. 2 demonstrates that VA-MoE outperforms competing models across multiple variables and settings, particularly in U10 and V10. Although VA-MoE shows marginally weaker performance in T2M compared to GraphCast [23] and comparable results to FengWu [9] and FuXi [10], it achieves superior long-term predictions for U10, V10, MSL, and SP. These indicate",
+        "bbox": [
+            88,
+            795,
+            482,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "that while VA-MoE exhibits slight limitations in short-term forecasting, it excels in long-term prediction tasks.",
+        "bbox": [
+            511,
+            583,
+            903,
+            613
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "When implementing the incremental training strategy, VA-MoE (IL) trained on a 40-year dataset with only half the training iterations achieves performance comparable to standard VA-MoE. Even with a reduced 20-year dataset, VA-MoE(IL) maintains acceptable performance using just a quarter of the iterations required by the standard approach. These results underscore the efficacy of incremental learning for atmospheric datasets and suggest its potential to replace conventional training methods in operational settings.",
+        "bbox": [
+            509,
+            614,
+            905,
+            750
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Results of Upper-Air Variables Prediction. This study evaluates model performance on upper-air variables, comparing three training approaches: (1) VA-MoE, trained exclusively on upper-air variables; (2) VA-MoE (IL), which incorporates incremental training on surface variables after initial upper-air training; and (3) competing models trained simultaneously on both upper-air and surface variables.",
+        "bbox": [
+            509,
+            750,
+            905,
+            854
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As shown in Tab. 3, VA-MoE achieves performance comparable to GraphCast [23] and outperforms IFS [32] and Pangu-Weather [3] across both short- and long-term",
+        "bbox": [
+            511,
+            854,
+            903,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "7921",
+        "bbox": [
+            482,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/a89841f541b62779568827b555ddd995ada9597a86b689891af324f299c2b25e.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td rowspan=\"2\">Para. (M)</td><td colspan=\"3\">Z500 (m2/s2) ↓</td><td colspan=\"3\">Q500 (×e-3, g/kg) ↓</td><td colspan=\"3\">U500 (m/s) ↓</td><td colspan=\"3\">V500 (m/s) ↓</td><td colspan=\"3\">T500 (K) ↓</td></tr><tr><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td></tr><tr><td>ViT*[13]</td><td>307</td><td>33.38</td><td>209.4</td><td>517.81</td><td>0.22</td><td>0.61</td><td>1.06</td><td>1.24</td><td>3.66</td><td>6.52</td><td>1.22</td><td>3.76</td><td>7.41</td><td>0.42</td><td>1.18</td><td>2.40</td></tr><tr><td>ViT+MoE (light)*[30]</td><td>609</td><td>37.92</td><td>207.11</td><td>405.73</td><td>0.22</td><td>0.60</td><td>0.78</td><td>1.30</td><td>3.84</td><td>5.87</td><td>1.27</td><td>3.89</td><td>6.11</td><td>0.46</td><td>1.23</td><td>2.02</td></tr><tr><td>ViT+MoE* [30]</td><td>1113</td><td>28.31</td><td>169.61</td><td>356.02</td><td>0.23</td><td>0.56</td><td>0.72</td><td>1.21</td><td>3.46</td><td>5.44</td><td>1.23</td><td>3.54</td><td>5.69</td><td>0.35</td><td>1.07</td><td>1.83</td></tr><tr><td>VA-MoE</td><td>665</td><td>20.59</td><td>139.02</td><td>302.13</td><td>0.18</td><td>0.49</td><td>0.62</td><td>0.91</td><td>3.02</td><td>4.76</td><td>0.91</td><td>3.08</td><td>4.97</td><td>0.27</td><td>0.92</td><td>1.59</td></tr><tr><td>VA-MoE (IL)</td><td>137</td><td>20.29</td><td>138.52</td><td>301.41</td><td>0.18</td><td>0.50</td><td>0.63</td><td>0.91</td><td>3.04</td><td>4.79</td><td>0.91</td><td>3.10</td><td>5.03</td><td>0.27</td><td>0.93</td><td>1.60</td></tr></table>",
+        "bbox": [
+            98,
+            89,
+            903,
+            196
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Architectural impact on 5 upper-air variables under  ${500}\\mathrm{{hPa}}$  . All experiments are in  ${1.5}^{ \\circ  }$  with  ${128} \\times  {256}$  resolutions.",
+        "bbox": [
+            127,
+            205,
+            866,
+            220
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/7a3f0e25dc8da0ded62ed82c5f1ef56cb67792b1eb0e64e3699a3a53e9edaebc.jpg",
+        "image_caption": [
+            "Figure 5. 6-hour global weather prediction of upper-air and surface variables forecast visualization generated by VA-MoE framework."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            223,
+            875,
+            455
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "prediction. Notably, VA-MoE (IL) demonstrates robust predictive capabilities for upper-air variables even after incremental training on surface variables, despite using only a 20-year dataset - half the size of the initial training data. Furthermore, VA-MoE (IL) exhibits marginal performance improvements over the baseline in long-term predictions of Z500, confirming the absence of catastrophic forgetting during incremental training. These results highlight the efficacy of incremental training in VA-MoE(IL) for maintaining and enhancing predictive accuracy on original variables.",
+        "bbox": [
+            89,
+            489,
+            482,
+            642
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.2. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            89,
+            650,
+            243,
+            666
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Impacts of Architecture. As shown in Tab. 4, we compare VA-MoE with other architectures, including ViT and ViT+MoE. VA-MoE significantly outperforms both ViT and ViT+MoE architectures across 6-hour, 72-hour, and 120-hour prediction horizons, despite ViT+MoE containing nearly twice the number of parameters. This performance gap highlights VA-MoE's suitability for weather forecasting tasks as its channel-adaptive expert design ensures parameter efficiency while maintaining accuracy, particularly in incremental learning. Considering the computational complexity, all ablation studies are in  $1.5^{\\circ}$  with  $128 \\times 256$ .",
+        "bbox": [
+            89,
+            672,
+            482,
+            839
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.3. Visualization",
+        "text_level": 1,
+        "bbox": [
+            89,
+            848,
+            227,
+            862
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In the work, we visualize some predicted results and parameters. In Fig. 5, we visualize the 6-hour predicted results of",
+        "bbox": [
+            89,
+            869,
+            483,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Z500, T850, U10, and T2M. There are max  $0.08\\%$ ,  $0.22\\%$ ,  $1.59\\%$ , and  $0.47\\%$  absolute errors across 4 variables, including 2 upper-air variables and 2 surface variables.",
+        "bbox": [
+            511,
+            489,
+            905,
+            535
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            511,
+            549,
+            633,
+            564
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this work, we proposed incremental weather forecasting, a novel task that addresses the challenge of dynamically expanding weather models to incorporate new variables without retraining from scratch. Our solution, Variables-Adaptive Mixture-of-Experts, enables parameter-efficient adaptation by selectively fine-tuning experts assigned to new variables through index embedding guidance, while preserving pretrained knowledge in existing parameters. Experimental results demonstrate that our approach achieves forecast accuracy comparable to fully retrained models while drastically reducing computational costs.",
+        "bbox": [
+            511,
+            574,
+            906,
+            741
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            511,
+            753,
+            700,
+            771
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This research was supported by the Hong Kong RGC General Research Fund (152169/22E, 152228/23E, 162161/24E), Research Impact Fund (No. R5060-19, No. R5011-23), Collaborative Research Fund (No. C1042-23GF), NSFC/RGC Collaborative Research Scheme (No. 62461160332 & CRS_HKUST602/24), Areas of Excellence Scheme (AoE/E-601/22-R), the InnoHK (HKGAI), and Key R&D Program of Zhejiang Province 2025C01075.",
+        "bbox": [
+            511,
+            779,
+            906,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "7922",
+        "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] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3366-3375, 2017. 3",
+            "[2] Peter Bauer, Alan Thorpe, and Gilbert Brunet. The quiet revolution of numerical weather prediction. Nature, 525(7567): 47-55, 2015. 1, 2",
+            "[3] Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3d neural networks. Nature, 619(7970): 533-538, 2023. 1, 2, 6, 7",
+            "[4] Cristian Bodnar, Wessel P Bruinsma, Ana Lucic, Megan Stanley, Johannes Brandstetter, Patrick Garvan, Maik Riechert, Jonathan Weyn, Haiyu Dong, Anna Vaughan, et al. Aurora: A foundation model of the atmosphere. arXiv:2405.13063, 2024. 1",
+            "[5] Boris Bonev, Thorsten Kurth, Christian Hundt, Jaideep Pathak, Maximilian Baust, Karthik Kashinath, and Anima Anandkumar. Spherical fourier neural operators: learning stable dynamics on the sphere. In Proceedings of International Conference on Machine Learning, 2023. 2",
+            "[6] Hao Chen, Yonghan Dong, Zhe-Ming Lu, Yunlong Yu, and Jungong Han. Self-prompting perceptual edge learning for dense prediction. IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4528-4541, 2023. 2",
+            "[7] Hao Chen, Yonghan Dong, Zheming Lu, Yunlong Yu, and Jungong Han. Pixel matching network for cross-domain few-shot segmentation. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pages 978-987, 2024. 6",
+            "[8] Hao Chen, Yunlong Yu, Yonghan Dong, Zheming Lu, Yingming Li, and Zhongfei Zhang. Multi-content interaction network for few-shot segmentation. ACM Transactions on Multimedia Computing, Communications and Applications, 20 (6):1-20, 2024. 6",
+            "[9] Kang Chen, Tao Han, Junchao Gong, Lei Bai, Fenghua Ling, Jing-Jia Luo, Xi Chen, Leiming Ma, Tianning Zhang, Rui Su, et al. Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead. arXiv:2304.02948, 2023. 1, 2, 7",
+            "[10] Lei Chen, Xiaohui Zhong, Feng Zhang, Yuan Cheng, Yinghui Xu, Yuan Qi, and Hao Li. Fuxi: A cascade machine learning forecasting system for 15-day global weather forecast. npj Climate and Atmospheric Science, 6(1):190, 2023. 2, 7",
+            "[11] Wuyang Chen, Yanqi Zhou, Nan Du, Yanping Huang, James Laudon, Zhifeng Chen, and Claire Cui. Lifelong language pretraining with distribution-specialized experts. In International Conference on Machine Learning, pages 5383-5395. PMLR, 2023. 3",
+            "[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. In Advances in Neural Information Processing Systems, pages 16344-16359, 2022. 6",
+            "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Trans-"
+        ],
+        "bbox": [
+            93,
+            114,
+            483,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "formers for image recognition at scale. arXiv:2010.11929, 2020.2,7,8",
+            "[14] Lihao Gan, Xin Man, Chenghong Zhang, and Jie Shao. Ew-moe: An effective model for global weather forecasting with mixture-of-experts. arXiv:2405.06004, 2024. 6, 7",
+            "[15] Yuan Gao, Hao Wu, Ruiqi Shu, Huanshuo Dong, Fan Xu, Rui Chen, Yibo Yan, Qingsong Wen, Xuming Hu, Kun Wang, et al. Oneforecast: A universal framework for global and regional weather forecasting. arXiv:2502.00338, 2025. 2",
+            "[16] Dipam Goswami, Albin Soutif-Cormerais, Yuyang Liu, Sandesh Kamath, Bart Twardowski, Joost van de Weijer, et al. Resurrecting old classes with new data for exemplar-free continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 28525-28534, 2024. 3",
+            "[17] 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",
+            "[18] Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, András Horányi, Joaquin Muñoz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al. The era5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730):1999-2049, 2020. 6",
+            "[19] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3",
+            "[20] Ryan Keisler. Forecasting global weather with graph neural networks. arXiv:2202.07575, 2022. 6",
+            "[21] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3346-3356, 2024. 3",
+            "[22] Thorsten Kurth, Shashank Subramanian, Peter Harrington, Jaideep Pathak, Morteza Mardani, David Hall, Andrea Miele, Karthik Kashinath, and Anima Anandkumar. Fourcastnet: Accelerating global high-resolution weather forecasting using adaptive fourier neural operators. In Proceedings of the Platform for Advanced Scientific Computing Conference, 2023. 1, 2, 6, 7",
+            "[23] 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, 6, 7",
+            "[24] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv:2010.08895, 2020. 2",
+            "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision, pages 10012-10022, 2021. 2"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "7923",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[26] Franco Molteni, Roberto Buizza, Tim N Palmer, and Thomas Petroliagis. The ecmwf ensemble prediction system: Methodology and validation. Quarterly journal of the royal meteorological society, 122(529):73-119, 1996. 2",
+            "[27] Tung Nguyen, Johannes Brandstetter, Ashish Kapoor, Jayesh K. Gupta, and Aditya Grover. Climax: A foundation model for weather and climate. In International Conference on Machine Learning, 2023. 6, 7",
+            "[28] Tung Nguyen, Rohan Shah, Hritik Bansal, Troy Arcomano, Romit Maulik, Rao Kotamarthi, Ian Foster, Sandeep Madireddy, and Aditya Grover. Scaling transformer neural networks for skillful and reliable medium-range weather forecasting. In Advances in Neural Information Processing Systems, pages 68740-68771, 2025. 6, 7",
+            "[29] Youqi Pan, Wugen Zhou, Yingdian Cao, and Hongbin Zha. Adaptive vio: Deep visual-inertial odometry with online continual learning. arXiv:2405.16754, 2024. 3",
+            "[30] Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. In European Conference on Computer Vision, 2024. 8",
+            "[31] Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, et al. Probabilistic weather forecasting with machine learning. Nature, 637(8044):84-90, 2025. 2",
+            "[32] D Richardson. The thorpex interactive grand global ensemble (tigge). In Geophysical Research Abstracts, page 02815, 2024. 7",
+            "[33] Harold Ritchie, Clive Temperton, Adrian Simmons, Mariano Hortal, Terry Davies, David Dent, and Mats Hamrud. Implementation of the semi-lagrangian method in a high-resolution version of the ecmwf forecast model. Monthly Weather Review, 123(2):489-514, 1995. 2",
+            "[34] Minhyuk Seo, Hyunseo Koh, Wonje Jeung, Minjae Lee, San Kim, Hankook Lee, Sungjun Cho, Sungik Choi, Hyunwoo Kim, and Jonghyun Choi. Learning equi-angular representations for online continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23933-23942, 2024. 3",
+            "[35] Yuwen Tan, Qinhao Zhou, Xiang Xiang, Ke Wang, Yuchuan Wu, and Yongbin Li. Semantically-shifted incremental adapter-tuning is a continual vitransformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23252-23262, 2024. 3",
+            "[36] Yogesh Verma, Markus Heinonen, and Vikas Garg. ClimODE: Climate forecasting with physics-informed neural ODEs. In International Conference on Learning Representations, 2024. 6",
+            "[37] Jialin Wu, Xia Hu, Yaqing Wang, Bo Pang, and Radu Soricut. Omni-smola: Boosting generalist multimodal models with soft mixture of low-rank experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14205-14215, 2024. 3",
+            "[38] Wei Xiong, Muyuan Ma, Xiaomeng Huang, Ziyang Zhang, Pei Sun, and Yang Tian. Koopmanlab: machine learning for"
+        ],
+        "bbox": [
+            91,
+            90,
+            480,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "solving complex physics equations. APL Machine Learning, 1(3), 2023. 2",
+            "[39] Wanghan Xu, Fenghua Ling, Wenlong Zhang, Tao Han, Hao Chen, Wanli Ouyang, and Lei Bai. Generalizing weather forecast to fine-grained temporal scales via physics-ai hybrid modeling. In Advances in Neural Information Processing Systems, 2024. 6",
+            "[40] Hongwei Yan, Liyuan Wang, Kaisheng Ma, and Yi Zhong. Orchestrate latent expertise: Advancing online continual learning with multi-level supervision and reverse self-distillation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23670-23680, 2024. 3",
+            "[41] Yuqi Yang, Peng-Tao Jiang, Qibin Hou, Hao Zhang, Jinwei Chen, and Bo Li. Multi-task dense prediction via mixture of low-rank experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27927-27937, 2024. 3",
+            "[42] Fei Ye and Adrian G Bors. Online task-free continual generative and discriminative learning via dynamic cluster memory. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 26202-26212, 2024. 3",
+            "[43] Donggeun Yoon, Minseok Seo, Doyi Kim, Yeji Choi, and Donghyeon Cho. Probabilistic weather forecasting with deterministic guidance-based diffusion model. In European Conference on Computer Vision, pages 108-124, 2024. 2",
+            "[44] Demin Yu, Xutao Li, Yunming Ye, Baoquan Zhang, Chuyao Luo, Kuai Dai, Rui Wang, and Xunlai Chen. Diffcast: A unified framework via residual diffusion for precipitation nowcasting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 27758-27767, 2024. 2",
+            "[45] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23219-23230, 2024. 3",
+            "[46] Yuchen Zhang, Mingsheng Long, Kaiyuan Chen, Lanxiang Xing, Ronghua Jin, Michael I Jordan, and Jianmin Wang. Skilful nowcasting of extreme precipitation with nowcastnet. Nature, 619(7970):526-532, 2023. 2",
+            "[47] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, Cheng-Lin Liu, and Zhaoxiang Zhang. Rcl: Reliable continual learning for unified failure detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12140-12150, 2024. 3"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            752
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "7924",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/122157a9-e33d-4cdd-b0f9-d8b983f2a104_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7eea66fdf5dc129b9d33f8e0ce279a1716764baf57971bd6accca28cff3537d
+size 2144797

2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/full.md

@@ -0,0 +1,350 @@
+# VA-MoE: Variables-Adaptive Mixture of Experts for
+
+# Incremental Weather Forecasting
+
+![](images/408a686bb8235a3b61e408da6aca9f7a13881c231183d5f5fbc8364f783e3092.jpg)
+
+Hao Chen $^{1\dagger}$  Han Tao $^{1\dagger}$  Guo Song $^{1*}$  Jie Zhang $^{1}$  Yonghan Dong $^{2}$  Yunlong Yu $^{3}$  Lei Bai $^{4}$ $^{1}$ Hong Kong University of Science and Technology (HKUST)  $^{2}$ Huawei Technologies Ltd.  
+ $^{3}$ Zhejiang University  $^{4}$ Shanghai AI Laboratory
+
+Equal contribution *Corresponding author {hchener, thanad}@connect.ust.hk
+
+{songguo, csejzhang}@ust.hk dongyonghan@huawei.com yuyunlong@zju.edu.cn bailei@pjlab.org.cn
+
+# Abstract
+
+This paper presents Variables-Adaptive Mixture of Experts (VA-MoE), a novel framework for incremental weather forecasting that dynamically adapts to evolving spatiotemporal patterns in real-time data. Traditional weather prediction models often struggle with exorbitant computational expenditure and the need to continuously update forecasts as new observations arrive. VA-MoE addresses these challenges by leveraging a hybrid architecture of experts, where each expert specializes in capturing distinct sub-patterns of atmospheric variables (e.g., temperature, humidity, wind speed). Moreover, the proposed method employs a variable-adaptive gating mechanism to dynamically select and combine relevant experts based on the input context, enabling efficient knowledge distillation and parameter sharing. This design significantly reduces computational overhead while maintaining high forecast accuracy. Experiments on ERA5 dataset demonstrate that VA-MoE performs comparable against state-of-the-art models in both short-term (e.g., 1-3 days) and long-term (e.g., 5 days) forecasting tasks, with only about  $25\%$  of trainable parameters and  $50\%$  of the initial training data. Code: https://github.com/chenhao-zju/VAMoE
+
+# 1. Introduction
+
+Weather forecasting remains one of the most critical scientific challenges with profound socio-economic impacts. While traditional Numerical Weather Prediction (NWP) [2] relies on solving complex partial different equations to simulate atmospheric dynamics, recent advances in data-driven artificial intelligence [3, 4, 9, 22] have revolutionized earth system modeling. Deep learning models, in particular, have demonstrated remarkable potential in capturing diverse spatiotemporal patterns across weather and climate phenomena.
+
+![](images/a597efb3a93059332d6e7ca7100ecc4ea802f35eb2a605ebaae8713017f4c997.jpg)  
+Figure 1. Illustration of two different training paradigms of weather forecasting when new variables arrive: (c) Full-Train the whole model (d) Incremental-Train only the module corresponding new variables. (b) Performances of these two paradigms.
+
+ena, giving rise to the emerging field of AI for Weather.
+
+However, a fundamental limitation of existing AI4Weather approaches lies in their assumption that all variables are available synchronously during training and inference. In reality, meteorological variables are often heterogeneous in terms of data sources, collection frequencies, and spatial-temporal distributions. For instance, the upper-air variables (e.g., temperature profiles) are sparse and sampled via radiosondes/satellites, while the surface variable (e.g., precipitation, wind) are dense but updated in near-real-time. This asynchrony poses significant challenges: when introducing new variables (e.g., satellite-derived aerosol data), existing models must be entirely retrained from scratch, incurring prohibitive computational costs. For example, Pangu [3] required 64 days and 192 V100 GPUs for a full retraining cycle.
+
+To address this challenge, we propose Incremental Weather Forecasting (IWF), a novel paradigm designed to dynamically expand meteorological forecast variables in response to increasing observational data. By enabling parameter-efficient variable scaling, IWF revolutionizes operational meteorology by offering a streamlined alternative
+
+to conventional models. This innovation tackles a critical limitation in the field: the need for scalable modeling solutions as datasets grow gradually, ensuring adaptability without compromising computational efficiency.
+
+A pivotal limitation of existing incremental approaches is catastrophic forgetting, i.e., when introducing new variables, pre-trained parameters drift towards new distributions, causing significant performance degradation on original variables. To overcome this, we present the Variables-Adaptive Mixture of Experts (VA-MoE) model, the first expert-based system explicitly designed for hierarchical atmospheric variables.
+
+Mixture of Experts (MoE) models inherently enable scalable representation learning for multi-variable atmospheric datasets by dynamically expanding expert modules as input dimensions grow. While traditional MoE architectures excel in representational capacity, their parallelized design which distributes uniform weights across experts, often leads to homogeneous outputs, limiting specialized knowledge acquisition. To address this limitation, we introduce VA-MoE, an auxiliary-loss-free framework that decouples expert specialization from computational efficiency. VA-MoE employs a phased training strategy: Initial experts are pretrained on base variables, then incrementally expanded to accommodate new variables. Crucially, the pretrained experts are frozen during expansion phases to prevent catastrophic forgetting, ensuring robust retention of prior expertise while adapting to novel inputs. To foster expert diversity without auxiliary losses, VA-MoE incorporates variable index embeddings, i.e., positional metadata encodings that guide experts to develop domain-specific specialization. These embeddings dynamically activate contextually relevant computational pathways for each variable, optimizing resource allocation at inference time. By integrating variable index embeddings that encode spatial-temporal metadata, VA-MoE directs experts to specialize in specific atmospheric domains. This context-aware routing reduces inference latency compared to static MoE models while maintaining expert diversity.
+
+Besides, we propose a gradient-scaled variable-adaptive loss function that dynamically aligns variable optimization rates with their inherent spatiotemporal characteristics. By quantifying the distinct temporal evolution patterns of atmospheric variables, our method allocates differential loss weights, i.e., larger gradient magnitudes prioritize fast-transient fields like temperature to capture immediate atmospheric dynamics, while gradual weight adjustments stabilize slow-changing fields such as geopotential height to preserve long-term system behavior.
+
+As shown in Fig. 1, our paradigm employs a two-phase training paradigm: an initial phase optimizing base atmospheric variables, followed by an incremental phase integrating new variables while freezing previously trained ex
+
+perts. Experiments with the same 40-year dataset reveal that our paradigm Fig. 1(a)+(d) achieves superior z500 predictions performance compared to full retraining Fig. 1(a)+(c).
+
+In conclusion, the contributions of this work include:
+
+- This work initiates systematic research on incremental learning paradigms for weather forecasting. We propose the quantitative benchmark to evaluate the trade-offs between model scalability and generalization in incremental weather modeling.  
+- We present Variables-Adaptive Mixture of Experts (VA-MoE), the first framework tailored for incremental atmospheric modeling. VA-MoE achieves expert specialization through contextual variable activation driven by variable index embeddings, enabling dynamic assignment of experts to variables during both training and inference.  
+- Extensive experiments on the ERA5 dataset demonstrate that VA-MoE achieves comparable performance for surface variables, while delivering superior accuracy in upper-air variables against the existing competitors under  $50\%$  reduced dataset size and  $25\%$  fewer parameters.
+
+# 2. Related work
+
+# 2.1. Data-Driven Weather Forecasting
+
+In recent decades, traditional NWP [2] methods have dominated the weather forecasting field due to the robust predictions and rigorous mathematical validation. However, NWP methods [26, 33] require training from scratch for new prediction, resulting in slow computation and high costs.
+
+Recent advances in deep learning have catalyzed a paradigm shift toward data-driven models for medium-range weather forecasting. Pioneering works like FourCastNet [22] leverage Fourier Neural Operators (FNOs) [24] to learn spatiotemporal weather patterns, while PanguWeather [3] employs a 3D vision transformer [13, 25] to model atmospheric dynamics. Subsequent innovations, including Neural Operators [5, 24, 38] and specialized architectures [9, 10, 15, 23, 31, 43, 44, 46], have achieved accuracy rivaling NWP at a fraction of computation.
+
+While existing data-driven models excel at fixed-variable forecasting after costly initial training, their inflexibility poses a critical barrier: retraining from scratch is required to incorporate new variables, rendering cross-domain adaptation impractical. In contrast, our work introduces incremental weather forecasting, a novel paradigm evolving spatiotemporal patterns in real-time data. By decoupling variable-specific expertise from shared dynamics, our approach supports dynamic expansion to new atmospheric parameters with minimal retraining overhead.
+
+# 2.2. Mixture of Experts for Incremental Learning
+
+Advances in deep learning[6, 17] have spurred interest in parameter-efficient transfer strategies, where models adapt
+
+![](images/0f6c306f6dc259a2ccf6fca6af91be6c6aae1b11d49c857416c88197d46d07c4.jpg)
+
+![](images/124c88fa5df353a7a3ccd95e9e7a9a285c004e52ab5edbcdfdd0876865d9dbfd.jpg)
+
+![](images/d77f605a7ffeb8dd32d6de34d598d180acb7152790dd370ed404951bb38af79d.jpg)  
+Figure 2. Illustration of the incremental weather forecasting paradigm. (a), (b) are the initial and incremental stages, respectively. (c) presents the detailed structure in both initial and incremental stages. During the initial stage, the model is trained with upper-air variables(UV). In the incremental stage, the fire components are trained while the snow components remain frozen. The index embedding is trained by the Index Encoder and divided into six parts based on the upper-air variables (UV: Z, Q, U, V, T) and surface variables (SV).
+
+to new domains by fine-tuning only a subset of parameters. A key paradigm, incremental learning, addresses the stability-plasticity dilemma: retaining knowledge of prior tasks while integrating new concepts. However, real-world data often involves domain shifts, such as variations in variable distributions or task objectives, leading to two dominant research threads. Class Incremental Learning (CIL) [16, 21, 35, 47], focuses on expanding classification tasks with new classes while Task Incremental Learning (TIL) [29, 34, 40, 42] adapts models to distinct task sequences.
+
+One of the most popular structures for addressing incremental learning tasks is the Mixture of Experts (MoE) model [19], which learns representations of new concepts through different experts. Due to its sparse architecture, many works have adopted the MoE structure to reduce inference costs and enhance model capacity [37, 41]. An early work introducing MoE to incremental learning is Expert Gate [1], which trains multiple backbones as different experts and assigns new domains to the relevant expert. Lifelong-MoE [11] leverages pretrained experts and gates to retain the representational knowledge of the training domain. In addition to these approaches, MoE is used as an adapter in MoE-Adapter [45], where adapters are attached to the main structure. These works have demonstrated the promising performance of MoE in visual and natural language incremental learning.
+
+In contrast to MoE designed for fixed-variable systems,
+
+our propose incremental learning of variable distributions by assigning distinct pressure-level groups to specialized experts. Unlike prior incremental learning literature that primarily focus on task or dataset-specific adaptation, our incremental paradigm targets variable-centric expansion, enabling models to adapt to evolving observational data without retraining from scratch.
+
+# 3. Methodology
+
+Our work focuses on a novel paradigm of incremental weather forecasting with VA-MoE. In this section, we first formally define the problem in Sec. 3.1. Then, we present the structure of VA-MoE, consisting with the Channel-Adaptive Expert (CAE) and the Shared Expert in Sec. 3.2. The model during incremental stage is introduced in Sec. 3.3. Finally, loss function is presented in Sec. 3.4.
+
+# 3.1. Overview
+
+For the weather forecasting task, the AI-based model  $\Phi$  aims to predict the future weather variables  $\mathbf{X}^{t + 1}$  with the historical weather variables  $\mathbf{X}^t$ , i.e.,  $\mathbf{X}^{t + 1} = \Phi (\mathbf{X}^t)$ .
+
+In the incremental weather forecasting task, the weather variables  $\mathbf{X}^t$  are divided into two sets: (i) the initial training variables at  $t$  time  $\mathbf{X}_h^t \in \mathbb{R}^{H \times W \times N}$ , and, (ii) the incremental variables at  $t$  time  $\mathbf{X}_{sv}^t \in \mathbb{R}^{H \times W \times M}$ . Unlike traditional weather forecasting tasks, where models are trained
+
+![](images/ec231ab293262e1e38372cb270935bd9ebecfb3ad7ef96b9c46580860695194a.jpg)  
+Figure 3. Illustration of the VA-MoE. In the left subgraph, the input features comprise both upper-air and surface variables. During the initial training stage, five distinct CAE modules process the upper-air variables. In the incremental stage, surface variables are handled by a dedicated module,  $\mathrm{CAESV}$ , while the original five CAEs remain frozen to preserve learned representations. The right subgraph details the CAE module. The Index Embedding is selected through a multi-layer process and applied to weight the Input Embedding via multiplication. Initially discrete, the index embedding is transformed into a continuous representation after passing through the GateEmbed layer, enabling it to dynamically reflect the relevance of specific channels to the input variables.
+
+with all variables available simultaneously, the incremental weather forecasting task sequentially incorporates dynamic variables as they become available over time. The incremental weather forecasting process is structured into a sequence of learning stages: an initial training stage followed by subsequent training stages.
+
+In the initial training stage, the available variables of 40 years are used to train a weather model, i.e.,  $\mathrm{X}_h^{t + 1} = \Phi_h(\mathrm{X}_h^t)$ , as shown in Fig. 2(a). In the incremental training stage, the half variables used at the initial stage and the surface variables at the incremental stage are incorporated together, to train the newly added modules for the surface variables while keeping the model parameters from the initial stage frozen, i.e.,  $\mathrm{X}^{t + 1} = \Phi (\mathrm{X}^t)$ , where  $\Phi$  consists of both the fixed parameters of the pre-trained model  $\Phi_h$  and the trainable parameters of  $\Phi_{sv}$ ,  $\mathrm{X}^t$  consists of all the available variables at time  $t$ ,  $\mathrm{X}^{t + 1}$  consists of the predicted variables at time  $t + 1$ , as shown in Fig. 2(b). In this work, the upper-air variables are utilized during the initial training stage while the surface variables become available and are incorporated during the incremental training stage. The upper-air variables include five different types, Z, Q, U, V, and T, each defined across 13 different levels.
+
+# 3.2. VA-MoE
+
+We introduce index embedding within transformer blocks, a novel mechanism to dynamically guide experts in learning hierarchical relationships between meteorological vari
+
+ables. As new variables are incrementally integrated, corresponding experts are added to the transformer architecture and optimized via index-based affinity assignments. Building on this framework, we adapt the Mixture-of-Experts (MoE) paradigm which enables task-specific specialization and flexible integration of new experts—into weather forecasting transformers.
+
+The structure of the proposed VA-MoE is illustrated in Fig. 2(c). The input features, extracted by the encoder, first pass through a normalization layer and a self-attention layer. The output of the self-attention layer is then combined with the residual connection. Then, another normalization layer is applied, followed by a VA-MoE module, which is also equipped with a residual connection. This process can be formulated as:
+
+$$
+\begin{array}{l} \mathbf {X} _ {m i d} = \mathbf {X} _ {i n} + \operatorname {S A} (\text {L a y e r N o r m} (\mathbf {X} _ {\text {i n}})), \\ \mathbf {X} _ {i n} = \mathbf {X} _ {i n - 1} + \operatorname {V A M E} (\mathbf {I} _ {i n - 1}, \mathbf {N} _ {i n - 1} (\mathbf {X} _ {i n})) \end{array} \tag {1}
+$$
+
+$$
+\mathbf {X} _ {\text {o u t}} = \mathbf {X} _ {\text {m i d}} + \operatorname {V A - M o E} (\operatorname {L a y e r N o r m} (\mathbf {X} _ {\text {m i d}})),
+$$
+
+where  $\mathbf{X}_{in}$  and  $\mathbf{X}_{out}$  are the input and the output of each Transformer Block, respectively.
+
+# 3.2.1. Variables and Index Embedding
+
+As illustrated in Fig. 3, during the training stage, the input feature  $\mathbf{X}_h^t\in \mathbb{R}^{H\times W\times N}$  consists of five different types of variables:  $\mathbf{X}_Z^t,\mathbf{X}_Q^t,\mathbf{X}_U^t,\mathbf{X}_V^t,\mathbf{X}_T^t\in \mathbb{R}^{H\times W\times (N / 5)}$ , where,  $H\times W$  denotes the spatial resolution. These five types of variables are concatenated and processed by a shared expert and five distinct Channel-Adaptive Experts (CAEs).
+
+In addition to these variables, the framework introduces an extra one-hot index embedding  $\mathbf{I}_h\in \mathbb{R}^{5\times N}$  to guide experts in learning variable affinity. The number  $N$  corresponds to the number of variables, and 5 corresponds to the number of variable types. Similar to the input weather features, the index embedding is composed of five separate embeddings:  $\mathbf{I}_Z,\mathbf{I}_Q,\mathbf{I}_U,\mathbf{I}_V,\mathbf{I}_T\in \mathbb{R}^{1\times 1\times N}$ . This initial index embedding is then encoded into a latent space using a Linear layer, expressed as  $\mathbf{I}_h = \mathrm{Linear}(\mathbf{I}_h)$ .
+
+# 3.2.2. Channel-Adaptive Expert (CAE)
+
+The key design of VA-MoE is the channel-adaptive expert, as shown in Fig. 3. For each variable  $Z$ , the corresponding expert module processes both the feature embeddings of the input variables  $\mathbf{X}_h^t$  and the index embedding  $\mathbf{I}_Z$ , which is formulated as:
+
+$$
+\mathbf {X} _ {Z} ^ {t, C A E} = \operatorname {C A E} _ {Z} \left(\mathbf {X} _ {h} ^ {t}, \mathbf {I} _ {Z}\right), \tag {2}
+$$
+
+where  $\mathrm{CAE}_Z$  is the abbreviation of channel-adaptive expert module for variable  $Z$ .
+
+In the CAE module for variable  $Z$ , the index embedding  $\mathbf{I}_Z$  is encoded through a linear layer for feature projection and then multiplied channel-wise with the input embedding  $\mathbf{X}_h^t$ . The resulting fused feature is processed
+
+by the GateEmbed layer, which produces a gate embedding to enhance the fusion of the index and variable embeddings. This enhanced feature is then normalized using a SoftMax layer and filtered through a TopK operation, which selects the top-K high-rank channels. This process yields the GateIndex  $\mathrm{GI}_Z\in \mathbb{R}^{H\times W\times K}$  and GateWeight  $\mathrm{GW}_Z\in \mathbb{R}^{H\times W\times K}$ . The process is formulated as:
+
+$$
+\mathbf {I} _ {Z} ^ {t o p k}, \mathbf {W} _ {Z} ^ {t o p k} = \mathrm {T O P} _ {k} (\operatorname {S o f t M a x} (\operatorname {M L P} _ {Z} (\mathbf {X} _ {h} ^ {t} \odot \mathbf {I} _ {Z}))), (3)
+$$
+
+where  $\mathbf{I}_Z^{topk}$  and  $\mathbf{W}_Z^{topk}$  denote the index and weight of the Top-k features. Once the GateIndex and GateWeight are computed, the input embedding is selectively processed based on the GateIndex, while the variable embedding is weighted channel-wise according to the GateWeight. The GateEmbed layer uses variable embeddings to guide the index embedding, enabling the creation of a variable-specific confidence matrix. Within the CAE module, this embedding enhances the confidence matrix by incorporating both semantic and positional information, rather than relying solely on positional information.
+
+The selected features for variable  $Z$  are obtained with:
+
+$$
+\mathbf {X} _ {Z} ^ {t, s e l e c t e d} = \mathbf {W} _ {Z} ^ {t o p k} * \operatorname {S e l e c t} _ {Z} \left(\mathbf {X} _ {h} ^ {t}, \mathbf {I} _ {Z} ^ {t o p k}\right), \tag {4}
+$$
+
+where  $\mathrm{Select}_Z$  denotes selecting Top-K embedding from the input feature  $\mathbf{X}_h^t$  within the  $\mathrm{CAE}_Z$  module.
+
+The selected and weighted embedding captures the key features associated with the specific type of variables. In the  $\mathrm{CAE}_Z$  module, these features serve as the training set to model the distribution of the variable  $Z$  with  $\mathrm{Expert}_Z$ . The expert module adopts a parallel structure, splitting the channels into multiple parts and enhancing each part with a sparse MoE. This process is formulated as:
+
+$$
+\mathbf {X} _ {Z} ^ {t, C A E} = \operatorname {E x p e r t} _ {Z} (\mathbf {X} _ {Z} ^ {t, s e l e c t e d}), \tag {5}
+$$
+
+where  $\mathbf{Expert}_Z$  is implemented with a multiple layer.
+
+# 3.2.3. Shared Expert.
+
+The shared expert processes the overall features of all variables and operates in parallel with the variable-specific modules. Each variable is handled by its corresponding CAE module, and their outputs are summarized to produce the fused features  $\mathbf{X}_h^{t,fused}$  for all variables. Since the channel count of the selected variable embeddings is lower than that of the original embeddings, an additional up-channel linear layer,  $\mathrm{Linear}_{\mathrm{up}}$ , is introduced. The fused features and the output from the shared expert are then combined via pixel-wise addition. This process is expressed as:
+
+$$
+\mathbf {X} _ {h} ^ {t, f u s e d} = \operatorname {C A E} _ {Z} \left(\mathbf {X} _ {h} ^ {t}, \mathbf {I} _ {Z}\right) + \dots + \operatorname {C A E} _ {T} \left(\mathbf {X} _ {h} ^ {t}, \mathbf {I} _ {T}\right), \tag {6}
+$$
+
+$$
+\left(\mathbf {X} ^ {\prime}\right) _ {h} ^ {t} = \operatorname {E x p e r t} _ {s h a r e d} \left(\mathbf {X} _ {h} ^ {t}\right) + \operatorname {L i n e a r} _ {u p} \left(\mathbf {X} _ {h} ^ {t, f u s e d}\right). \tag {7}
+$$
+
+# 3.3. Incremental Learning Stage
+
+When new variables are incrementally introduced to the model, specialized experts are dynamically integrated into the transformer blocks to process these variables. Concurrently, both the encoder and decoder architectures undergo synchronized updates to maintain compatibility with the expanded variable set.
+
+# 3.3.1. Encoder and Decoder Modules.
+
+When  $K$  new variables are introduced incrementally, the encoder's input dimension expands to  $H \times W \times (N + M)$ , necessitating an adjustment to the convolutional layer's kernel from  $(3, 3, N, C)$  to  $(3, 3, N + M, C)$ . To accommodate this, the layer's parameters are partitioned into two components: The  $(3, 3, N, C)$  segment retains pretrained weights from earlier stages, ensuring continuity. The  $(3, 3, M, C)$  segment, corresponding to the new variables, is initialized randomly to learn emergent patterns.
+
+The same strategy is applied to the position embedding and the Decoder module.
+
+# 3.3.2. VA-MoE Module.
+
+During the incremental learning phase,  $M$  new experts are dynamically integrated into the MoE layer of each transformer block to accommodate  $M$  newly introduced variables. In this phase, only the new experts and the shared expert are actively trained, while all preexisting experts remain frozen to preserve previously learned knowledge.
+
+# 3.3.3. Index Embedding Module.
+
+When  $M$  new variables become available during the incremental stage, the index embedding changes from  $\mathbf{I}_h\in$ $\mathbb{R}^{(N\times l)\times N}$  to  $\mathbf{I}_h\in \mathbb{R}^{(N\times l + M\times r)\times (N + M)}$ , where each  $M$  new variables has  $r$  levels. Additionally, the index encoder needs to be retrained.
+
+# 3.4. Objective Function
+
+To train the model, we introduce a novel dynamic prediction loss combined with a reconstruction loss.
+
+# 3.4.1. Dynamic Prediction Loss.
+
+The prediction loss aims to minimize the difference between the model's predicted results and the actual future weather variables. Unlike existing weather forecasting methods that treat all the variables equally, we argue that different variables, such as temperature and surface pressure, follow distinct distributions. Thus, we introduce a novel dynamic prediction loss that dynamically assigns a weight to each channel of the input data. This loss is formulated as:
+
+$$
+O b j _ {p r e d} = \left(\hat {\mathbf {X}} ^ {t + 1} - \mathbf {X} ^ {t + 1}\right) \odot \left(\hat {\mathbf {X}} ^ {t + 1} - \mathbf {X} ^ {t + 1}\right) / e ^ {\mathbf {w}} + \mathbf {w}, (8)
+$$
+
+where  $\odot$  denotes Hadamard product,  $e$  denotes the base of the natural logarithm,  $\mathbf{X}^{t + 1}\in \mathbb{R}^{H\times W\times C}$ , and  $\mathbf{w}\in$
+
+<table><tr><td>Name</td><td>Description</td><td>Levels</td></tr><tr><td colspan="3">Upper-Air</td></tr><tr><td>Z</td><td>Geopotential</td><td>13</td></tr><tr><td>Q</td><td>Specific humidity</td><td>13</td></tr><tr><td>U</td><td>x-direction wind</td><td>13</td></tr><tr><td>V</td><td>y-direction wind</td><td>13</td></tr><tr><td>T</td><td>Temperature</td><td>13</td></tr><tr><td colspan="3">Surface-Incremental Learning</td></tr><tr><td>u10</td><td>x-direction wind at 10m height</td><td>Single</td></tr><tr><td>v10</td><td>y-direction wind at 10m height</td><td>Single</td></tr><tr><td>t2m</td><td>Temperature at 2m height</td><td>Single</td></tr><tr><td>msl</td><td>Mean sea-level pressure</td><td>Single</td></tr><tr><td>sp</td><td>Surface pressure</td><td>Single</td></tr></table>
+
+$\mathbb{R}^{1\times 1\times C}$  is a learnable vector. With this loss function, the model could dynamically prioritize variables based on their distributional characteristics.
+
+# 3.4.2. Reconstruction Loss.
+
+We introduce an additional reconstruction loss that connects the encoder and decoder directly, formulated as:
+
+$$
+O b j _ {r e c o n} = \left(\hat {\mathbf {X}} ^ {t} - \mathbf {X} ^ {t}\right) \odot \left(\hat {\mathbf {X}} ^ {t} - \mathbf {X} ^ {t}\right), \tag {9}
+$$
+
+where  $\hat{\mathbf{X}}^t = \mathrm{Dec}(\mathrm{Enc}(\mathbf{X}^t))$  represents the reconstructed variables. With the loss function, the encoder and decoder modules specialize in encoding and decoding features, while the intermediate transformer blocks focus on learning the data distribution. This ensures that the decoding process is entirely handled by decoder, allowing the intermediate blocks to focus on their task and not participate in decoding, thereby optimizing network efficiency.
+
+To this end, the model is trained with the final objective function, a combination of the dynamic prediction loss and reconstruction loss. The final objective is formulated as:
+
+$$
+O b j _ {f i n a l} = O b j _ {p r e d} + \lambda O b j _ {r e c o n}, \tag {10}
+$$
+
+where  $\lambda$  denotes a hyper-parameter.
+
+# 4. Experiments
+
+Dataset. In this work, we conduct experiments on a popular weather dataset, i.e., ERA5[18], provided by the ECMWF. ERA5 dataset is a reanalysis atmospheric dataset, consisting of the atmospheric variables from 1979 to the present day with a  $0.25^{\circ}$  spatial resolution with  $721 \times 1440$ . We train the model on 40-year dataset in the initial stage and continuously train the model on 20-year dataset in the incremental stage, which contains the weather variables from
+
+![](images/5c32d76ef3271d634424358f221973a4c9a674afb759538b61aacedf3e343fc8.jpg)
+
+![](images/a0d1aa500d574a7e95efd8c0c9b28ef1e85a3a448b254b80980622a66a0bc2a2.jpg)
+
+![](images/052ad5c9ca6a4a26d4335087addf28eac50c9b1211888b4204be97c3469dbdba.jpg)  
+Figure 4. Comparative analysis of RMSE  $\downarrow$  across 10 data-driven models for four variables, including Z500 and T850 (upper-air variables) in the initial stage, as well as T2M and U10 (incremental surface variables) in the incremental stage.
+
+![](images/bd3379cdf299482f4eddfea62f542bb052f9854ee8afec711d077e1cd24bf117.jpg)
+
+Table 1. A summary of atmospheric variables. The 13 levels are 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000 hPa. 'Single' denotes the variables under earth's surface.  
+
+<table><tr><td>——ClimaX</td><td>——EWMoE</td><td>——WeatherGFT</td><td>——Stormer</td></tr><tr><td>——FourCastNet</td><td>——ClimODE</td><td>——GraphCast</td><td>——Ours(IL)</td></tr><tr><td>——Panguo-Weather</td><td>——Keisler</td><td></td><td></td></tr></table>
+
+1979 to 2020 and 2000 to 2020 year, respectively. And we test the model on one-year dataset of weather variables in 2021. In this work, the model processes 5 upper-air variables and 5 surface variables as Tab. 1, where the upper-air variables are exploited in the training stage, and the surface variables are exploited in the incremental stage.
+
+Implementation Details. The main structure of this work follows the novel backbones [7, 8, 12]. We apply the AdamW optimizer with 0.0002 and 0.00005 learning rates to initial and incremental stages, respectively. In two stages, we train 100 epochs and set batch size to 16. Our model is trained with PyTorch using 16 A100 GPUs.
+
+# 4.1. Main Results
+
+Fig. 4 presents a quantitative comparison of our structure with existing state-of-the-art approaches, including ClimaX [27], Pangu-Weather [3], ClimODE [36], WeatherGFT [39], FourCastNet [22], EWMoE [14], Keisler [20], GraphCast [23], and Stormer [28], for the 5-day weather prediction task. All the competitors are trained with both upper-air and surface variables on a  $0.25^{\circ}$  weather dataset with a resolution of  $721 \times 1440$ .
+
+For the surface variables, including T2M and U10, our VA-MoE demonstrates performance similar to the leading methods Stormer [28] and Graphcast [23]. Compared to other competitors, our structure shows significant advantages in both short-term and long-term predictions. For the upper-air variable Z500, our structure delivers one of the best performances among all 10 methods, outperforming
+
+<table><tr><td rowspan="2"></td><td rowspan="2">Dataset(years)</td><td rowspan="2">Iteration\( (×10^4) \)</td><td colspan="3">T2M(K)↓</td><td colspan="3">U10(m/s)↓</td><td colspan="3">V10(m/s)↓</td><td colspan="3">MSL(Pa)↓</td><td colspan="3">SP(Pa)↓</td></tr><tr><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td></tr><tr><td colspan="18">Plain Training</td></tr><tr><td>ViT* [13]</td><td>1979-2020</td><td>40</td><td>0.72</td><td>1.35</td><td>1.86</td><td>0.66</td><td>1.98</td><td>3.01</td><td>0.68</td><td>2.02</td><td>3.11</td><td>40.2</td><td>208.5</td><td>393.9</td><td>63.3</td><td>222.1</td><td>397.0</td></tr><tr><td>IFS [32]</td><td>1979-2020</td><td>40</td><td>1.09</td><td>1.38</td><td>1.74</td><td>0.96</td><td>1.87</td><td>2.78</td><td>0.99</td><td>1.93</td><td>2.87</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Pangu-Weather [3]</td><td>1979-2020</td><td>40</td><td>0.82</td><td>1.09</td><td>1.53</td><td>0.77</td><td>1.63</td><td>2.54</td><td>0.79</td><td>1.68</td><td>2.65</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>FourCastNet [22]</td><td>1979-2020</td><td>40</td><td>0.82</td><td>1.02</td><td>1.77</td><td>0.82</td><td>2.08</td><td>3.34</td><td>0.84</td><td>2.11</td><td>3.41</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>ClimaX [27]</td><td>1979-2020</td><td>40</td><td>1.11</td><td>1.47</td><td>1.83</td><td>1.04</td><td>2.02</td><td>2.79</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Graphcast [23]</td><td>1979-2020</td><td>40</td><td>0.51</td><td>0.94</td><td>1.37</td><td>0.38</td><td>1.51</td><td>2.37</td><td>-</td><td>-</td><td>-</td><td>23.4</td><td>135.2</td><td>278.2</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Fengwu [9]</td><td>1979-2020</td><td>40</td><td>0.58</td><td>1.03</td><td>1.41</td><td>0.42</td><td>1.53</td><td>2.32</td><td>-</td><td>-</td><td>-</td><td>23.2</td><td>137.1</td><td>276.9</td><td>-</td><td>-</td><td>-</td></tr><tr><td>FuXi [10]</td><td>1979-2020</td><td>40</td><td>0.55</td><td>0.99</td><td>1.41</td><td>0.42</td><td>1.50</td><td>2.36</td><td>0.43</td><td>1.54</td><td>2.44</td><td>27.2</td><td>136.7</td><td>282.9</td><td>-</td><td>-</td><td>-</td></tr><tr><td>VA-MoE</td><td>1979-2020</td><td>40</td><td>0.57</td><td>1.03</td><td>1.42</td><td>0.43</td><td>1.41</td><td>2.25</td><td>0.44</td><td>1.46</td><td>2.34</td><td>27.5</td><td>131.1</td><td>275.9</td><td>57.1</td><td>168.9</td><td>302.4</td></tr><tr><td colspan="18">Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (79-20)</td></tr><tr><td>VA-MoE (IL)</td><td>1979-2020</td><td>20</td><td>0.58</td><td>1.05</td><td>1.45</td><td>0.48</td><td>1.47</td><td>2.33</td><td>0.47</td><td>1.54</td><td>2.41</td><td>27.9</td><td>137.3</td><td>281.6</td><td>59.3</td><td>173.4</td><td>312.4</td></tr><tr><td colspan="18">Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (00-20)</td></tr><tr><td>VA-MoE (IL)</td><td>2000-2020</td><td>10</td><td>0.73</td><td>1.17</td><td>1.57</td><td>0.54</td><td>1.58</td><td>2.49</td><td>0.55</td><td>1.63</td><td>2.57</td><td>30.0</td><td>148.8</td><td>304.7</td><td>60.6</td><td>171.4</td><td>314.8</td></tr></table>
+
+Table 2. Prediction performances on Incremental Training with 5 surface variables, i.e., T2M, U10, V10, MSL, and SP. * denotes running by ourselves. The best results are marked in bold. All experiments are in  ${0.25}^{ \circ  }$  with  ${721} \times  {1440}$  resolutions.  
+
+<table><tr><td rowspan="2"></td><td rowspan="2">Para. (M)</td><td colspan="3">Z500(m2/s2) ↓</td><td colspan="3">Q500(×e-3, g/kg) ↓</td><td colspan="3">U500(m/s) ↓</td><td colspan="3">V500(m/s) ↓</td><td colspan="3">T500(K) ↓</td></tr><tr><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td></tr><tr><td>IFS [32]</td><td>-</td><td>28.31</td><td>154.08</td><td>333.96</td><td>0.31</td><td>0.61</td><td>0.75</td><td>1.43</td><td>3.23</td><td>5.12</td><td>1.40</td><td>3.58</td><td>5.64</td><td>0.36</td><td>0.98</td><td>1.70</td></tr><tr><td>Pangu-Weather [3]</td><td>-</td><td>24.88</td><td>167.90</td><td>391.26</td><td>0.25</td><td>0.55</td><td>0.69</td><td>0.96</td><td>3.13</td><td>4.73</td><td>0.91</td><td>3.52</td><td>5.15</td><td>0.27</td><td>0.94</td><td>1.56</td></tr><tr><td>Graphcast [23]</td><td>-</td><td>15.23</td><td>125.42</td><td>275.35</td><td>-</td><td>-</td><td>-</td><td>0.77</td><td>2.86</td><td>4.49</td><td>0.74</td><td>2.92</td><td>4.67</td><td>0.23</td><td>0.87</td><td>1.48</td></tr><tr><td>VA-MoE</td><td>665</td><td>19.28</td><td>134.63</td><td>295.52</td><td>0.17</td><td>0.49</td><td>0.62</td><td>0.84</td><td>2.99</td><td>4.71</td><td>0.84</td><td>3.04</td><td>4.89</td><td>0.25</td><td>0.76</td><td>1.36</td></tr><tr><td colspan="17">Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (00-20)</td></tr><tr><td>VA-MoE (IL)</td><td>137</td><td>18.23</td><td>133.14</td><td>292.63</td><td>0.17</td><td>0.49</td><td>0.61</td><td>0.84</td><td>3.01</td><td>4.74</td><td>0.84</td><td>3.51</td><td>4.93</td><td>0.25</td><td>0.76</td><td>1.37</td></tr></table>
+
+Table 3. Prediction performances on Initial Training with 5 upper-air variables. All experiments are in  ${0.25}^{ \circ  }$  with  ${721} \times  {1440}$  resolutions.
+
+Pangu-Weather [3] and Stormer [28], and significantly surpassing ClimaX [27] and FourCastNet [22], while slightly inferior to GraphCast [23]. For T850, our structure is slightly inferior to EWMoE [14], GraphCast [23], and Stormer [28] in the short-term predictions within 48 hours. After that period, our structure achieves the best performance in the long-term predictions.
+
+Results of Surface Variables Prediction. This study evaluates model performance on five surface variables: T2M, U10, V10, MSL, and SP. While competing structures are trained with both upper-air and surface variables, VA-MoE (IL) employs a two-stage training approach: initial training with upper-air variables followed by incremental training with all five surface variables.
+
+Analysis of the results from Tab. 2 demonstrates that VA-MoE outperforms competing models across multiple variables and settings, particularly in U10 and V10. Although VA-MoE shows marginally weaker performance in T2M compared to GraphCast [23] and comparable results to FengWu [9] and FuXi [10], it achieves superior long-term predictions for U10, V10, MSL, and SP. These indicate
+
+that while VA-MoE exhibits slight limitations in short-term forecasting, it excels in long-term prediction tasks.
+
+When implementing the incremental training strategy, VA-MoE (IL) trained on a 40-year dataset with only half the training iterations achieves performance comparable to standard VA-MoE. Even with a reduced 20-year dataset, VA-MoE(IL) maintains acceptable performance using just a quarter of the iterations required by the standard approach. These results underscore the efficacy of incremental learning for atmospheric datasets and suggest its potential to replace conventional training methods in operational settings.
+
+Results of Upper-Air Variables Prediction. This study evaluates model performance on upper-air variables, comparing three training approaches: (1) VA-MoE, trained exclusively on upper-air variables; (2) VA-MoE (IL), which incorporates incremental training on surface variables after initial upper-air training; and (3) competing models trained simultaneously on both upper-air and surface variables.
+
+As shown in Tab. 3, VA-MoE achieves performance comparable to GraphCast [23] and outperforms IFS [32] and Pangu-Weather [3] across both short- and long-term
+
+<table><tr><td rowspan="2"></td><td rowspan="2">Para. (M)</td><td colspan="3">Z500 (m2/s2) ↓</td><td colspan="3">Q500 (×e-3, g/kg) ↓</td><td colspan="3">U500 (m/s) ↓</td><td colspan="3">V500 (m/s) ↓</td><td colspan="3">T500 (K) ↓</td></tr><tr><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td><td>6h</td><td>72h</td><td>120h</td></tr><tr><td>ViT*[13]</td><td>307</td><td>33.38</td><td>209.4</td><td>517.81</td><td>0.22</td><td>0.61</td><td>1.06</td><td>1.24</td><td>3.66</td><td>6.52</td><td>1.22</td><td>3.76</td><td>7.41</td><td>0.42</td><td>1.18</td><td>2.40</td></tr><tr><td>ViT+MoE (light)*[30]</td><td>609</td><td>37.92</td><td>207.11</td><td>405.73</td><td>0.22</td><td>0.60</td><td>0.78</td><td>1.30</td><td>3.84</td><td>5.87</td><td>1.27</td><td>3.89</td><td>6.11</td><td>0.46</td><td>1.23</td><td>2.02</td></tr><tr><td>ViT+MoE* [30]</td><td>1113</td><td>28.31</td><td>169.61</td><td>356.02</td><td>0.23</td><td>0.56</td><td>0.72</td><td>1.21</td><td>3.46</td><td>5.44</td><td>1.23</td><td>3.54</td><td>5.69</td><td>0.35</td><td>1.07</td><td>1.83</td></tr><tr><td>VA-MoE</td><td>665</td><td>20.59</td><td>139.02</td><td>302.13</td><td>0.18</td><td>0.49</td><td>0.62</td><td>0.91</td><td>3.02</td><td>4.76</td><td>0.91</td><td>3.08</td><td>4.97</td><td>0.27</td><td>0.92</td><td>1.59</td></tr><tr><td>VA-MoE (IL)</td><td>137</td><td>20.29</td><td>138.52</td><td>301.41</td><td>0.18</td><td>0.50</td><td>0.63</td><td>0.91</td><td>3.04</td><td>4.79</td><td>0.91</td><td>3.10</td><td>5.03</td><td>0.27</td><td>0.93</td><td>1.60</td></tr></table>
+
+Table 4. Architectural impact on 5 upper-air variables under  ${500}\mathrm{{hPa}}$  . All experiments are in  ${1.5}^{ \circ  }$  with  ${128} \times  {256}$  resolutions.
+
+![](images/7a3f0e25dc8da0ded62ed82c5f1ef56cb67792b1eb0e64e3699a3a53e9edaebc.jpg)  
+Figure 5. 6-hour global weather prediction of upper-air and surface variables forecast visualization generated by VA-MoE framework.
+
+prediction. Notably, VA-MoE (IL) demonstrates robust predictive capabilities for upper-air variables even after incremental training on surface variables, despite using only a 20-year dataset - half the size of the initial training data. Furthermore, VA-MoE (IL) exhibits marginal performance improvements over the baseline in long-term predictions of Z500, confirming the absence of catastrophic forgetting during incremental training. These results highlight the efficacy of incremental training in VA-MoE(IL) for maintaining and enhancing predictive accuracy on original variables.
+
+# 4.2. Ablation Study
+
+Impacts of Architecture. As shown in Tab. 4, we compare VA-MoE with other architectures, including ViT and ViT+MoE. VA-MoE significantly outperforms both ViT and ViT+MoE architectures across 6-hour, 72-hour, and 120-hour prediction horizons, despite ViT+MoE containing nearly twice the number of parameters. This performance gap highlights VA-MoE's suitability for weather forecasting tasks as its channel-adaptive expert design ensures parameter efficiency while maintaining accuracy, particularly in incremental learning. Considering the computational complexity, all ablation studies are in  $1.5^{\circ}$  with  $128 \times 256$ .
+
+# 4.3. Visualization
+
+In the work, we visualize some predicted results and parameters. In Fig. 5, we visualize the 6-hour predicted results of
+
+Z500, T850, U10, and T2M. There are max  $0.08\%$ ,  $0.22\%$ ,  $1.59\%$ , and  $0.47\%$  absolute errors across 4 variables, including 2 upper-air variables and 2 surface variables.
+
+# 5. Conclusion
+
+In this work, we proposed incremental weather forecasting, a novel task that addresses the challenge of dynamically expanding weather models to incorporate new variables without retraining from scratch. Our solution, Variables-Adaptive Mixture-of-Experts, enables parameter-efficient adaptation by selectively fine-tuning experts assigned to new variables through index embedding guidance, while preserving pretrained knowledge in existing parameters. Experimental results demonstrate that our approach achieves forecast accuracy comparable to fully retrained models while drastically reducing computational costs.
+
+# 6. Acknowledgements
+
+This research was supported by the Hong Kong RGC General Research Fund (152169/22E, 152228/23E, 162161/24E), Research Impact Fund (No. R5060-19, No. R5011-23), Collaborative Research Fund (No. C1042-23GF), NSFC/RGC Collaborative Research Scheme (No. 62461160332 & CRS_HKUST602/24), Areas of Excellence Scheme (AoE/E-601/22-R), the InnoHK (HKGAI), and Key R&D Program of Zhejiang Province 2025C01075.
+
+# References
+
+[1] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3366-3375, 2017. 3  
+[2] Peter Bauer, Alan Thorpe, and Gilbert Brunet. The quiet revolution of numerical weather prediction. Nature, 525(7567): 47-55, 2015. 1, 2  
+[3] Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3d neural networks. Nature, 619(7970): 533-538, 2023. 1, 2, 6, 7  
+[4] Cristian Bodnar, Wessel P Bruinsma, Ana Lucic, Megan Stanley, Johannes Brandstetter, Patrick Garvan, Maik Riechert, Jonathan Weyn, Haiyu Dong, Anna Vaughan, et al. Aurora: A foundation model of the atmosphere. arXiv:2405.13063, 2024. 1  
+[5] Boris Bonev, Thorsten Kurth, Christian Hundt, Jaideep Pathak, Maximilian Baust, Karthik Kashinath, and Anima Anandkumar. Spherical fourier neural operators: learning stable dynamics on the sphere. In Proceedings of International Conference on Machine Learning, 2023. 2  
+[6] Hao Chen, Yonghan Dong, Zhe-Ming Lu, Yunlong Yu, and Jungong Han. Self-prompting perceptual edge learning for dense prediction. IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4528-4541, 2023. 2  
+[7] Hao Chen, Yonghan Dong, Zheming Lu, Yunlong Yu, and Jungong Han. Pixel matching network for cross-domain few-shot segmentation. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pages 978-987, 2024. 6  
+[8] Hao Chen, Yunlong Yu, Yonghan Dong, Zheming Lu, Yingming Li, and Zhongfei Zhang. Multi-content interaction network for few-shot segmentation. ACM Transactions on Multimedia Computing, Communications and Applications, 20 (6):1-20, 2024. 6  
+[9] Kang Chen, Tao Han, Junchao Gong, Lei Bai, Fenghua Ling, Jing-Jia Luo, Xi Chen, Leiming Ma, Tianning Zhang, Rui Su, et al. Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead. arXiv:2304.02948, 2023. 1, 2, 7  
+[10] Lei Chen, Xiaohui Zhong, Feng Zhang, Yuan Cheng, Yinghui Xu, Yuan Qi, and Hao Li. Fuxi: A cascade machine learning forecasting system for 15-day global weather forecast. npj Climate and Atmospheric Science, 6(1):190, 2023. 2, 7  
+[11] Wuyang Chen, Yanqi Zhou, Nan Du, Yanping Huang, James Laudon, Zhifeng Chen, and Claire Cui. Lifelong language pretraining with distribution-specialized experts. In International Conference on Machine Learning, pages 5383-5395. PMLR, 2023. 3  
+[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. In Advances in Neural Information Processing Systems, pages 16344-16359, 2022. 6  
+[13] Alexey Dosovitskiy. An image is worth 16x16 words: Trans-
+
+formers for image recognition at scale. arXiv:2010.11929, 2020.2,7,8  
+[14] Lihao Gan, Xin Man, Chenghong Zhang, and Jie Shao. Ew-moe: An effective model for global weather forecasting with mixture-of-experts. arXiv:2405.06004, 2024. 6, 7  
+[15] Yuan Gao, Hao Wu, Ruiqi Shu, Huanshuo Dong, Fan Xu, Rui Chen, Yibo Yan, Qingsong Wen, Xuming Hu, Kun Wang, et al. Oneforecast: A universal framework for global and regional weather forecasting. arXiv:2502.00338, 2025. 2  
+[16] Dipam Goswami, Albin Soutif-Cormerais, Yuyang Liu, Sandesh Kamath, Bart Twardowski, Joost van de Weijer, et al. Resurrecting old classes with new data for exemplar-free continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 28525-28534, 2024. 3  
+[17] 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  
+[18] Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, András Horányi, Joaquin Muñoz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al. The era5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730):1999-2049, 2020. 6  
+[19] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3  
+[20] Ryan Keisler. Forecasting global weather with graph neural networks. arXiv:2202.07575, 2022. 6  
+[21] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3346-3356, 2024. 3  
+[22] Thorsten Kurth, Shashank Subramanian, Peter Harrington, Jaideep Pathak, Morteza Mardani, David Hall, Andrea Miele, Karthik Kashinath, and Anima Anandkumar. Fourcastnet: Accelerating global high-resolution weather forecasting using adaptive fourier neural operators. In Proceedings of the Platform for Advanced Scientific Computing Conference, 2023. 1, 2, 6, 7  
+[23] 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, 6, 7  
+[24] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv:2010.08895, 2020. 2  
+[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision, pages 10012-10022, 2021. 2
+
+[26] Franco Molteni, Roberto Buizza, Tim N Palmer, and Thomas Petroliagis. The ecmwf ensemble prediction system: Methodology and validation. Quarterly journal of the royal meteorological society, 122(529):73-119, 1996. 2  
+[27] Tung Nguyen, Johannes Brandstetter, Ashish Kapoor, Jayesh K. Gupta, and Aditya Grover. Climax: A foundation model for weather and climate. In International Conference on Machine Learning, 2023. 6, 7  
+[28] Tung Nguyen, Rohan Shah, Hritik Bansal, Troy Arcomano, Romit Maulik, Rao Kotamarthi, Ian Foster, Sandeep Madireddy, and Aditya Grover. Scaling transformer neural networks for skillful and reliable medium-range weather forecasting. In Advances in Neural Information Processing Systems, pages 68740-68771, 2025. 6, 7  
+[29] Youqi Pan, Wugen Zhou, Yingdian Cao, and Hongbin Zha. Adaptive vio: Deep visual-inertial odometry with online continual learning. arXiv:2405.16754, 2024. 3  
+[30] Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. In European Conference on Computer Vision, 2024. 8  
+[31] Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, et al. Probabilistic weather forecasting with machine learning. Nature, 637(8044):84-90, 2025. 2  
+[32] D Richardson. The thorpex interactive grand global ensemble (tigge). In Geophysical Research Abstracts, page 02815, 2024. 7  
+[33] Harold Ritchie, Clive Temperton, Adrian Simmons, Mariano Hortal, Terry Davies, David Dent, and Mats Hamrud. Implementation of the semi-lagrangian method in a high-resolution version of the ecmwf forecast model. Monthly Weather Review, 123(2):489-514, 1995. 2  
+[34] Minhyuk Seo, Hyunseo Koh, Wonje Jeung, Minjae Lee, San Kim, Hankook Lee, Sungjun Cho, Sungik Choi, Hyunwoo Kim, and Jonghyun Choi. Learning equi-angular representations for online continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23933-23942, 2024. 3  
+[35] Yuwen Tan, Qinhao Zhou, Xiang Xiang, Ke Wang, Yuchuan Wu, and Yongbin Li. Semantically-shifted incremental adapter-tuning is a continual vitransformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23252-23262, 2024. 3  
+[36] Yogesh Verma, Markus Heinonen, and Vikas Garg. ClimODE: Climate forecasting with physics-informed neural ODEs. In International Conference on Learning Representations, 2024. 6  
+[37] Jialin Wu, Xia Hu, Yaqing Wang, Bo Pang, and Radu Soricut. Omni-smola: Boosting generalist multimodal models with soft mixture of low-rank experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14205-14215, 2024. 3  
+[38] Wei Xiong, Muyuan Ma, Xiaomeng Huang, Ziyang Zhang, Pei Sun, and Yang Tian. Koopmanlab: machine learning for
+
+solving complex physics equations. APL Machine Learning, 1(3), 2023. 2  
+[39] Wanghan Xu, Fenghua Ling, Wenlong Zhang, Tao Han, Hao Chen, Wanli Ouyang, and Lei Bai. Generalizing weather forecast to fine-grained temporal scales via physics-ai hybrid modeling. In Advances in Neural Information Processing Systems, 2024. 6  
+[40] Hongwei Yan, Liyuan Wang, Kaisheng Ma, and Yi Zhong. Orchestrate latent expertise: Advancing online continual learning with multi-level supervision and reverse self-distillation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23670-23680, 2024. 3  
+[41] Yuqi Yang, Peng-Tao Jiang, Qibin Hou, Hao Zhang, Jinwei Chen, and Bo Li. Multi-task dense prediction via mixture of low-rank experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27927-27937, 2024. 3  
+[42] Fei Ye and Adrian G Bors. Online task-free continual generative and discriminative learning via dynamic cluster memory. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 26202-26212, 2024. 3  
+[43] Donggeun Yoon, Minseok Seo, Doyi Kim, Yeji Choi, and Donghyeon Cho. Probabilistic weather forecasting with deterministic guidance-based diffusion model. In European Conference on Computer Vision, pages 108-124, 2024. 2  
+[44] Demin Yu, Xutao Li, Yunming Ye, Baoquan Zhang, Chuyao Luo, Kuai Dai, Rui Wang, and Xunlai Chen. Diffcast: A unified framework via residual diffusion for precipitation nowcasting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 27758-27767, 2024. 2  
+[45] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23219-23230, 2024. 3  
+[46] Yuchen Zhang, Mingsheng Long, Kaiyuan Chen, Lanxiang Xing, Ronghua Jin, Michael I Jordan, and Jianmin Wang. Skilful nowcasting of extreme precipitation with nowcastnet. Nature, 619(7970):526-532, 2023. 2  
+[47] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, Cheng-Lin Liu, and Zhaoxiang Zhang. Rcl: Reliable continual learning for unified failure detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12140-12150, 2024. 3

2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bce7286370f0dfd322f20959f8c31c8d2d26e21b49e7e7bbc481daaefe53f0aa
+size 818512

2025/VACE_ All-in-One Video Creation and Editing/bb0c9cf5-4f56-4fe0-bdd0-bf9454fc4819_content_list.json

@@ -0,0 +1,1633 @@
+[
+    {
+        "type": "text",
+        "text": "VACE: All-in-One Video Creation and Editing",
+        "text_level": 1,
+        "bbox": [
+            263,
+            130,
+            736,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zeyinzi Jiang*",
+        "bbox": [
+            153,
+            180,
+            272,
+            199
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zhen Han*",
+        "bbox": [
+            285,
+            181,
+            372,
+            196
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Chaojie Mao*†",
+        "bbox": [
+            395,
+            181,
+            514,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jingfeng Zhang",
+        "bbox": [
+            535,
+            181,
+            661,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yulin Pan",
+        "bbox": [
+            683,
+            181,
+            763,
+            196
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yu Liu",
+        "bbox": [
+            784,
+            181,
+            841,
+            196
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Tongyi Lab, Alibaba Group",
+        "text_level": 1,
+        "bbox": [
+            387,
+            208,
+            609,
+            227
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/6dc0ac2e79e0ba9bb358ce5358c79ab788f0b963e087806e235c7b43592a9f61.jpg",
+        "image_caption": [
+            "Reference-to-video generation"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            250,
+            486,
+            295
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/01187571914a5e6d6532ea027270a9ab6354e126bbeba8577fb665a9462afeee.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            542,
+            252,
+            602,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/85545d4d442ad80243ffb8b78dc40cee7e78444acc8b7f750789189b0c03bbf4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            617,
+            250,
+            901,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/6c9c116f5690eba3d9ee44b3b791916ff76bb2ec58621dbbe86a0ae44980aa0d.jpg",
+        "image_caption": [
+            "Video-to-video editing",
+            "Pose Transfer (Pose Control)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            310,
+            486,
+            349
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/1d10ef66bf9e895e32d99c420bde8e9d16ef77d13a2666119f69ac5d62c76b41.jpg",
+        "image_caption": [
+            "Motion Transfer (Layout Control)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            310,
+            901,
+            349
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/58ad5b44490705232a973cf7a586334b054d0bc8c494901a11fe885919a638d8.jpg",
+        "image_caption": [
+            "Structure Transfer (Depth Control)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            369,
+            486,
+            410
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/80df6d6119b73f2d4005bce8b9c89505b8abb30c091bc04ee8329de809340e56.jpg",
+        "image_caption": [
+            "Colorization (Gray Control)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            511,
+            369,
+            901,
+            410
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/95d1d327420388d5737f00901b38c9c4081cca6ef3806a9ed2ec1ed8f43224ce.jpg",
+        "image_caption": [
+            "Masked Video-to-video editing",
+            "Subject Recreation (Inpainting)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            431,
+            486,
+            473
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/0f9436c54eb21d2c2ced5d62e408773bb87a5fad5ad4a71c1a7bcecc0f3dff35.jpg",
+        "image_caption": [
+            "Subject Removal (Inpainting)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            431,
+            901,
+            473
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/6e856a972006ce8b10c473c7deb34dd2b9911915774764a54da3b6adf70ffa9b.jpg",
+        "image_caption": [
+            "Canvas Extension (Outpainting)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            491,
+            486,
+            532
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/f1285142f3ed6719d624eecacd52d7f643cbad1f1346811431897c717ae26fcc.jpg",
+        "image_caption": [
+            "Temporal Extension (First Clip)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            491,
+            901,
+            532
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/5a8e8ec3fd8352db7fcaa2531b94053d11e401af6e453edc2f6ee0834d1ac198.jpg",
+        "image_caption": [
+            "Task Composition",
+            "Swap Anything (Subject Reference + Inpainting)",
+            "Figure 1. Comprehensive capabilities of VACE. We present the outstanding generation results based on Wan2.1 (left) and LTX-Video (right). For each task, the original input image or video (left), the context video (top left corner), and the generated frames are illustrated."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            551,
+            486,
+            597
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/ca5b623143e9929a16ef37e21b1cf7268db1c9a956e3c85c812cdbf5bb037f27.jpg",
+        "image_caption": [
+            "Animate Anything (Frame Reference + Pose Control)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            511,
+            555,
+            901,
+            597
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            248,
+            666,
+            326,
+            681
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Diffusion Transformer has demonstrated powerful capability and scalability in generating high-quality images and videos. Further pursuing the unification of generation and editing tasks has yielded significant progress in the domain of image content creation. However, due to the intrinsic demands for consistency across both temporal and spatial dynamics, achieving a unified approach for video synthesis remains challenging. We introduce VACE, which enables users to perform Video tasks within an All-in-one framework for Creation and Editing. These tasks include reference-to-video generation, video-to-video editing, and masked video-to-video editing. Specifically, we effectively",
+        "bbox": [
+            88,
+            696,
+            485,
+            878
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "integrate the requirements of various tasks by organizing video task inputs, such as editing, reference, and masking, into a unified interface referred to as the Video Condition Unit (VCU). Furthermore, by utilizing a Context Adapter structure, we inject different task concepts into the model using formalized representations of temporal and spatial dimensions, allowing it to handle arbitrary video synthesis tasks flexibly. Extensive experiments demonstrate that the unified model of VACE achieves performance on par with task-specific models across various subtasks. Simultaneously, it enables diverse applications through versatile task combinations. Project page: https://alivilab.github.io/VACE-Page/.",
+        "bbox": [
+            511,
+            667,
+            908,
+            864
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV 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": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Equal Contribution. †Project lead.",
+        "bbox": [
+            107,
+            886,
+            297,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "17191",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            91,
+            90,
+            222,
+            104
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In recent years, the domain of visual generation tasks has witnessed remarkable advancements, driven in particular by the rapid evolution of diffusion models [24, 25, 48, 53, 54, 56, 57]. Beyond the early foundational pre-trained models for text-to-image [7, 16, 33] or text-to-video [9, 22, 64] generation in the field, there has been a proliferation of downstream tasks and applications, such as painting [3, 82], editing [4, 42, 68, 70, 75], controllable generation [30, 76], frame reference generation [20, 73], and ID-referenced video synthesis [11, 35, 47, 74]. This array of developments highlights the dynamic and complex nature of the visual generation field. To enhance task flexibility and reduce the overhead associated with deploying multiple models, researchers have begun to focus on constructing unified model architectures [12, 63] (e.g., ACE [23, 41] and Omni-Gen [71]), aiming to integrate different tasks into a single image model, facilitating the creation of various application workflows while maintaining simplicity in usage. In the field of video, due to the collaborative transformations in both temporal and spatial dimensions, leveraging a unified model can present endless possibilities for video creation. However, leveraging diverse input modalities and ensuring spatiotemporal consistency are still challenging for unified video generation and editing.",
+        "bbox": [
+            89,
+            114,
+            485,
+            477
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We propose VACE, an all-in-one model for video creation and editing that performs tasks including reference-to-video generation, video-to-video editing, masked video-to-video editing, and free composition of these tasks, as illustrated in Fig. 1. The aggregation of various capabilities reduces the costs of service deployment and user interaction, and by combining the capabilities of different tasks within a single model, it addresses challenges faced by existing video generation models, such as controllable generation of long videos, multi-condition and reference-based generation, and continuous video editing, thereby empowering users with greater creativity. To achieve this, we utilize the current mainstream Diffusion Transformers (DiTs) structure as the foundational video framework and pre-trained text-to-video generation models [22, 64], which provide better basic capabilities and scalability for handling long video sequences. Specifically, VACE takes into account the needs of different tasks during its construction and designs a unified interface, dubbed the Video Condition Unit (VCU), which integrates multiple modalities such as images or videos for editing, references, and masks. Furthermore, to differentiate the visual modality information in editing and reference tasks, we introduce the concept decoupling strategy, enabling the model to understand what aspects need to be retained and what should be modified. Meanwhile, by employing a pluggable Context Adapter structure, concepts from different tasks (e.g., the areas or ranges of editing or reference) are injected into the model through col",
+        "bbox": [
+            89,
+            478,
+            483,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "laborative spatiotemporal representation, enabling it to possess the capability of adaptive processing for unified tasks.",
+        "bbox": [
+            511,
+            90,
+            903,
+            119
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Due to the lack of existing multi-task benchmarks in video synthesis, we construct a dataset of 480 evaluation samples containing 12 different tasks, while evaluating the performance of the VACE unified model by comparing it with existing specialized models. Experimental results demonstrate that our framework exhibits sufficient competitiveness in both quantitative and qualitative analyses. To the best of our knowledge, we are the first all-in-one model based on the video DiT architecture that concurrently supports such a wide range of tasks. Notably, this innovative framework allows for the compositional expansion of base tasks, enabling the construction of scenarios such as long video re-rendering, which provides a versatile and efficient solution for video synthesis, opening new possibilities for user-side video content creation and editing.",
+        "bbox": [
+            511,
+            122,
+            906,
+            348
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            511,
+            361,
+            655,
+            376
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Visual Generation and Editing. With the rapid development of image [2, 7, 16, 18, 58, 59] and video [22, 32, 73, 77] generation models, they are being used to create high-quality visual content and are widely applied in fields such as advertising, film special effects, game development, and animation production [13, 43-45, 55]. Meanwhile, to meet the diverse needs of visual media production and to enhance efficiency and quality, precise generation and editing methods have emerged. Models are required to perform generative creation based on multimodal inputs, such as depth, structure, pose, scene, and characters. According to the purposes of the input conditions, we can categorize them into two types: editing of the input and concept-guided re-creation. A significant portion of the work, such as ControlNet [76], ControlVideo,Composer [26], VideoComposer [68], and SCEdit [30], focuses on single-condition editing and multi-condition composite editing based on temporal and spatial alignment conditions. Additionally, some works that focus on interactive local editing scenarios, such as DragGAN [46] and MagicBrush [75]. Methods that guide generation based on semantic information from the input, such as Cone [38], Cone2 [39], InstantID [67], and PuLID [21], can achieve conceptual understanding of the input and inject it into the model for creative purposes.",
+        "bbox": [
+            511,
+            387,
+            906,
+            763
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Task-unified Visual Generative Model. As the complexity and diversity of user creations increase, relying solely on a single model or a complicated chain of multiple models can no longer provide a convenient and efficient path for implementing creative ideas. In image generation, a unified generation and editing framework has begun to emerge, allowing for more flexible creative approaches. Methods such as UltraEdit [81] and SEED-Data-Edit [19] provide general-purpose editing datasets, while techniques like In",
+        "bbox": [
+            511,
+            765,
+            908,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "17192",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/2bdea4dc0e5c25dd5529b7d9f6e57e9084415154530503123357e880e21f4386.jpg",
+        "image_caption": [
+            "Figure 2. Task categories covered by VACE. Four basic tasks can be combined to create a vast number of possibilities."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            99,
+            89,
+            897,
+            290
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "structPix2Pix [4], MagicBrush [61], and CosXL [60] offer general instruction-based editing features. Additionally, methods like UniControl [50] and UNIC-Adapter [15] have unified controllable generation. Further advancements have led to the development of ACE [23, 41], OmniGen [71], OmniControl [63], and UniReal [12], which expand the scope of tasks by providing flexible controllable generation, local editing, and reference-guided generation. In the video domain, due to the increased difficulty of generation, approaches often manifest as single-task single-model frameworks, offering capabilities for editing or reference generation, as seen in Video-P2P [37], MagicEdit [34], MotionCtrl [69], Magic Mirror [80], and Phantom [35]. VACE aims to fill the gap for a unified model within the video domain, providing possibilities for complex creative scenarios.",
+        "bbox": [
+            89,
+            330,
+            483,
+            559
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            89,
+            577,
+            181,
+            593
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "VACE is designed as a multimodal-to-video generation model, where text, image, video, and mask are integrated into a unified conditioning input. To cover as many video generation and editing tasks as possible, we conduct in-depth research into existing tasks, then divide them into 4 categories according to their individual requirements of multimodal inputs. Without losing generality, we specifically design a novel multimodal input format for each category under a Video Condition Unit (VCU) paradigm. Finally, we restructure the DiT model for VCU inputs, making it a versatile model for a wide range of video tasks.",
+        "bbox": [
+            89,
+            604,
+            483,
+            771
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Multimodal Inputs and Video Tasks.",
+        "text_level": 1,
+        "bbox": [
+            89,
+            786,
+            406,
+            801
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Despite existing video tasks varying in complex user inputs and ambitious creative goals, we found that most of their inputs can be fully represented in 4 modalities: text, image, video, and mask. Overall, as illustrated in Fig. 2, we group these video tasks into 5 categories based on their requirements of these four multimodal inputs.",
+        "bbox": [
+            89,
+            810,
+            483,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Text-to-Video Generation (T2V) is a basic video creation task, and text is the only input.",
+            "- Reference-to-Video Generation (R2V) requires additional images as reference inputs, making sure that specified contents, such as subjects of faces, animals, and other objects, or video frames, appear in the generated video.",
+            "- Video-to-Video Editing (V2V) makes an entire change to a provided video, such as colorization, stylization, controllable generation, etc. We use video control types whose control signals can be represented and stored as RGB videos, including depth, grayscale, pose, scribble, optical flow, and layout; however, the method itself is not limited to these.",
+            "- Masked Video-to-Video Editing (MV2V) makes changes to an input video only within the provided 3D regions of interest (3D ROI), seamlessly blending in with the other unchanged regions, such as inpainting, outpainting, video extension, etc. We use an extra spatiotemporal mask to represent the 3D ROI.",
+            "- Task Composition includes all the compositional possibilities of the 4 types of video tasks above."
+        ],
+        "bbox": [
+            513,
+            330,
+            903,
+            648
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Video Condition Unit",
+        "text_level": 1,
+        "bbox": [
+            513,
+            657,
+            714,
+            672
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We propose an input paradigm, Video Condition Unit (VCU), to unify diverse input conditions into textual input, frame sequence, and mask sequence. A VCU can be denoted as",
+        "bbox": [
+            511,
+            680,
+            906,
+            739
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nV = [ T; F; M ], \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            653,
+            741,
+            903,
+            758
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $T$  is a text prompt, while  $F$  and  $M$  are sequences of context video frames  $\\{u_1, u_2, \\dots, u_n\\}$  and masks  $\\{m_1, m_2, \\dots, m_n\\}$  respectively. Here,  $u$  is in RGB space, normalized to  $[-1, 1]$  and  $m$  is binary, in which \"1\"s and \"0\"s symbolize where to edit or not.  $F$  and  $M$  are aligned both in spatial size  $h \\times w$  and temporal size  $n$ . In T2V, no context frame or mask is required. To keep generality, we assign a default value  $0_{h \\times w}$  to each  $u$  denoting empty input, and set every  $m$  to  $1_{h \\times w}$  meaning that all these pixels",
+        "bbox": [
+            511,
+            763,
+            906,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "17193",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/2a2cdb8973973f507ef1873f7f8e7f67b97950d6faa55f40b8c825386eb1ca8e.jpg",
+        "image_caption": [
+            "Figure 3. Overview of VACE Framework. Frames and masks are tokenized through Concept Decoupling, Context Latent Encode, and Context Embedder. To achieve training with VCU as input, we employ two strategies: (a) Fully Fine-tuning and (b) Context Adapter Tuning. The latter converges faster and supports pluggable features."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            98,
+            88,
+            486,
+            256
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/572bd3aa43bbd8d28bf5ab05a3e7a83f3135cec10cac1bc8df624e2eb426a3ba.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            490,
+            88,
+            901,
+            257
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/92a45626850099a38ad40633be9b967221966e8df2a8a8367cad01631bff0ba4.jpg",
+        "table_caption": [
+            "Table 1. The formal representation of frames  $(F\\mathrm{s})$  and masks  $(M\\mathrm{s})$  under the four basic tasks. Frames and masks are aligned spatially and temporally."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Tasks</td><td>Frames (Fs) &amp; Masks (Ms)</td></tr><tr><td>T2V</td><td>F = {0h×w} × n\nM = {1h×w} × n</td></tr><tr><td>R2V</td><td>F = {r1, r2, ..., rl} + {0h×w} × n\nM = {0h×w} × l + {1h×w} × n</td></tr><tr><td>V2V</td><td>F = {u1, u2, ..., un}\nM = {1h×w} × n</td></tr><tr><td>MV2V</td><td>F = {u1, u2, ..., un}\nM = {m1, m2, ..., mn}</td></tr></table>",
+        "bbox": [
+            91,
+            376,
+            470,
+            549
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "are about to be re-generated. For R2V, additional reference frames  $r_i$  are inserted in front of the default frame sequence, while all-zero masks  $0_{h \\times w}$  are inserted in front of the mask sequence. These all-zero masks mean that the corresponding frames should be kept unchanged. In V2V, the context frame sequence is the input video frames, and the context mask is a sequence of  $1_{h \\times w}$ . For MV2V, both context video and mask are required. The formal mathematical representations are shown in Tab. 1.",
+        "bbox": [
+            89,
+            575,
+            482,
+            710
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "VCU can also support task composition. For example, the context frames of reference-inpainting task are  $\\{r_1,r_2,\\dots,r_l,u_1,u_2,\\dots,u_n\\}$  and the context masks are  $\\{0_{h\\times w}\\} \\times l + \\{m_1,m_2,\\dots,m_n\\}$ . In this case, users can modify  $l$  objects in the video and regenerate based on the provided reference images. For another example, users only have a scribble image and want to generate a video beginning with the contents described by this scribble image, which is a scribble-based video extension task. The context frames are  $\\{u\\} +\\{0_{h\\times w}\\} \\times (n - 1)$  and the context masks are  $\\{1_{h\\times w}\\} \\times n$ . In this way, we can achieve multi-condition and reference control generation for long videos.",
+        "bbox": [
+            89,
+            719,
+            482,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Architecture",
+        "text_level": 1,
+        "bbox": [
+            511,
+            325,
+            648,
+            339
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We restructure the DiT model for VACE, as shown in Fig. 3, aiming to support multimodal VCU inputs. Since there is an existing pipeline for text tokenization, we only consider the tokenization of context frames and masks. After being tokenized, the context tokens are combined with noisy video tokens and used to fine-tune the DiT model. Different from that, we also propose a Context Adapter Tuning strategy, which allows context tokens to pass through Context Blocks and be added back to the original DiT Blocks.",
+        "bbox": [
+            511,
+            347,
+            906,
+            483
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3.1. Context Tokenization",
+        "text_level": 1,
+        "bbox": [
+            511,
+            489,
+            709,
+            502
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Concept Decoupling. Two different visual concepts of natural video and control signals, like depth, pose, are encoded in  $F$  simultaneously. We believe that explicitly separating these data of different modalities and distributions is essential for model convergence. The concept decoupling is based on masks and yields two frame sequences identical in shape:  $F_{c} = F \\times M$  and  $F_{k} = F \\times (1 - M)$ , where  $F_{c}$  is called the reactive frames contain all the pixels to be changed, while all the pixels to be kept are stored in  $F_{k}$ , named inactive frames. Specifically, the reference images and the unchanged part of V2V and MV2V go to  $F_{k}$ , while control signals and those pixels about to change, such as gray pixels, are collected to  $F_{c}$ .",
+        "bbox": [
+            511,
+            508,
+            906,
+            703
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Context Latent Encoding. A typical DiT processes noisy video latents  $X \\in \\mathbb{R}^{n' \\times h' \\times w' \\times d}$ , where  $n'$ ,  $h'$ , and  $w'$  are the temporal and spatial shapes of the latent space. Similar to  $X$ ,  $F_{c}$ ,  $F_{k}$ , and  $M$  need to be encoded into a high-dimensional feature space to ensure the property of significant spatiotemporal correlations. Therefore, we reorganize them together with  $X$  into a hierarchical and spatiotemporal aligned visual features.  $F_{c}$ ,  $F_{k}$  are processed by video VAE and mapped into the same latent space of  $X$ , maintaining their spatiotemporal consistency. To avoid any unexpected mishmash of images and videos, reference images are separately encoded by VAE encoder and concatenated back along the temporal dimension, while the correspond",
+        "bbox": [
+            511,
+            704,
+            908,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "17194",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ing parts need to be removed during decoding.  $M$  is directly reshaped and interpolated. After that,  $F_{c}, F_{k}$ , and  $M$  are all mapping into latent spaces and are spatiotemporal aligned with  $X$  in the shape of  $n' \\times h' \\times w'$ .",
+        "bbox": [
+            89,
+            90,
+            480,
+            151
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Context Embedder. We extend the embedder layer by concatenating  $F_{c}$ ,  $F_{k}$ , and  $M$  in the channel dimension and tokenizing them into context tokens, which we refer to as the Context Embedder. The corresponding weights to tokenize  $F_{c}$  and  $F_{k}$  are directly copied from the original video embedder, and the weights to tokenize  $M$  are initialized to zeros.",
+        "bbox": [
+            89,
+            152,
+            480,
+            256
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3.2. Fully Fine-Tuning and Context Adapter Tuning",
+        "text_level": 1,
+        "bbox": [
+            89,
+            263,
+            467,
+            279
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To achieve training with VCU as input, a simple methodology is to fully fine-tune the whole DiT model, as shown in Fig. 3a. Context tokens are added together with noisy tokens  $X$ , and all the parameters in DiT and the newly introduced Context Embedder will be updated during training. To avoid fully fine-tuning and achieve faster convergence, as well as to establish a pluggable feature with the foundation model, we also propose another methodology processing the context token in a Res-Tuning [29] manner, as shown in Fig. 3b. Particularly, we select and copy several Transformer Blocks from the original DiT, forming a distributed and cascade-type Context Blocks. The original DiT processes video tokens and text tokens, while the newly added Transformer Blocks process context tokens and text tokens. The output of each Context Block is inserted back into the DiT blocks as an additive signal to assist the main branch in performing generation and editing tasks. In this manner, the parameters of DiT are frozen. Only the Context Embedder and Context Blocks are trainable.",
+        "bbox": [
+            89,
+            282,
+            482,
+            568
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Datasets",
+        "text_level": 1,
+        "bbox": [
+            89,
+            582,
+            187,
+            597
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Data Construction",
+        "text_level": 1,
+        "bbox": [
+            89,
+            606,
+            269,
+            621
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To obtain an all-in-one model, the diversity and complexity of the required data construction also increase. Existing common text-to-video and image-to-video tasks only require constructing pairs of text and video. However, for the tasks in VACE, the modalities need to be further expanded to include target videos, source videos, local masks, reference, and more. To efficiently and rapidly acquire data for various tasks, it is imperative to maintain video quality while also conducting instance-level analysis and understanding of the video data.",
+        "bbox": [
+            89,
+            628,
+            482,
+            777
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To achieve this, we first analyze the video data itself by performing shot slicing and preliminarily filtering out data based on resolution, aesthetic score, and motion amplitude. Next, we label the first frame of the video using RAM [78] and combine it with Grounding DINO [36] for detection, utilizing the localization results to perform secondary filtering on videos with target areas that are either too small or too large. Furthermore, we employ the propagate operation",
+        "bbox": [
+            89,
+            780,
+            482,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "of SAM2 [52] for video segmentation to obtain instance-level information across the video. Leveraging the results of video segmentation, we filter instances in the temporal dimension by calculating the effective frame ratio based on the mask area threshold.",
+        "bbox": [
+            511,
+            90,
+            903,
+            165
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the actual training process, the construction for different tasks also needs to be tailored according to the characteristics of each task: 1) For some controllable video generation tasks, we pre-extract depth [51], scribble [6], pose [5, 72], and optical flow [65] from the filtered videos. For gray and layout tasks, we create data on the fly. 2) For repainting tasks, random instances from the videos can be masked for inpainting, while the inverse of the mask enables the construction of outpainting data. Augmentation of the masks [62] allows for unconditional inpainting. 3) In the case of extension tasks, we extract key frames such as the first frame, last frame, frames from both ends, random frames, and segments from both ends to support a wider variety of extension types. 4) For reference tasks, we can extract several face or object instances from the videos and apply offline or online augmentation operations to create paired data. Notably, we randomly combine all the previously mentioned tasks for training to accommodate a broader range of model application scenarios. Additionally, for all operations involving masks, we perform arbitrary augmentation to satisfy various granular local generation requirements.",
+        "bbox": [
+            511,
+            167,
+            906,
+            500
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. VACE-Benchmark",
+        "text_level": 1,
+        "bbox": [
+            511,
+            513,
+            692,
+            527
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Significant progress has been made in the field of video generation. However, a scientific and thorough evaluation of the performance of these models remains an urgent issue that needs to be addressed. VBench [27] and VBench++ [28] have established a precise evaluation framework for text-to-video and image-to-video tasks through an extensive assessment suite and dimensional design. Nevertheless, as the video generation ecosystem continues to evolve, more derivative tasks have begun to emerge, such as video reference generation and video editing, for which a comprehensive benchmark is still lacking. To address this gap, we propose VACE-Benchmark to evaluate various downstream tasks related to video in a systematic manner.",
+        "bbox": [
+            511,
+            537,
+            906,
+            747
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Starting from the data sources, we recognize that real videos and generated videos may exhibit different performance characteristics during evaluation. Thus, we collected a total of 240 high-quality videos categorized by their sources, encompassing various data types, including text-to-video, inpainting, outpainting, extension, grayscale, depth, scribble, pose, optical flow, layout, reference face, and reference object tasks, with an average of 20 samples for each task. The input modalities include videos, masks, and references, and we also provide the original videos",
+        "bbox": [
+            511,
+            750,
+            908,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "17195",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/1fa4a54822a61bf1cc9e1b5b49f7de831ea2fd8e0e4c0076d2e8a8520a89fe66.jpg",
+        "table_caption": [
+            "Table 2. Quantitative evaluations on VACE-Benchmark. We compare the automated score metrics of the unified VACE based on LTX-Video and the proprietary model on the dimensions of video quality and video consistency, as well as results of human user studies."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Type</td><td rowspan=\"2\">Method</td><td colspan=\"10\">Video Quality &amp; Video Consistency</td><td colspan=\"4\">User Study</td></tr><tr><td>Aesthetic Quality</td><td>Background Consistency</td><td>Dynamic Degree</td><td>Imaging Quality</td><td>Motion Smoothness</td><td>Overall Consistency</td><td>Subject Consistency</td><td>Temporal Flickering</td><td>Normalized Average</td><td>Prompt Following</td><td>Temporal Consistency</td><td>Video Quality</td><td>Average</td><td></td></tr><tr><td rowspan=\"4\">I2V</td><td>I2VGenXL [77]</td><td>55.20%</td><td>92.87%</td><td>60.00%</td><td>63.31%</td><td>97.43%</td><td>23.78%</td><td>89.58%</td><td>95.67%</td><td>71.54%</td><td>2.65</td><td>1.60</td><td>2.34</td><td>2.20</td><td></td></tr><tr><td>CogVideoX-I2V [73]</td><td>57.78%</td><td>94.80%</td><td>40.00%</td><td>68.23%</td><td>98.69%</td><td>24.38%</td><td>93.84%</td><td>97.84%</td><td>73.66%</td><td>3.30</td><td>2.28</td><td>3.19</td><td>2.92</td><td></td></tr><tr><td>LTX-Video [22]</td><td>56.12%</td><td>94.57%</td><td>35.00%</td><td>62.72%</td><td>99.27%</td><td>24.92%</td><td>92.83%</td><td>98.41%</td><td>72.89%</td><td>2.95</td><td>2.28</td><td>2.28</td><td>2.50</td><td></td></tr><tr><td>VACE (Ours)</td><td>57.53%</td><td>95.32%</td><td>45.00%</td><td>68.03%</td><td>99.08%</td><td>25.13%</td><td>93.61%</td><td>97.80%</td><td>74.38%</td><td>3.20</td><td>4.00</td><td>2.54</td><td>3.24</td><td></td></tr><tr><td rowspan=\"2\">Inpaint</td><td>ProPainter [82]</td><td>44.70%</td><td>95.64%</td><td>50.00%</td><td>61.57%</td><td>99.01%</td><td>18.48%</td><td>92.99%</td><td>98.47%</td><td>70.15%</td><td>2.35</td><td>4.00</td><td>2.99</td><td>3.11</td><td></td></tr><tr><td>VACE (Ours)</td><td>51.30%</td><td>96.30%</td><td>50.00%</td><td>60.39%</td><td>99.12%</td><td>21.12%</td><td>94.59%</td><td>98.21%</td><td>72.05%</td><td>2.40</td><td>4.00</td><td>2.60</td><td>3.00</td><td></td></tr><tr><td rowspan=\"3\">Outpaint</td><td>Follow-Your-Canvas [8]</td><td>53.30%</td><td>95.99%</td><td>5.00%</td><td>69.53%</td><td>98.08%</td><td>25.90%</td><td>95.38%</td><td>97.20%</td><td>71.54%</td><td>3.05</td><td>2.00</td><td>1.63</td><td>2.23</td><td></td></tr><tr><td>M3DDM [17]</td><td>53.34%</td><td>95.87%</td><td>30.00%</td><td>65.07%</td><td>99.22%</td><td>25.43%</td><td>93.65%</td><td>98.85%</td><td>73.16%</td><td>3.70</td><td>3.88</td><td>2.28</td><td>3.29</td><td></td></tr><tr><td>VACE (Ours)</td><td>57.04%</td><td>96.55%</td><td>30.00%</td><td>69.49%</td><td>99.20%</td><td>25.36%</td><td>94.47%</td><td>98.47%</td><td>74.25%</td><td>3.90</td><td>3.92</td><td>3.58</td><td>3.80</td><td></td></tr><tr><td rowspan=\"4\">Depth</td><td>Control-A-Video [10]</td><td>50.62%</td><td>91.71%</td><td>70.00%</td><td>67.76%</td><td>97.58%</td><td>24.48%</td><td>88.10%</td><td>96.58%</td><td>72.35%</td><td>2.70</td><td>2.28</td><td>1.54</td><td>2.17</td><td></td></tr><tr><td>VideoComposer [68]</td><td>50.03%</td><td>94.18%</td><td>70.00%</td><td>59.44%</td><td>96.23%</td><td>24.95%</td><td>89.79%</td><td>94.38%</td><td>70.74%</td><td>2.60</td><td>2.44</td><td>2.17</td><td>2.40</td><td></td></tr><tr><td>ControlVideo [79]</td><td>63.30%</td><td>95.02%</td><td>10.00%</td><td>65.13%</td><td>96.49%</td><td>24.20%</td><td>92.29%</td><td>95.42%</td><td>70.07%</td><td>2.55</td><td>2.50</td><td>1.82</td><td>2.29</td><td></td></tr><tr><td>VACE (Ours)</td><td>56.72%</td><td>96.12%</td><td>60.00%</td><td>66.41%</td><td>98.84%</td><td>25.27%</td><td>94.09%</td><td>97.27%</td><td>74.99%</td><td>3.10</td><td>3.92</td><td>2.66</td><td>3.23</td><td></td></tr><tr><td rowspan=\"4\">Pose</td><td>Text2Video-Zero [31]</td><td>57.63%</td><td>87.67%</td><td>100.00%</td><td>70.74%</td><td>79.65%</td><td>23.94%</td><td>84.82%</td><td>76.57%</td><td>59.69%</td><td>2.15</td><td>2.00</td><td>1.88</td><td>2.01</td><td></td></tr><tr><td>ControlVideo [79]</td><td>65.37%</td><td>94.56%</td><td>25.00%</td><td>65.28%</td><td>97.32%</td><td>25.19%</td><td>92.76%</td><td>96.82%</td><td>72.45%</td><td>2.15</td><td>1.80</td><td>2.03</td><td>1.99</td><td></td></tr><tr><td>Follow-Your-Pose [40]</td><td>48.79%</td><td>86.80%</td><td>100.00%</td><td>67.41%</td><td>90.12%</td><td>26.10%</td><td>80.18%</td><td>88.02%</td><td>66.43%</td><td>2.00</td><td>2.60</td><td>1.58</td><td>2.06</td><td></td></tr><tr><td>VACE (Ours)</td><td>60.17%</td><td>94.92%</td><td>75.00%</td><td>64.71%</td><td>98.63%</td><td>26.44%</td><td>94.82%</td><td>96.60%</td><td>76.13%</td><td>2.95</td><td>3.96</td><td>2.63</td><td>3.18</td><td></td></tr><tr><td rowspan=\"2\">Flow</td><td>FLATTEN [14]</td><td>56.23%</td><td>95.80%</td><td>70.00%</td><td>61.65%</td><td>97.86%</td><td>26.23%</td><td>93.94%</td><td>96.17%</td><td>74.42%</td><td>3.50</td><td>2.40</td><td>3.19</td><td>3.03</td><td></td></tr><tr><td>VACE (Ours)</td><td>55.76%</td><td>96.07%</td><td>75.00%</td><td>65.37%</td><td>98.98%</td><td>25.89%</td><td>94.63%</td><td>96.93%</td><td>75.90%</td><td>2.90</td><td>3.75</td><td>2.60</td><td>3.08</td><td></td></tr><tr><td rowspan=\"4\">R2V</td><td>Kling1.6 [1]</td><td>62.13%</td><td>96.04%</td><td>85.00%</td><td>69.27%</td><td>99.38%</td><td>27.82%</td><td>93.79%</td><td>97.79%</td><td>78.81%</td><td>4.22</td><td>4.10</td><td>3.80</td><td>4.04</td><td></td></tr><tr><td>Pika2.2 [49]</td><td>62.48%</td><td>96.79%</td><td>65.00%</td><td>69.87%</td><td>99.37%</td><td>26.02%</td><td>95.93%</td><td>98.90%</td><td>77.87%</td><td>4.00</td><td>3.85</td><td>3.87</td><td>3.91</td><td></td></tr><tr><td>Vidu2.0 [66]</td><td>64.30%</td><td>96.85%</td><td>35.00%</td><td>67.03%</td><td>99.66%</td><td>26.53%</td><td>96.73%</td><td>99.41%</td><td>76.47%</td><td>3.90</td><td>3.85</td><td>3.77</td><td>3.84</td><td></td></tr><tr><td>VACE (Ours)</td><td>63.25%</td><td>98.03%</td><td>30.00%</td><td>72.29%</td><td>99.51%</td><td>25.85%</td><td>98.54%</td><td>99.15%</td><td>76.76%</td><td>3.47</td><td>3.42</td><td>3.30</td><td>3.40</td><td></td></tr></table>",
+        "bbox": [
+            91,
+            128,
+            913,
+            537
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "to enable further processing by developers based on the specific characteristics of each task. Regarding the data prompts, we supply the original captions of the videos for quantitative assessment, as well as rewritten prompts tailored to the specific tasks to evaluate the models' creativity.",
+        "bbox": [
+            89,
+            554,
+            482,
+            630
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            642,
+            223,
+            659
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            89,
+            666,
+            282,
+            683
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. VACE is trained based on Diffusion Transformers for text-to-video generation at various scales. It utilizes LTX-Video-2B [22] for faster generation, while Wan-T2V-14B [64] is used specifically for higher-quality outputs, supporting resolutions of up to 720p. The training employs a phased approach. Initially, we focus on foundational tasks such as inpainting and extension, which are considered modal complementary to the pre-trained text-to-video models. This includes the incorporation of masks and the learning of contextual generation in both spatial and temporal dimensions. Next, from a task expansion perspective, we progressively transition from single input reference frames to multiple input reference frames and from single tasks to composite tasks. Fi",
+        "bbox": [
+            89,
+            688,
+            483,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "nally, we fine-tune the model's quality using higher-quality data and longer sequences. The input for model training accommodates arbitrary resolutions, dynamic durations, and variable frame rates to support diverse input needs of users.",
+        "bbox": [
+            511,
+            554,
+            906,
+            614
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Baselines. Our goal is to achieve the unification of video creation and editing tasks, and currently, there is no comparable all-in-one video generation model available, which leads us to focus our evaluation on comparing our general model with proprietary task-specific models. Moreover, due to the numerous tasks involved and the lack of open-sourced methods for many of them, we conduct our comparisons on models that are available either offline or online. Specifically for the tasks, we compare the following: 1) For the I2V task, we examine I2VGenXL [77], CogVideoX-I2V [73], and LTX-Video-I2V [22]; 2) In the repainting task, we compare the ProPainter [82] for removal inpainting, while Follow-Your-Canvas [8] and M3DDM [17] are compared for outpainting; 3) For controllable task, we use Control-A-Video [10], VideoComposer [68], and ControlVideo [79] under depth conditions, and compare Text2Video-Zero [31], ControlVideo [79], and Follow-Your-Pose [40] under pose conditions, as well as FLAT",
+        "bbox": [
+            511,
+            628,
+            908,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "17196",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/c7fa9ea987d1bb71f5051d3e521300faee221cee1c690174509157a93ceb2484.jpg",
+        "image_caption": [
+            "Reference Anything"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            98,
+            93,
+            243,
+            148
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f6416d4a007a14dd68aefc6bfd0f1fe6cb773c3b035d97636f24ebb9a0c54d90.jpg",
+        "image_caption": [],
+        "image_footnote": [
+            "The elegant lady carefully selects bags in the boutique, and she shows the charm of a mature woman in a black slim dress with a pearl necklace. Holding a vintage-inspired brown leather handbag, she is carefully observing its craftsmanship and texture. The interior of the store."
+        ],
+        "bbox": [
+            254,
+            88,
+            897,
+            160
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/8434dacb2e4a7709e3f78e2229e213784dba4a91b4b1f46122d498181dd52856.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            183,
+            254,
+            247
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7755152ca82d7f0fbc13b2b7b203bfcb16d5f3adeb3ff41ce7b58ef1ace9a95a.jpg",
+        "image_caption": [],
+        "image_footnote": [
+            "A young boy rises from his chair and walks briskly to the right side of the frame towards the edge of the sun-drenched frame, as if chasing a new adventure. His eyes were bright, and the corners of his mouth were slightly upturned, revealing curiosity and excitement about the unknown..."
+        ],
+        "bbox": [
+            254,
+            180,
+            898,
+            251
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4a2a7dc2056bb9eefe68f4fd0f438de07109f6e820e51ec5c8399a3d85f38d3d.jpg",
+        "image_caption": [
+            "Move Anything"
+        ],
+        "image_footnote": [
+            "\"Anime-style, hot-blooded teenager in bright orange long-sleeved pants sportswear, standing on a surfboard, facing the golden sunshine in the rough sea. The teenager's short yellow hair is flying in the wind, his eyes are firm, and he has a confident smile on the corner of his mouth.\""
+        ],
+        "bbox": [
+            98,
+            271,
+            254,
+            342
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2ddbe8eeb9b05ec19d551d6405830a378b3c62dc317578be5cdea44cf0957355.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            254,
+            271,
+            898,
+            342
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/04fcf82eeef0b0e26c17bc7b5f184eadced2d8d689a0a494b4537ebd63f4d9cb.jpg",
+        "image_caption": [
+            "Animate Anything"
+        ],
+        "image_footnote": [
+            "The video shows an old movie-style scene in retro tones, with a little penguin and a kitten having a joyous bike race. Little penguins and kittens, both dressed in orange-and-red race suits, ride vintage multi-wheeled bikes on a nostalgic dirt road flanked by spectators..."
+        ],
+        "bbox": [
+            96,
+            362,
+            254,
+            433
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/a20893df72f446a06fe28fb383ae5b5f8fa37e8addd8069ae865846de816e160.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            254,
+            362,
+            897,
+            433
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/5105cdd07923abce73ff9a89c0e5639e4918777310ce414986bec976cd654a25.jpg",
+        "image_caption": [
+            "Swap Anything"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            462,
+            254,
+            516
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/9751db892a0a52597f8e5f5d005db3142b8067aa9d728a4fd80dd8f895ed60a5.jpg",
+        "image_caption": [
+            "Expand Anything",
+            "Figure 4. Visualization results of compositional tasks. VACE creatively enables reference-, move-, animate-, swap-, and expand-anything."
+        ],
+        "image_footnote": [
+            "\"In the style of classical oil painting, the background is a river, and in the center of the picture is a mature and elegant woman, wearing a long skirt and sitting on a chair. She took the open red heart-shaped sunglasses from her arms with both hands and put them on...\""
+        ],
+        "bbox": [
+            254,
+            452,
+            897,
+            523
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "TEN [14] under optical flow conditions; 4) In reference generation, given the absence of open-sourced models, we compare commercial products Kling1.6 [1], Pika2.2 [49], and Vidu2.0 [66].",
+        "bbox": [
+            89,
+            597,
+            482,
+            657
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Evaluation. To comprehensively evaluate the performance of various tasks, we employ the VACE-Benchmark for assessment. Specifically, we divide the evaluation into automatic scoring and a user study for manual assessment. For the automatic scoring, we utilize select metrics from VBench [27] to assess video quality and video consistency, including eight indicators: aesthetic quality, background consistency, dynamic degree, imaging quality, motion smoothness, overall consistency, subject consistency, and temporal flickering. For the manual assessment, we utilize the mean opinion score (MOS) as our evaluation metric, focusing on three aspects: prompt following, temporal consistency, and video quality. In practice, we anonymize the generated data and randomly distribute it to different participants for scoring on a scale from 1 to 5.",
+        "bbox": [
+            89,
+            674,
+            482,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2. Main Results",
+        "text_level": 1,
+        "bbox": [
+            513,
+            595,
+            651,
+            611
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Quantitative Evaluation. We compare VACE comprehensive model based on LTX-Video with task proprietary approaches on VACE-Benchmark. For certain tasks, we follow existing methods; for example, although we support generating based on any frame, we conduct comparisons using the first-frame reference approach from current open-source methods to ensure fairness. From Tab. 2, we can see that for the tasks of I2V, inpainting, outpainting, depth, pose, and optical flow, our method demonstrates better performance than other open-source methods across eight indicators of video quality and video consistency, with normalized average metrics showing superior results. Some competing methods can only generate at a resolution of 256, have very short generation durations, and exhibit instability in temporal coherence, resulting in poorer performance on automatic metric calculations. For the R2V task, there is still a certain gap in metrics compared to commercial models for a small-scale model that aims for fast generation,",
+        "bbox": [
+            511,
+            628,
+            906,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "17197",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/cfdbd7dd56c30c3a6af8e7f6a707e50ca6f36261f186b92d725ce4882711d134.jpg",
+        "image_caption": [
+            "(a) Base structure setting."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            98,
+            92,
+            285,
+            210
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/271e1884078ac46369e29ea040268eec90e4a0725fcd2514331506d8f08d1a3c.jpg",
+        "image_caption": [
+            "(b) Hyperparameter settings."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            287,
+            95,
+            473,
+            210
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/95c6ecc97504c4d853d22a32ff59fd3c44365b48d077815be8a79fc91f3a7089.jpg",
+        "image_caption": [
+            "(c) Context adapter configurations."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            98,
+            227,
+            284,
+            342
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/14fdebd9ed061f7d0e1c6ae123885833ea7a70db27684d65351cc2fd5fe7e5d8.jpg",
+        "image_caption": [
+            "(d) Concept decouple setting.",
+            "Figure 5. Ablation Studies of the VACE regarding structures, hyperparameters, and module configurations."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            287,
+            227,
+            473,
+            342
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "while being comparable to the metrics of Vidu 2.0. According to the results of human user studies, our method consistently performs better in evaluation metrics across multiple tasks, aligning well with user preferences.",
+        "bbox": [
+            89,
+            422,
+            482,
+            483
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Qualitative Results. In Fig. 1, we present the results of the VACE single model across various tasks. It is evident that the model achieves a high level of performance in video quality and temporal consistency. Furthermore, in composition tasks shown in Fig. 4, our model showcases impressive abilities, effectively integrating different modalities and tasks to produce results that cannot be generated by existing single or multiple models, thereby demonstrating its strong potential in the fields of video generation and editing. For example, in the \"Move Anything\" case, by providing a single input image and a movement trajectory, we are able to precisely move the characters in the scene with a specified direction while maintaining coherence and narrative consistency.",
+        "bbox": [
+            89,
+            484,
+            482,
+            696
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3. Ablation Studies",
+        "text_level": 1,
+        "bbox": [
+            89,
+            709,
+            254,
+            724
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To better understand the impact of different independent modules on a unified video generation framework, we conducted a series of systematic comparative experiments based on the LTX-Video model to achieve a better model structure and configuration. To accurately assess the different experimental settings, we sample 250 data points for each task as a validation set and calculate the training loss, reflecting the model's training progress through the mean curve changes of different tasks.",
+        "bbox": [
+            89,
+            732,
+            482,
+            868
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Base Structure. Text-guided image or video generation models only take noise as inference input. When extended",
+        "bbox": [
+            89,
+            869,
+            482,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "to our unified input paradigm, VCU, we can conduct training using fully fine-tuning or by incorporating additional parameter fine-tuning. Specifically, as shown in Fig. 5a, we compare the concatenation of different inputs along the channel dimension and modify the input dimensions of the patchify projection layer to achieve the loading and fully fine-tuning of the pre-trained model. Additionally, we introduce some extra training parameters in the form of RestTuning [29], which serializes VCU in a bypass branch and injects information into the main branch. The results indicate that both methods yielded similar effects; however, since the additional parameter fine-tuning converges faster, we base our subsequent experiments on this approach. As shown in Fig. 5b, we further conduct hyperparameter experiments based on this structure, focusing on aspects such as weighting schemes, timestamp shifting, and p-zero.",
+        "bbox": [
+            511,
+            90,
+            903,
+            332
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Context Adapter. Since the number of context blocks will significantly affect the model size and inference time consumption, we attempt to find an optimal number and distribution of context blocks. We begin with selecting continuous blocks at the input side and make comparisons between the first 1/4 blocks, 1/2 blocks, and all blocks. Inspired by the Res-Tuning [29] method, we also experiment with evenly distributing the injection blocks instead of selecting a continuous block series. As shown in Fig. 5c, we can see that when using the same number of blocks, the distributed arrangement of blocks outperforms the continuous arrangement in shallow blocks. Furthermore, a greater number of blocks generally yields better results, but due to the limited improvement in effectiveness and the constraints of training resources, we adopt a partially distributed arrangement of blocks.",
+        "bbox": [
+            511,
+            332,
+            906,
+            571
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Concept Decouple. During training, we introduce a Concept Decouple processing module to further disassemble the visual units, clarifying what content the model needs to learn to modify or retain. As shown in Fig. 5d, using this module results in a more significant reduction in loss.",
+        "bbox": [
+            511,
+            574,
+            905,
+            650
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            511,
+            664,
+            633,
+            679
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This paper introduces VACE, an all-in-one video generation and editing framework. It unifies the diverse and complex multimodal inputs required for various video tasks, bridging the gap between specialized models for each individual task. This enables most video AI creation tasks to be completed with a single inference of a single model. While broadly covering various video tasks, VACE also supports flexible and free combinations of these tasks, greatly expanding the application scenarios of video generation models and meeting a wide range of user creative needs. The VACE framework paves the way for the development of unified visual generative models with multimodal inputs and represents a significant milestone in the field of visual generation.",
+        "bbox": [
+            511,
+            689,
+            905,
+            898
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "17198",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments. We would like to express our sincere appreciation for the contributions of many colleagues for their insightful discussions, valuable suggestions, and constructive feedback, including: Yuwei Wang, Haiming Zhao, Chenwei Xie, and Sheng Yao for their data contributions, and Shiwei Zhang, Tao Fang, and Xiang Wang for their discussions and suggestions.",
+        "bbox": [
+            89,
+            90,
+            485,
+            191
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            91,
+            219,
+            187,
+            234
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] KLING AI. KLING AI, https://klingai.com/, 2025.6,7,13",
+            "[2] Runway AI. Stable Diffusion v1.5 Model Card, https://huggingface.co/runwayml/stable-diffusion-v1-5, 2022.2",
+            "[3] Runway AI. Stable Diffusion Inpainting Model Card, https://huggingface.co/runwayml/stable-diffusion-inpainting, 2022.2",
+            "[4] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning To Follow Image Editing Instructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18392-18402, 2023. 2, 3",
+            "[5] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. OpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields. IEEE Trans. Pattern Anal. Mach. Intell., 43(1):172-186, 2021. 5",
+            "[6] Caroline Chan, Frédo Durand, and Phillip Isola. Learning To Generate Line Drawings That Convey Geometry and Semantics. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7915-7925, 2022. 5",
+            "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt- $\\alpha$ : Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis. arXiv preprint arXiv:2310.00426, 2023. 2",
+            "[8] Qihua Chen, Yue Ma, Hongfa Wang, Junkun Yuan, Wenzhe Zhao, Qi Tian, Hongmei Wang, Shaobo Min, Qifeng Chen, and Wei Liu. Follow-Your-Canvas: Higher-Resolution Video Outpainting with Extensive Content Generation. In Assoc. Adv. Artif. Intell., 2025. 6, 13",
+            "[9] Shoufa Chen, Chongjian Ge, Yuqi Zhang, Yida Zhang, Fengda Zhu, Hao Yang, Hongxiang Hao, Hui Wu, Zhichao Lai, Yifei Hu, Ting-Che Lin, Shilong Zhang, Fu Li, Chuan Li, Xing Wang, Yanghua Peng, Peize Sun, Ping Luo, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Goku: Flow Based Video Generative Foundation Models. arXiv preprint arXiv:2502.04896, 2025. 2",
+            "[10] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Ji-ashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-A-Video: Controllable Text-to-Video Diffusion Models with Motion Prior and Reward Feedback Learning. arXiv preprint arXiv:2305.13840, 2023. 6, 13",
+            "[11] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. AnyDoor: Zero-shot Object-level Image Customization. arXiv preprint arXiv:2307.09481, 2023. 2"
+        ],
+        "bbox": [
+            93,
+            243,
+            483,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[12] Xi Chen, Zhifei Zhang, He Zhang, Yuqian Zhou, Soo Ye Kim, Qing Liu, Yijun Li, Jianming Zhang, Nanxuan Zhao, Yilin Wang, Hui Ding, Zhe Lin, and Hengshuang Zhao. UniReal: Universal Image Generation and Editing via Learning Real-world Dynamics. arXiv preprint arXiv:2412.07774, 2024. 2, 3",
+            "[13] Alibaba Cloud. Tongyi Wanxiang, https://tongyi. aliyun.com/wanxiang, 2023. 2",
+            "[14] Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: Optical FLOW-guided ATTENtion for consistent text-to-video editing. In Int. Conf. Learn. Represent., 2024. 6, 7, 13",
+            "[15] Lunhao Duan, Shanshan Zhao, Wenjun Yan, Yinglun Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, Mingming Gong, and Gui-Song Xia. UNIC-Adapter: Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation. arXiv preprint arXiv:2412.18928, 2024.3",
+            "[16] 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, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. In Int. Conf. Mach. Learn., 2024. 2",
+            "[17] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierarchical Masked 3D Diffusion Model for Video Outpainting. In ACM Int. Conf. Multimedia, pages 7890-7900, 2023. 6, 13",
+            "[18] FLUX. FLUX, https://blackforestlabs.ai/, 2024.2",
+            "[19] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. SEED-Data-Edit Technical Report: A Hybrid Dataset for Instructional Image Editing. arXiv preprint arXiv:2405.04007, 2024. 2",
+            "[20] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Pengfei Wan, Di Zhang, Yufan Liu, Weiming Hu, Zhengjun Zha, Haibin Huang, and Chongyang Ma. 12V-Adapter: A General Image-to-Video Adapter for Diffusion Models. In ACM SIGGRAPH, pages 1-12, 2024. 2",
+            "[21] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, Peng Zhang, and Qian He. PuLID: Pure and Lightning ID Customization via Contrastive Alignment. In Adv. Neural Inform. Process. Syst., 2024. 2",
+            "[22] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. LTX-Video: Realtime Video Latent Diffusion. arXiv preprint arXiv:2501.00103, 2025. 2, 6, 13",
+            "[23] Zhen Han, Zeyinzi Jiang, Yulin Pan, Jingfeng Zhang, Chaojie Mao, Chenwei Xie, Yu Liu, and Jingren Zhou. ACE: All-round Creator and Editor Following Instructions via Diffusion Transformer. In Int. Conf. Learn. Represent., 2025. 2, 3"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            897
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "17199",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[24] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. In Adv. Neural Inform. Process. Syst., 2021. 2",
+            "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In Adv. Neural Inform. Process. Syst. Curran Associates, Inc., 2020. 2",
+            "[26] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and Controllable Image Synthesis with Composable Conditions. In Int. Conf. Mach.Learn., 2023.2",
+            "[27] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive Benchmark Suite for Video Generative Models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21807-21818, 2024. 5, 7",
+            "[28] Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, Yaohui Wang, Xinyuan Chen, YingCong Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench++: Comprehensive and Versatile Benchmark Suite for Video Generative Models. arXiv preprint arXiv:2411.13503, 2024. 5",
+            "[29] Zeyinzi Jiang, Chaojie Mao, Ziyuan Huang, Ao Ma, Yiliang Lv, Yujun Shen, Deli Zhao, and Jingren Zhou. Res-Tuning: A Flexible and Efficient Tuning Paradigm via Unbinding Tuner from Backbone. In Adv. Neural Inform. Process. Syst., 2023. 5, 8",
+            "[30] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. SCEdit: Efficient and Controllable Image Diffusion Generation via Skip Connection Editing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8995-9004, 2024. 2",
+            "[31] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators. In Int. Conf. Comput. Vis., pages 15954–15964, 2023. 6, 13",
+            "[32] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyuan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Dax Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. HunyuanVideo: A Systematic Framework For Large Video Generative Models. arXiv preprint arXiv:2412.03603, 2024. 2",
+            "[33] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding. arXiv preprint arXiv:2405.08748, 2024. 2"
+        ],
+        "bbox": [
+            89,
+            92,
+            482,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[34] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. MagicEdit: High-Fidelity and Temporally Coherent Video Editing. arXiv preprint arXiv:2308.14749, 2023. 3",
+            "[35] Lijie Liu, Tianxiang Ma, Bingchuan Li, Zhuowei Chen, Jiawei Liu, Qian He, and Xinglong Wu. Phantom: Subject-consistent video generation via cross-modal alignment. arXiv preprint arXiv:2502.11079, 2025. 2, 3",
+            "[36] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. arXiv preprint arXiv:2303.05499, 2023. 5",
+            "[37] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-P2P: Video Editing with Cross-attention Control. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8599–8608, 2024. 3",
+            "[38] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept Neurons in Diffusion Models for Customized Generation. In Int. Conf. Mach. Learn., 2023. 2",
+            "[39] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones 2: Customizable Image Synthesis with Multiple Subjects. In Adv. Neural Inform. Process. Syst., 2023. 2",
+            "[40] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Ying Shan, Xiu Li, and Qifeng Chen. Follow Your Pose: Pose-Guided Text-to-Video Generation using Pose-Free Videos. In Assoc. Adv. Artif. Intell., 2024. 6, 13",
+            "[41] Chaojie Mao, Jingfeng Zhang, Yulin Pan, Zeyinzi Jiang, Zhen Han, Yu Liu, and Jingren Zhou. ACE++: Instruction-Based Image Creation and Editing via Context-Aware Content Filling. arXiv preprint arXiv:2501.02487, 2025. 2, 3",
+            "[42] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. In Int. Conf. Learn. Represent., 2021. 2",
+            "[43] Midjourney. Midjourney, https://www.midjourney.com, 2023.2",
+            "[44] MiniMax. Hailuo AI Video, https://hailuoai.com/video, 2024.",
+            "[45] OpenAI. DALL·E 3, https://openai.com/dall-e-3, 2023.2",
+            "[46] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold. In ACM SIGGRAPH, 2023. 2",
+            "[47] Yulin Pan, Chaojie Mao, Zeyinzi Jiang, Zhen Han, and Jingfeng Zhang. Locate, Assign, Refine: Taming Customized Image Inpainting with Text-Subject Guidance. arXiv preprint arXiv:2403.19534, 2024. 2",
+            "[48] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers. In Int. Conf. Comput. Vis., pages 4195-4305, 2023. 2",
+            "[49] PiKa. PiKa, https://pika.art/, 2025.6, 7, 13",
+            "[50] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming"
+        ],
+        "bbox": [
+            514,
+            92,
+            903,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "17200",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Xiong, Silvio Savarese, Stefano Ermon, Yun Fu, and Ran Xu. UniControl: A Unified Diffusion Model for Controllable Visual Generation In the Wild. In Adv. Neural Inform. Process. Syst., 2023. 3",
+            "[51] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer. IEEE Trans. Pattern Anal. Mach. Intell., pages 1623-1637, 2022. 5",
+            "[52] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. SAM 2: Segment Anything in Images and Videos. In Int. Conf. Learn. Represent., 2025. 5",
+            "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684-10695, 2022. 2",
+            "[54] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. Med. Image Comput. Computer-Assisted Interv., 2015. 2",
+            "[55] Runway. Gen-3, https://app.runwayml.com/video-tools, 2025.2",
+            "[56] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models. In Int. Conf. Learn. Represent., 2021. 2",
+            "[57] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. In Int. Conf. Learn. Represent., 2021. 2",
+            "[58] StabilityAI. Stable Diffusion v2-1 Model Card, https: //huggingface.co/stabilityai/stable-diffusion-2-1,2022.2",
+            "[59] StabilityAI. Stable Diffusion XL Model Card, https: //huggingface.co/stabilityai/stable-diffusion-x1-base-1.0,2022.2",
+            "[60] StabilityAI. CosXL Model Card, https://huggingface.co/stabilityai/cosxl, 2024.3",
+            "[61] Ya Sheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. ImageBrush: Learning Visual In-Context Instructions for Exemplar-Based Image Manipulation. In Adv. Neural Inform. Process. Syst., 2023. 3",
+            "[62] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-Robust Large Mask Inpainting With Fourier Convolutions. In IEEE Winter Conf. Appl. Comput. Vis., pages 2149-2159, 2022. 5",
+            "[63] Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. OminiControl: Minimal and Universal Control for Diffusion Transformer. arXiv preprint arXiv:2411.15098, 2024. 2, 3",
+            "[64] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 2, 6, 13"
+        ],
+        "bbox": [
+            91,
+            90,
+            482,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[65] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. In Eur. Conf. Comput. Vis., pages 402-419, 2020. 5",
+            "[66] Vidu. Vidu, https://www.vidu.cn/, 2025. 6, 7, 13",
+            "[67] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. InstantID: Zero-shot Identity-Preserving Generation in Seconds. arXiv preprint arXiv:2401.07519, 2024. 2",
+            "[68] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional Video Synthesis with Motion Controllability. In Adv. Neural Inform. Process. Syst., 2023. 2, 6, 13",
+            "[69] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A Unified and Flexible Motion Controller for Video Generation. In ACM SIGGRAPH, pages 1-11, 2024. 3",
+            "[70] Yujie Wei, Shiwei Zhang, Hangjie Yuan, Xiang Wang, Haonan Qiu, Rui Zhao, Yutong Feng, Feng Liu, Zhizhong Huang, Jiaxin Ye, Yingya Zhang, and Hongming Shan. DreamVideo-2: Zero-Shot Subject-Driven Video Customization with Precise Motion Control. arXiv preprint arXiv:2410.13830, 2024. 2",
+            "[71] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. OmniGen: Unified Image Generation. arXiv preprint arXiv:2409.11340, 2024. 2, 3",
+            "[72] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective Whole-body Pose Estimation with Two-stages Distillation. In Int. Conf. Comput. Vis., pages 4210-4220, 2023. 5",
+            "[73] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer. In Int. Conf. Learn. Represent., 2025. 2, 6, 13",
+            "[74] Shenghai Yuan, Jinfa Huang, Xianyi He, Yunyuan Ge, Yu-jun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. Identity-Preserving Text-to-Video Generation by Frequency Decomposition. In IEEE Conf. Comput. Vis. Pattern Recog., 2025. 2",
+            "[75] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. MagicBrush: A Manually Annotated Dataset for Instruction-Guided Image Editing. In Adv. Neural Inform. Process. Syst., 2023. 2",
+            "[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In Int. Conf. Comput. Vis., pages 3836–3847, 2023. 2",
+            "[77] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2VGen-XL: High-Quality Image-to-Video Synthesis via Cascaded Diffusion Models. arXiv preprint arXiv:2311.04145, 2023. 2, 6, 13",
+            "[78] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo,"
+        ],
+        "bbox": [
+            516,
+            90,
+            903,
+            900
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "17201",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize Anything: A Strong Image Tagging Model. arXiv preprint arXiv:2306.03514, 2023. 5",
+            "[79] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. ControlVideo: Training-free Controllable Text-to-Video Generation. In Int. Conf. Learn. Represent., 2024. 6, 13",
+            "[80] Yuechen Zhang, Yaoyang Liu, Bin Xia, Bohao Peng, Zexin Yan, Eric Lo, and Jiaya Jia. Magic Mirror: ID-Preserved Video Generation in Video Diffusion Transformers. arXiv preprint arXiv:2411.13503, 2025. 3",
+            "[81] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based Fine-Grained Image Editing at Scale. arXiv preprint arXiv:2407.05282v1, 2024. 2",
+            "[82] Shangchen Zhou, Chongyi Li, Kelvin C. K. Chan, and Chen Change Loy. ProPainter: Improving Propagation and Transformer for Video Inpainting. In Int. Conf. Comput. Vis., pages 10477-10486, 2023. 2, 6, 13"
+        ],
+        "bbox": [
+            91,
+            90,
+            480,
+            373
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "17202",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 11
+    }
+]

2025/VACE_ All-in-One Video Creation and Editing/bb0c9cf5-4f56-4fe0-bdd0-bf9454fc4819_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b659ea208919e8bc322093dde0fe027f5c2956ee57f56e559fe1aad0dedca67
+size 5291197

2025/VACE_ All-in-One Video Creation and Editing/full.md

@@ -0,0 +1,343 @@
+# VACE: All-in-One Video Creation and Editing
+
+Zeyinzi Jiang*
+
+Zhen Han*
+
+Chaojie Mao*†
+
+Jingfeng Zhang
+
+Yulin Pan
+
+Yu Liu
+
+# Tongyi Lab, Alibaba Group
+
+![](images/6dc0ac2e79e0ba9bb358ce5358c79ab788f0b963e087806e235c7b43592a9f61.jpg)  
+Reference-to-video generation
+
+![](images/01187571914a5e6d6532ea027270a9ab6354e126bbeba8577fb665a9462afeee.jpg)
+
+![](images/85545d4d442ad80243ffb8b78dc40cee7e78444acc8b7f750789189b0c03bbf4.jpg)
+
+![](images/6c9c116f5690eba3d9ee44b3b791916ff76bb2ec58621dbbe86a0ae44980aa0d.jpg)  
+Video-to-video editing  
+Pose Transfer (Pose Control)
+
+![](images/1d10ef66bf9e895e32d99c420bde8e9d16ef77d13a2666119f69ac5d62c76b41.jpg)  
+Motion Transfer (Layout Control)
+
+![](images/58ad5b44490705232a973cf7a586334b054d0bc8c494901a11fe885919a638d8.jpg)  
+Structure Transfer (Depth Control)
+
+![](images/80df6d6119b73f2d4005bce8b9c89505b8abb30c091bc04ee8329de809340e56.jpg)  
+Colorization (Gray Control)
+
+![](images/95d1d327420388d5737f00901b38c9c4081cca6ef3806a9ed2ec1ed8f43224ce.jpg)  
+Masked Video-to-video editing  
+Subject Recreation (Inpainting)
+
+![](images/0f9436c54eb21d2c2ced5d62e408773bb87a5fad5ad4a71c1a7bcecc0f3dff35.jpg)  
+Subject Removal (Inpainting)
+
+![](images/6e856a972006ce8b10c473c7deb34dd2b9911915774764a54da3b6adf70ffa9b.jpg)  
+Canvas Extension (Outpainting)
+
+![](images/f1285142f3ed6719d624eecacd52d7f643cbad1f1346811431897c717ae26fcc.jpg)  
+Temporal Extension (First Clip)
+
+![](images/5a8e8ec3fd8352db7fcaa2531b94053d11e401af6e453edc2f6ee0834d1ac198.jpg)  
+Task Composition  
+Swap Anything (Subject Reference + Inpainting)  
+Figure 1. Comprehensive capabilities of VACE. We present the outstanding generation results based on Wan2.1 (left) and LTX-Video (right). For each task, the original input image or video (left), the context video (top left corner), and the generated frames are illustrated.
+
+![](images/ca5b623143e9929a16ef37e21b1cf7268db1c9a956e3c85c812cdbf5bb037f27.jpg)  
+Animate Anything (Frame Reference + Pose Control)
+
+# Abstract
+
+Diffusion Transformer has demonstrated powerful capability and scalability in generating high-quality images and videos. Further pursuing the unification of generation and editing tasks has yielded significant progress in the domain of image content creation. However, due to the intrinsic demands for consistency across both temporal and spatial dynamics, achieving a unified approach for video synthesis remains challenging. We introduce VACE, which enables users to perform Video tasks within an All-in-one framework for Creation and Editing. These tasks include reference-to-video generation, video-to-video editing, and masked video-to-video editing. Specifically, we effectively
+
+integrate the requirements of various tasks by organizing video task inputs, such as editing, reference, and masking, into a unified interface referred to as the Video Condition Unit (VCU). Furthermore, by utilizing a Context Adapter structure, we inject different task concepts into the model using formalized representations of temporal and spatial dimensions, allowing it to handle arbitrary video synthesis tasks flexibly. Extensive experiments demonstrate that the unified model of VACE achieves performance on par with task-specific models across various subtasks. Simultaneously, it enables diverse applications through versatile task combinations. Project page: https://alivilab.github.io/VACE-Page/.
+
+# 1. Introduction
+
+In recent years, the domain of visual generation tasks has witnessed remarkable advancements, driven in particular by the rapid evolution of diffusion models [24, 25, 48, 53, 54, 56, 57]. Beyond the early foundational pre-trained models for text-to-image [7, 16, 33] or text-to-video [9, 22, 64] generation in the field, there has been a proliferation of downstream tasks and applications, such as painting [3, 82], editing [4, 42, 68, 70, 75], controllable generation [30, 76], frame reference generation [20, 73], and ID-referenced video synthesis [11, 35, 47, 74]. This array of developments highlights the dynamic and complex nature of the visual generation field. To enhance task flexibility and reduce the overhead associated with deploying multiple models, researchers have begun to focus on constructing unified model architectures [12, 63] (e.g., ACE [23, 41] and Omni-Gen [71]), aiming to integrate different tasks into a single image model, facilitating the creation of various application workflows while maintaining simplicity in usage. In the field of video, due to the collaborative transformations in both temporal and spatial dimensions, leveraging a unified model can present endless possibilities for video creation. However, leveraging diverse input modalities and ensuring spatiotemporal consistency are still challenging for unified video generation and editing.
+
+We propose VACE, an all-in-one model for video creation and editing that performs tasks including reference-to-video generation, video-to-video editing, masked video-to-video editing, and free composition of these tasks, as illustrated in Fig. 1. The aggregation of various capabilities reduces the costs of service deployment and user interaction, and by combining the capabilities of different tasks within a single model, it addresses challenges faced by existing video generation models, such as controllable generation of long videos, multi-condition and reference-based generation, and continuous video editing, thereby empowering users with greater creativity. To achieve this, we utilize the current mainstream Diffusion Transformers (DiTs) structure as the foundational video framework and pre-trained text-to-video generation models [22, 64], which provide better basic capabilities and scalability for handling long video sequences. Specifically, VACE takes into account the needs of different tasks during its construction and designs a unified interface, dubbed the Video Condition Unit (VCU), which integrates multiple modalities such as images or videos for editing, references, and masks. Furthermore, to differentiate the visual modality information in editing and reference tasks, we introduce the concept decoupling strategy, enabling the model to understand what aspects need to be retained and what should be modified. Meanwhile, by employing a pluggable Context Adapter structure, concepts from different tasks (e.g., the areas or ranges of editing or reference) are injected into the model through col
+
+laborative spatiotemporal representation, enabling it to possess the capability of adaptive processing for unified tasks.
+
+Due to the lack of existing multi-task benchmarks in video synthesis, we construct a dataset of 480 evaluation samples containing 12 different tasks, while evaluating the performance of the VACE unified model by comparing it with existing specialized models. Experimental results demonstrate that our framework exhibits sufficient competitiveness in both quantitative and qualitative analyses. To the best of our knowledge, we are the first all-in-one model based on the video DiT architecture that concurrently supports such a wide range of tasks. Notably, this innovative framework allows for the compositional expansion of base tasks, enabling the construction of scenarios such as long video re-rendering, which provides a versatile and efficient solution for video synthesis, opening new possibilities for user-side video content creation and editing.
+
+# 2. Related Work
+
+Visual Generation and Editing. With the rapid development of image [2, 7, 16, 18, 58, 59] and video [22, 32, 73, 77] generation models, they are being used to create high-quality visual content and are widely applied in fields such as advertising, film special effects, game development, and animation production [13, 43-45, 55]. Meanwhile, to meet the diverse needs of visual media production and to enhance efficiency and quality, precise generation and editing methods have emerged. Models are required to perform generative creation based on multimodal inputs, such as depth, structure, pose, scene, and characters. According to the purposes of the input conditions, we can categorize them into two types: editing of the input and concept-guided re-creation. A significant portion of the work, such as ControlNet [76], ControlVideo,Composer [26], VideoComposer [68], and SCEdit [30], focuses on single-condition editing and multi-condition composite editing based on temporal and spatial alignment conditions. Additionally, some works that focus on interactive local editing scenarios, such as DragGAN [46] and MagicBrush [75]. Methods that guide generation based on semantic information from the input, such as Cone [38], Cone2 [39], InstantID [67], and PuLID [21], can achieve conceptual understanding of the input and inject it into the model for creative purposes.
+
+Task-unified Visual Generative Model. As the complexity and diversity of user creations increase, relying solely on a single model or a complicated chain of multiple models can no longer provide a convenient and efficient path for implementing creative ideas. In image generation, a unified generation and editing framework has begun to emerge, allowing for more flexible creative approaches. Methods such as UltraEdit [81] and SEED-Data-Edit [19] provide general-purpose editing datasets, while techniques like In
+
+![](images/2bdea4dc0e5c25dd5529b7d9f6e57e9084415154530503123357e880e21f4386.jpg)  
+Figure 2. Task categories covered by VACE. Four basic tasks can be combined to create a vast number of possibilities.
+
+structPix2Pix [4], MagicBrush [61], and CosXL [60] offer general instruction-based editing features. Additionally, methods like UniControl [50] and UNIC-Adapter [15] have unified controllable generation. Further advancements have led to the development of ACE [23, 41], OmniGen [71], OmniControl [63], and UniReal [12], which expand the scope of tasks by providing flexible controllable generation, local editing, and reference-guided generation. In the video domain, due to the increased difficulty of generation, approaches often manifest as single-task single-model frameworks, offering capabilities for editing or reference generation, as seen in Video-P2P [37], MagicEdit [34], MotionCtrl [69], Magic Mirror [80], and Phantom [35]. VACE aims to fill the gap for a unified model within the video domain, providing possibilities for complex creative scenarios.
+
+# 3. Method
+
+VACE is designed as a multimodal-to-video generation model, where text, image, video, and mask are integrated into a unified conditioning input. To cover as many video generation and editing tasks as possible, we conduct in-depth research into existing tasks, then divide them into 4 categories according to their individual requirements of multimodal inputs. Without losing generality, we specifically design a novel multimodal input format for each category under a Video Condition Unit (VCU) paradigm. Finally, we restructure the DiT model for VCU inputs, making it a versatile model for a wide range of video tasks.
+
+# 3.1. Multimodal Inputs and Video Tasks.
+
+Despite existing video tasks varying in complex user inputs and ambitious creative goals, we found that most of their inputs can be fully represented in 4 modalities: text, image, video, and mask. Overall, as illustrated in Fig. 2, we group these video tasks into 5 categories based on their requirements of these four multimodal inputs.
+
+- Text-to-Video Generation (T2V) is a basic video creation task, and text is the only input.  
+- Reference-to-Video Generation (R2V) requires additional images as reference inputs, making sure that specified contents, such as subjects of faces, animals, and other objects, or video frames, appear in the generated video.  
+- Video-to-Video Editing (V2V) makes an entire change to a provided video, such as colorization, stylization, controllable generation, etc. We use video control types whose control signals can be represented and stored as RGB videos, including depth, grayscale, pose, scribble, optical flow, and layout; however, the method itself is not limited to these.  
+- Masked Video-to-Video Editing (MV2V) makes changes to an input video only within the provided 3D regions of interest (3D ROI), seamlessly blending in with the other unchanged regions, such as inpainting, outpainting, video extension, etc. We use an extra spatiotemporal mask to represent the 3D ROI.  
+- Task Composition includes all the compositional possibilities of the 4 types of video tasks above.
+
+# 3.2. Video Condition Unit
+
+We propose an input paradigm, Video Condition Unit (VCU), to unify diverse input conditions into textual input, frame sequence, and mask sequence. A VCU can be denoted as
+
+$$
+V = [ T; F; M ], \tag {1}
+$$
+
+where  $T$  is a text prompt, while  $F$  and  $M$  are sequences of context video frames  $\{u_1, u_2, \dots, u_n\}$  and masks  $\{m_1, m_2, \dots, m_n\}$  respectively. Here,  $u$  is in RGB space, normalized to  $[-1, 1]$  and  $m$  is binary, in which "1"s and "0"s symbolize where to edit or not.  $F$  and  $M$  are aligned both in spatial size  $h \times w$  and temporal size  $n$ . In T2V, no context frame or mask is required. To keep generality, we assign a default value  $0_{h \times w}$  to each  $u$  denoting empty input, and set every  $m$  to  $1_{h \times w}$  meaning that all these pixels
+
+![](images/2a2cdb8973973f507ef1873f7f8e7f67b97950d6faa55f40b8c825386eb1ca8e.jpg)  
+Figure 3. Overview of VACE Framework. Frames and masks are tokenized through Concept Decoupling, Context Latent Encode, and Context Embedder. To achieve training with VCU as input, we employ two strategies: (a) Fully Fine-tuning and (b) Context Adapter Tuning. The latter converges faster and supports pluggable features.
+
+![](images/572bd3aa43bbd8d28bf5ab05a3e7a83f3135cec10cac1bc8df624e2eb426a3ba.jpg)
+
+Table 1. The formal representation of frames  $(F\mathrm{s})$  and masks  $(M\mathrm{s})$  under the four basic tasks. Frames and masks are aligned spatially and temporally.  
+
+<table><tr><td>Tasks</td><td>Frames (Fs) &amp; Masks (Ms)</td></tr><tr><td>T2V</td><td>F = {0h×w} × n
+M = {1h×w} × n</td></tr><tr><td>R2V</td><td>F = {r1, r2, ..., rl} + {0h×w} × n
+M = {0h×w} × l + {1h×w} × n</td></tr><tr><td>V2V</td><td>F = {u1, u2, ..., un}
+M = {1h×w} × n</td></tr><tr><td>MV2V</td><td>F = {u1, u2, ..., un}
+M = {m1, m2, ..., mn}</td></tr></table>
+
+are about to be re-generated. For R2V, additional reference frames  $r_i$  are inserted in front of the default frame sequence, while all-zero masks  $0_{h \times w}$  are inserted in front of the mask sequence. These all-zero masks mean that the corresponding frames should be kept unchanged. In V2V, the context frame sequence is the input video frames, and the context mask is a sequence of  $1_{h \times w}$ . For MV2V, both context video and mask are required. The formal mathematical representations are shown in Tab. 1.
+
+VCU can also support task composition. For example, the context frames of reference-inpainting task are  $\{r_1,r_2,\dots,r_l,u_1,u_2,\dots,u_n\}$  and the context masks are  $\{0_{h\times w}\} \times l + \{m_1,m_2,\dots,m_n\}$ . In this case, users can modify  $l$  objects in the video and regenerate based on the provided reference images. For another example, users only have a scribble image and want to generate a video beginning with the contents described by this scribble image, which is a scribble-based video extension task. The context frames are  $\{u\} +\{0_{h\times w}\} \times (n - 1)$  and the context masks are  $\{1_{h\times w}\} \times n$ . In this way, we can achieve multi-condition and reference control generation for long videos.
+
+# 3.3. Architecture
+
+We restructure the DiT model for VACE, as shown in Fig. 3, aiming to support multimodal VCU inputs. Since there is an existing pipeline for text tokenization, we only consider the tokenization of context frames and masks. After being tokenized, the context tokens are combined with noisy video tokens and used to fine-tune the DiT model. Different from that, we also propose a Context Adapter Tuning strategy, which allows context tokens to pass through Context Blocks and be added back to the original DiT Blocks.
+
+# 3.3.1. Context Tokenization
+
+Concept Decoupling. Two different visual concepts of natural video and control signals, like depth, pose, are encoded in  $F$  simultaneously. We believe that explicitly separating these data of different modalities and distributions is essential for model convergence. The concept decoupling is based on masks and yields two frame sequences identical in shape:  $F_{c} = F \times M$  and  $F_{k} = F \times (1 - M)$ , where  $F_{c}$  is called the reactive frames contain all the pixels to be changed, while all the pixels to be kept are stored in  $F_{k}$ , named inactive frames. Specifically, the reference images and the unchanged part of V2V and MV2V go to  $F_{k}$ , while control signals and those pixels about to change, such as gray pixels, are collected to  $F_{c}$ .
+
+Context Latent Encoding. A typical DiT processes noisy video latents  $X \in \mathbb{R}^{n' \times h' \times w' \times d}$ , where  $n'$ ,  $h'$ , and  $w'$  are the temporal and spatial shapes of the latent space. Similar to  $X$ ,  $F_{c}$ ,  $F_{k}$ , and  $M$  need to be encoded into a high-dimensional feature space to ensure the property of significant spatiotemporal correlations. Therefore, we reorganize them together with  $X$  into a hierarchical and spatiotemporal aligned visual features.  $F_{c}$ ,  $F_{k}$  are processed by video VAE and mapped into the same latent space of  $X$ , maintaining their spatiotemporal consistency. To avoid any unexpected mishmash of images and videos, reference images are separately encoded by VAE encoder and concatenated back along the temporal dimension, while the correspond
+
+ing parts need to be removed during decoding.  $M$  is directly reshaped and interpolated. After that,  $F_{c}, F_{k}$ , and  $M$  are all mapping into latent spaces and are spatiotemporal aligned with  $X$  in the shape of  $n' \times h' \times w'$ .
+
+Context Embedder. We extend the embedder layer by concatenating  $F_{c}$ ,  $F_{k}$ , and  $M$  in the channel dimension and tokenizing them into context tokens, which we refer to as the Context Embedder. The corresponding weights to tokenize  $F_{c}$  and  $F_{k}$  are directly copied from the original video embedder, and the weights to tokenize  $M$  are initialized to zeros.
+
+# 3.3.2. Fully Fine-Tuning and Context Adapter Tuning
+
+To achieve training with VCU as input, a simple methodology is to fully fine-tune the whole DiT model, as shown in Fig. 3a. Context tokens are added together with noisy tokens  $X$ , and all the parameters in DiT and the newly introduced Context Embedder will be updated during training. To avoid fully fine-tuning and achieve faster convergence, as well as to establish a pluggable feature with the foundation model, we also propose another methodology processing the context token in a Res-Tuning [29] manner, as shown in Fig. 3b. Particularly, we select and copy several Transformer Blocks from the original DiT, forming a distributed and cascade-type Context Blocks. The original DiT processes video tokens and text tokens, while the newly added Transformer Blocks process context tokens and text tokens. The output of each Context Block is inserted back into the DiT blocks as an additive signal to assist the main branch in performing generation and editing tasks. In this manner, the parameters of DiT are frozen. Only the Context Embedder and Context Blocks are trainable.
+
+# 4. Datasets
+
+# 4.1. Data Construction
+
+To obtain an all-in-one model, the diversity and complexity of the required data construction also increase. Existing common text-to-video and image-to-video tasks only require constructing pairs of text and video. However, for the tasks in VACE, the modalities need to be further expanded to include target videos, source videos, local masks, reference, and more. To efficiently and rapidly acquire data for various tasks, it is imperative to maintain video quality while also conducting instance-level analysis and understanding of the video data.
+
+To achieve this, we first analyze the video data itself by performing shot slicing and preliminarily filtering out data based on resolution, aesthetic score, and motion amplitude. Next, we label the first frame of the video using RAM [78] and combine it with Grounding DINO [36] for detection, utilizing the localization results to perform secondary filtering on videos with target areas that are either too small or too large. Furthermore, we employ the propagate operation
+
+of SAM2 [52] for video segmentation to obtain instance-level information across the video. Leveraging the results of video segmentation, we filter instances in the temporal dimension by calculating the effective frame ratio based on the mask area threshold.
+
+In the actual training process, the construction for different tasks also needs to be tailored according to the characteristics of each task: 1) For some controllable video generation tasks, we pre-extract depth [51], scribble [6], pose [5, 72], and optical flow [65] from the filtered videos. For gray and layout tasks, we create data on the fly. 2) For repainting tasks, random instances from the videos can be masked for inpainting, while the inverse of the mask enables the construction of outpainting data. Augmentation of the masks [62] allows for unconditional inpainting. 3) In the case of extension tasks, we extract key frames such as the first frame, last frame, frames from both ends, random frames, and segments from both ends to support a wider variety of extension types. 4) For reference tasks, we can extract several face or object instances from the videos and apply offline or online augmentation operations to create paired data. Notably, we randomly combine all the previously mentioned tasks for training to accommodate a broader range of model application scenarios. Additionally, for all operations involving masks, we perform arbitrary augmentation to satisfy various granular local generation requirements.
+
+# 4.2. VACE-Benchmark
+
+Significant progress has been made in the field of video generation. However, a scientific and thorough evaluation of the performance of these models remains an urgent issue that needs to be addressed. VBench [27] and VBench++ [28] have established a precise evaluation framework for text-to-video and image-to-video tasks through an extensive assessment suite and dimensional design. Nevertheless, as the video generation ecosystem continues to evolve, more derivative tasks have begun to emerge, such as video reference generation and video editing, for which a comprehensive benchmark is still lacking. To address this gap, we propose VACE-Benchmark to evaluate various downstream tasks related to video in a systematic manner.
+
+Starting from the data sources, we recognize that real videos and generated videos may exhibit different performance characteristics during evaluation. Thus, we collected a total of 240 high-quality videos categorized by their sources, encompassing various data types, including text-to-video, inpainting, outpainting, extension, grayscale, depth, scribble, pose, optical flow, layout, reference face, and reference object tasks, with an average of 20 samples for each task. The input modalities include videos, masks, and references, and we also provide the original videos
+
+Table 2. Quantitative evaluations on VACE-Benchmark. We compare the automated score metrics of the unified VACE based on LTX-Video and the proprietary model on the dimensions of video quality and video consistency, as well as results of human user studies.  
+
+<table><tr><td rowspan="2">Type</td><td rowspan="2">Method</td><td colspan="10">Video Quality &amp; Video Consistency</td><td colspan="4">User Study</td></tr><tr><td>Aesthetic Quality</td><td>Background Consistency</td><td>Dynamic Degree</td><td>Imaging Quality</td><td>Motion Smoothness</td><td>Overall Consistency</td><td>Subject Consistency</td><td>Temporal Flickering</td><td>Normalized Average</td><td>Prompt Following</td><td>Temporal Consistency</td><td>Video Quality</td><td>Average</td><td></td></tr><tr><td rowspan="4">I2V</td><td>I2VGenXL [77]</td><td>55.20%</td><td>92.87%</td><td>60.00%</td><td>63.31%</td><td>97.43%</td><td>23.78%</td><td>89.58%</td><td>95.67%</td><td>71.54%</td><td>2.65</td><td>1.60</td><td>2.34</td><td>2.20</td><td></td></tr><tr><td>CogVideoX-I2V [73]</td><td>57.78%</td><td>94.80%</td><td>40.00%</td><td>68.23%</td><td>98.69%</td><td>24.38%</td><td>93.84%</td><td>97.84%</td><td>73.66%</td><td>3.30</td><td>2.28</td><td>3.19</td><td>2.92</td><td></td></tr><tr><td>LTX-Video [22]</td><td>56.12%</td><td>94.57%</td><td>35.00%</td><td>62.72%</td><td>99.27%</td><td>24.92%</td><td>92.83%</td><td>98.41%</td><td>72.89%</td><td>2.95</td><td>2.28</td><td>2.28</td><td>2.50</td><td></td></tr><tr><td>VACE (Ours)</td><td>57.53%</td><td>95.32%</td><td>45.00%</td><td>68.03%</td><td>99.08%</td><td>25.13%</td><td>93.61%</td><td>97.80%</td><td>74.38%</td><td>3.20</td><td>4.00</td><td>2.54</td><td>3.24</td><td></td></tr><tr><td rowspan="2">Inpaint</td><td>ProPainter [82]</td><td>44.70%</td><td>95.64%</td><td>50.00%</td><td>61.57%</td><td>99.01%</td><td>18.48%</td><td>92.99%</td><td>98.47%</td><td>70.15%</td><td>2.35</td><td>4.00</td><td>2.99</td><td>3.11</td><td></td></tr><tr><td>VACE (Ours)</td><td>51.30%</td><td>96.30%</td><td>50.00%</td><td>60.39%</td><td>99.12%</td><td>21.12%</td><td>94.59%</td><td>98.21%</td><td>72.05%</td><td>2.40</td><td>4.00</td><td>2.60</td><td>3.00</td><td></td></tr><tr><td rowspan="3">Outpaint</td><td>Follow-Your-Canvas [8]</td><td>53.30%</td><td>95.99%</td><td>5.00%</td><td>69.53%</td><td>98.08%</td><td>25.90%</td><td>95.38%</td><td>97.20%</td><td>71.54%</td><td>3.05</td><td>2.00</td><td>1.63</td><td>2.23</td><td></td></tr><tr><td>M3DDM [17]</td><td>53.34%</td><td>95.87%</td><td>30.00%</td><td>65.07%</td><td>99.22%</td><td>25.43%</td><td>93.65%</td><td>98.85%</td><td>73.16%</td><td>3.70</td><td>3.88</td><td>2.28</td><td>3.29</td><td></td></tr><tr><td>VACE (Ours)</td><td>57.04%</td><td>96.55%</td><td>30.00%</td><td>69.49%</td><td>99.20%</td><td>25.36%</td><td>94.47%</td><td>98.47%</td><td>74.25%</td><td>3.90</td><td>3.92</td><td>3.58</td><td>3.80</td><td></td></tr><tr><td rowspan="4">Depth</td><td>Control-A-Video [10]</td><td>50.62%</td><td>91.71%</td><td>70.00%</td><td>67.76%</td><td>97.58%</td><td>24.48%</td><td>88.10%</td><td>96.58%</td><td>72.35%</td><td>2.70</td><td>2.28</td><td>1.54</td><td>2.17</td><td></td></tr><tr><td>VideoComposer [68]</td><td>50.03%</td><td>94.18%</td><td>70.00%</td><td>59.44%</td><td>96.23%</td><td>24.95%</td><td>89.79%</td><td>94.38%</td><td>70.74%</td><td>2.60</td><td>2.44</td><td>2.17</td><td>2.40</td><td></td></tr><tr><td>ControlVideo [79]</td><td>63.30%</td><td>95.02%</td><td>10.00%</td><td>65.13%</td><td>96.49%</td><td>24.20%</td><td>92.29%</td><td>95.42%</td><td>70.07%</td><td>2.55</td><td>2.50</td><td>1.82</td><td>2.29</td><td></td></tr><tr><td>VACE (Ours)</td><td>56.72%</td><td>96.12%</td><td>60.00%</td><td>66.41%</td><td>98.84%</td><td>25.27%</td><td>94.09%</td><td>97.27%</td><td>74.99%</td><td>3.10</td><td>3.92</td><td>2.66</td><td>3.23</td><td></td></tr><tr><td rowspan="4">Pose</td><td>Text2Video-Zero [31]</td><td>57.63%</td><td>87.67%</td><td>100.00%</td><td>70.74%</td><td>79.65%</td><td>23.94%</td><td>84.82%</td><td>76.57%</td><td>59.69%</td><td>2.15</td><td>2.00</td><td>1.88</td><td>2.01</td><td></td></tr><tr><td>ControlVideo [79]</td><td>65.37%</td><td>94.56%</td><td>25.00%</td><td>65.28%</td><td>97.32%</td><td>25.19%</td><td>92.76%</td><td>96.82%</td><td>72.45%</td><td>2.15</td><td>1.80</td><td>2.03</td><td>1.99</td><td></td></tr><tr><td>Follow-Your-Pose [40]</td><td>48.79%</td><td>86.80%</td><td>100.00%</td><td>67.41%</td><td>90.12%</td><td>26.10%</td><td>80.18%</td><td>88.02%</td><td>66.43%</td><td>2.00</td><td>2.60</td><td>1.58</td><td>2.06</td><td></td></tr><tr><td>VACE (Ours)</td><td>60.17%</td><td>94.92%</td><td>75.00%</td><td>64.71%</td><td>98.63%</td><td>26.44%</td><td>94.82%</td><td>96.60%</td><td>76.13%</td><td>2.95</td><td>3.96</td><td>2.63</td><td>3.18</td><td></td></tr><tr><td rowspan="2">Flow</td><td>FLATTEN [14]</td><td>56.23%</td><td>95.80%</td><td>70.00%</td><td>61.65%</td><td>97.86%</td><td>26.23%</td><td>93.94%</td><td>96.17%</td><td>74.42%</td><td>3.50</td><td>2.40</td><td>3.19</td><td>3.03</td><td></td></tr><tr><td>VACE (Ours)</td><td>55.76%</td><td>96.07%</td><td>75.00%</td><td>65.37%</td><td>98.98%</td><td>25.89%</td><td>94.63%</td><td>96.93%</td><td>75.90%</td><td>2.90</td><td>3.75</td><td>2.60</td><td>3.08</td><td></td></tr><tr><td rowspan="4">R2V</td><td>Kling1.6 [1]</td><td>62.13%</td><td>96.04%</td><td>85.00%</td><td>69.27%</td><td>99.38%</td><td>27.82%</td><td>93.79%</td><td>97.79%</td><td>78.81%</td><td>4.22</td><td>4.10</td><td>3.80</td><td>4.04</td><td></td></tr><tr><td>Pika2.2 [49]</td><td>62.48%</td><td>96.79%</td><td>65.00%</td><td>69.87%</td><td>99.37%</td><td>26.02%</td><td>95.93%</td><td>98.90%</td><td>77.87%</td><td>4.00</td><td>3.85</td><td>3.87</td><td>3.91</td><td></td></tr><tr><td>Vidu2.0 [66]</td><td>64.30%</td><td>96.85%</td><td>35.00%</td><td>67.03%</td><td>99.66%</td><td>26.53%</td><td>96.73%</td><td>99.41%</td><td>76.47%</td><td>3.90</td><td>3.85</td><td>3.77</td><td>3.84</td><td></td></tr><tr><td>VACE (Ours)</td><td>63.25%</td><td>98.03%</td><td>30.00%</td><td>72.29%</td><td>99.51%</td><td>25.85%</td><td>98.54%</td><td>99.15%</td><td>76.76%</td><td>3.47</td><td>3.42</td><td>3.30</td><td>3.40</td><td></td></tr></table>
+
+to enable further processing by developers based on the specific characteristics of each task. Regarding the data prompts, we supply the original captions of the videos for quantitative assessment, as well as rewritten prompts tailored to the specific tasks to evaluate the models' creativity.
+
+# 5. Experiments
+
+# 5.1. Experimental Setup
+
+Implementation Details. VACE is trained based on Diffusion Transformers for text-to-video generation at various scales. It utilizes LTX-Video-2B [22] for faster generation, while Wan-T2V-14B [64] is used specifically for higher-quality outputs, supporting resolutions of up to 720p. The training employs a phased approach. Initially, we focus on foundational tasks such as inpainting and extension, which are considered modal complementary to the pre-trained text-to-video models. This includes the incorporation of masks and the learning of contextual generation in both spatial and temporal dimensions. Next, from a task expansion perspective, we progressively transition from single input reference frames to multiple input reference frames and from single tasks to composite tasks. Fi
+
+nally, we fine-tune the model's quality using higher-quality data and longer sequences. The input for model training accommodates arbitrary resolutions, dynamic durations, and variable frame rates to support diverse input needs of users.
+
+Baselines. Our goal is to achieve the unification of video creation and editing tasks, and currently, there is no comparable all-in-one video generation model available, which leads us to focus our evaluation on comparing our general model with proprietary task-specific models. Moreover, due to the numerous tasks involved and the lack of open-sourced methods for many of them, we conduct our comparisons on models that are available either offline or online. Specifically for the tasks, we compare the following: 1) For the I2V task, we examine I2VGenXL [77], CogVideoX-I2V [73], and LTX-Video-I2V [22]; 2) In the repainting task, we compare the ProPainter [82] for removal inpainting, while Follow-Your-Canvas [8] and M3DDM [17] are compared for outpainting; 3) For controllable task, we use Control-A-Video [10], VideoComposer [68], and ControlVideo [79] under depth conditions, and compare Text2Video-Zero [31], ControlVideo [79], and Follow-Your-Pose [40] under pose conditions, as well as FLAT
+
+![](images/c7fa9ea987d1bb71f5051d3e521300faee221cee1c690174509157a93ceb2484.jpg)  
+Reference Anything
+
+![](images/f6416d4a007a14dd68aefc6bfd0f1fe6cb773c3b035d97636f24ebb9a0c54d90.jpg)  
+The elegant lady carefully selects bags in the boutique, and she shows the charm of a mature woman in a black slim dress with a pearl necklace. Holding a vintage-inspired brown leather handbag, she is carefully observing its craftsmanship and texture. The interior of the store.
+
+![](images/8434dacb2e4a7709e3f78e2229e213784dba4a91b4b1f46122d498181dd52856.jpg)
+
+![](images/7755152ca82d7f0fbc13b2b7b203bfcb16d5f3adeb3ff41ce7b58ef1ace9a95a.jpg)  
+A young boy rises from his chair and walks briskly to the right side of the frame towards the edge of the sun-drenched frame, as if chasing a new adventure. His eyes were bright, and the corners of his mouth were slightly upturned, revealing curiosity and excitement about the unknown...
+
+Move Anything  
+![](images/4a2a7dc2056bb9eefe68f4fd0f438de07109f6e820e51ec5c8399a3d85f38d3d.jpg)  
+"Anime-style, hot-blooded teenager in bright orange long-sleeved pants sportswear, standing on a surfboard, facing the golden sunshine in the rough sea. The teenager's short yellow hair is flying in the wind, his eyes are firm, and he has a confident smile on the corner of his mouth."
+
+![](images/2ddbe8eeb9b05ec19d551d6405830a378b3c62dc317578be5cdea44cf0957355.jpg)
+
+Animate Anything  
+![](images/04fcf82eeef0b0e26c17bc7b5f184eadced2d8d689a0a494b4537ebd63f4d9cb.jpg)  
+The video shows an old movie-style scene in retro tones, with a little penguin and a kitten having a joyous bike race. Little penguins and kittens, both dressed in orange-and-red race suits, ride vintage multi-wheeled bikes on a nostalgic dirt road flanked by spectators...
+
+![](images/a20893df72f446a06fe28fb383ae5b5f8fa37e8addd8069ae865846de816e160.jpg)
+
+![](images/5105cdd07923abce73ff9a89c0e5639e4918777310ce414986bec976cd654a25.jpg)  
+Swap Anything
+
+Expand Anything  
+Figure 4. Visualization results of compositional tasks. VACE creatively enables reference-, move-, animate-, swap-, and expand-anything.  
+![](images/9751db892a0a52597f8e5f5d005db3142b8067aa9d728a4fd80dd8f895ed60a5.jpg)  
+"In the style of classical oil painting, the background is a river, and in the center of the picture is a mature and elegant woman, wearing a long skirt and sitting on a chair. She took the open red heart-shaped sunglasses from her arms with both hands and put them on..."
+
+TEN [14] under optical flow conditions; 4) In reference generation, given the absence of open-sourced models, we compare commercial products Kling1.6 [1], Pika2.2 [49], and Vidu2.0 [66].
+
+Evaluation. To comprehensively evaluate the performance of various tasks, we employ the VACE-Benchmark for assessment. Specifically, we divide the evaluation into automatic scoring and a user study for manual assessment. For the automatic scoring, we utilize select metrics from VBench [27] to assess video quality and video consistency, including eight indicators: aesthetic quality, background consistency, dynamic degree, imaging quality, motion smoothness, overall consistency, subject consistency, and temporal flickering. For the manual assessment, we utilize the mean opinion score (MOS) as our evaluation metric, focusing on three aspects: prompt following, temporal consistency, and video quality. In practice, we anonymize the generated data and randomly distribute it to different participants for scoring on a scale from 1 to 5.
+
+# 5.2. Main Results
+
+Quantitative Evaluation. We compare VACE comprehensive model based on LTX-Video with task proprietary approaches on VACE-Benchmark. For certain tasks, we follow existing methods; for example, although we support generating based on any frame, we conduct comparisons using the first-frame reference approach from current open-source methods to ensure fairness. From Tab. 2, we can see that for the tasks of I2V, inpainting, outpainting, depth, pose, and optical flow, our method demonstrates better performance than other open-source methods across eight indicators of video quality and video consistency, with normalized average metrics showing superior results. Some competing methods can only generate at a resolution of 256, have very short generation durations, and exhibit instability in temporal coherence, resulting in poorer performance on automatic metric calculations. For the R2V task, there is still a certain gap in metrics compared to commercial models for a small-scale model that aims for fast generation,
+
+![](images/cfdbd7dd56c30c3a6af8e7f6a707e50ca6f36261f186b92d725ce4882711d134.jpg)  
+(a) Base structure setting.
+
+![](images/271e1884078ac46369e29ea040268eec90e4a0725fcd2514331506d8f08d1a3c.jpg)  
+(b) Hyperparameter settings.
+
+![](images/95c6ecc97504c4d853d22a32ff59fd3c44365b48d077815be8a79fc91f3a7089.jpg)  
+(c) Context adapter configurations.
+
+![](images/14fdebd9ed061f7d0e1c6ae123885833ea7a70db27684d65351cc2fd5fe7e5d8.jpg)  
+(d) Concept decouple setting.  
+Figure 5. Ablation Studies of the VACE regarding structures, hyperparameters, and module configurations.
+
+while being comparable to the metrics of Vidu 2.0. According to the results of human user studies, our method consistently performs better in evaluation metrics across multiple tasks, aligning well with user preferences.
+
+Qualitative Results. In Fig. 1, we present the results of the VACE single model across various tasks. It is evident that the model achieves a high level of performance in video quality and temporal consistency. Furthermore, in composition tasks shown in Fig. 4, our model showcases impressive abilities, effectively integrating different modalities and tasks to produce results that cannot be generated by existing single or multiple models, thereby demonstrating its strong potential in the fields of video generation and editing. For example, in the "Move Anything" case, by providing a single input image and a movement trajectory, we are able to precisely move the characters in the scene with a specified direction while maintaining coherence and narrative consistency.
+
+# 5.3. Ablation Studies
+
+To better understand the impact of different independent modules on a unified video generation framework, we conducted a series of systematic comparative experiments based on the LTX-Video model to achieve a better model structure and configuration. To accurately assess the different experimental settings, we sample 250 data points for each task as a validation set and calculate the training loss, reflecting the model's training progress through the mean curve changes of different tasks.
+
+Base Structure. Text-guided image or video generation models only take noise as inference input. When extended
+
+to our unified input paradigm, VCU, we can conduct training using fully fine-tuning or by incorporating additional parameter fine-tuning. Specifically, as shown in Fig. 5a, we compare the concatenation of different inputs along the channel dimension and modify the input dimensions of the patchify projection layer to achieve the loading and fully fine-tuning of the pre-trained model. Additionally, we introduce some extra training parameters in the form of RestTuning [29], which serializes VCU in a bypass branch and injects information into the main branch. The results indicate that both methods yielded similar effects; however, since the additional parameter fine-tuning converges faster, we base our subsequent experiments on this approach. As shown in Fig. 5b, we further conduct hyperparameter experiments based on this structure, focusing on aspects such as weighting schemes, timestamp shifting, and p-zero.
+
+Context Adapter. Since the number of context blocks will significantly affect the model size and inference time consumption, we attempt to find an optimal number and distribution of context blocks. We begin with selecting continuous blocks at the input side and make comparisons between the first 1/4 blocks, 1/2 blocks, and all blocks. Inspired by the Res-Tuning [29] method, we also experiment with evenly distributing the injection blocks instead of selecting a continuous block series. As shown in Fig. 5c, we can see that when using the same number of blocks, the distributed arrangement of blocks outperforms the continuous arrangement in shallow blocks. Furthermore, a greater number of blocks generally yields better results, but due to the limited improvement in effectiveness and the constraints of training resources, we adopt a partially distributed arrangement of blocks.
+
+Concept Decouple. During training, we introduce a Concept Decouple processing module to further disassemble the visual units, clarifying what content the model needs to learn to modify or retain. As shown in Fig. 5d, using this module results in a more significant reduction in loss.
+
+# 6. Conclusion
+
+This paper introduces VACE, an all-in-one video generation and editing framework. It unifies the diverse and complex multimodal inputs required for various video tasks, bridging the gap between specialized models for each individual task. This enables most video AI creation tasks to be completed with a single inference of a single model. While broadly covering various video tasks, VACE also supports flexible and free combinations of these tasks, greatly expanding the application scenarios of video generation models and meeting a wide range of user creative needs. The VACE framework paves the way for the development of unified visual generative models with multimodal inputs and represents a significant milestone in the field of visual generation.
+
+Acknowledgments. We would like to express our sincere appreciation for the contributions of many colleagues for their insightful discussions, valuable suggestions, and constructive feedback, including: Yuwei Wang, Haiming Zhao, Chenwei Xie, and Sheng Yao for their data contributions, and Shiwei Zhang, Tao Fang, and Xiang Wang for their discussions and suggestions.
+
+# References
+
+[1] KLING AI. KLING AI, https://klingai.com/, 2025.6,7,13  
+[2] Runway AI. Stable Diffusion v1.5 Model Card, https://huggingface.co/runwayml/stable-diffusion-v1-5, 2022.2  
+[3] Runway AI. Stable Diffusion Inpainting Model Card, https://huggingface.co/runwayml/stable-diffusion-inpainting, 2022.2  
+[4] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning To Follow Image Editing Instructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18392-18402, 2023. 2, 3  
+[5] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. OpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields. IEEE Trans. Pattern Anal. Mach. Intell., 43(1):172-186, 2021. 5  
+[6] Caroline Chan, Frédo Durand, and Phillip Isola. Learning To Generate Line Drawings That Convey Geometry and Semantics. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7915-7925, 2022. 5  
+[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt- $\alpha$ : Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis. arXiv preprint arXiv:2310.00426, 2023. 2  
+[8] Qihua Chen, Yue Ma, Hongfa Wang, Junkun Yuan, Wenzhe Zhao, Qi Tian, Hongmei Wang, Shaobo Min, Qifeng Chen, and Wei Liu. Follow-Your-Canvas: Higher-Resolution Video Outpainting with Extensive Content Generation. In Assoc. Adv. Artif. Intell., 2025. 6, 13  
+[9] Shoufa Chen, Chongjian Ge, Yuqi Zhang, Yida Zhang, Fengda Zhu, Hao Yang, Hongxiang Hao, Hui Wu, Zhichao Lai, Yifei Hu, Ting-Che Lin, Shilong Zhang, Fu Li, Chuan Li, Xing Wang, Yanghua Peng, Peize Sun, Ping Luo, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Goku: Flow Based Video Generative Foundation Models. arXiv preprint arXiv:2502.04896, 2025. 2  
+[10] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Ji-ashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-A-Video: Controllable Text-to-Video Diffusion Models with Motion Prior and Reward Feedback Learning. arXiv preprint arXiv:2305.13840, 2023. 6, 13  
+[11] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. AnyDoor: Zero-shot Object-level Image Customization. arXiv preprint arXiv:2307.09481, 2023. 2
+
+[12] Xi Chen, Zhifei Zhang, He Zhang, Yuqian Zhou, Soo Ye Kim, Qing Liu, Yijun Li, Jianming Zhang, Nanxuan Zhao, Yilin Wang, Hui Ding, Zhe Lin, and Hengshuang Zhao. UniReal: Universal Image Generation and Editing via Learning Real-world Dynamics. arXiv preprint arXiv:2412.07774, 2024. 2, 3  
+[13] Alibaba Cloud. Tongyi Wanxiang, https://tongyi. aliyun.com/wanxiang, 2023. 2  
+[14] Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: Optical FLOW-guided ATTENtion for consistent text-to-video editing. In Int. Conf. Learn. Represent., 2024. 6, 7, 13  
+[15] Lunhao Duan, Shanshan Zhao, Wenjun Yan, Yinglun Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, Mingming Gong, and Gui-Song Xia. UNIC-Adapter: Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation. arXiv preprint arXiv:2412.18928, 2024.3  
+[16] 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, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. In Int. Conf. Mach. Learn., 2024. 2  
+[17] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierarchical Masked 3D Diffusion Model for Video Outpainting. In ACM Int. Conf. Multimedia, pages 7890-7900, 2023. 6, 13  
+[18] FLUX. FLUX, https://blackforestlabs.ai/, 2024.2  
+[19] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. SEED-Data-Edit Technical Report: A Hybrid Dataset for Instructional Image Editing. arXiv preprint arXiv:2405.04007, 2024. 2  
+[20] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Pengfei Wan, Di Zhang, Yufan Liu, Weiming Hu, Zhengjun Zha, Haibin Huang, and Chongyang Ma. 12V-Adapter: A General Image-to-Video Adapter for Diffusion Models. In ACM SIGGRAPH, pages 1-12, 2024. 2  
+[21] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, Peng Zhang, and Qian He. PuLID: Pure and Lightning ID Customization via Contrastive Alignment. In Adv. Neural Inform. Process. Syst., 2024. 2  
+[22] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. LTX-Video: Realtime Video Latent Diffusion. arXiv preprint arXiv:2501.00103, 2025. 2, 6, 13  
+[23] Zhen Han, Zeyinzi Jiang, Yulin Pan, Jingfeng Zhang, Chaojie Mao, Chenwei Xie, Yu Liu, and Jingren Zhou. ACE: All-round Creator and Editor Following Instructions via Diffusion Transformer. In Int. Conf. Learn. Represent., 2025. 2, 3
+
+[24] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. In Adv. Neural Inform. Process. Syst., 2021. 2  
+[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In Adv. Neural Inform. Process. Syst. Curran Associates, Inc., 2020. 2  
+[26] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and Controllable Image Synthesis with Composable Conditions. In Int. Conf. Mach.Learn., 2023.2  
+[27] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive Benchmark Suite for Video Generative Models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21807-21818, 2024. 5, 7  
+[28] Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, Yaohui Wang, Xinyuan Chen, YingCong Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench++: Comprehensive and Versatile Benchmark Suite for Video Generative Models. arXiv preprint arXiv:2411.13503, 2024. 5  
+[29] Zeyinzi Jiang, Chaojie Mao, Ziyuan Huang, Ao Ma, Yiliang Lv, Yujun Shen, Deli Zhao, and Jingren Zhou. Res-Tuning: A Flexible and Efficient Tuning Paradigm via Unbinding Tuner from Backbone. In Adv. Neural Inform. Process. Syst., 2023. 5, 8  
+[30] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. SCEdit: Efficient and Controllable Image Diffusion Generation via Skip Connection Editing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8995-9004, 2024. 2  
+[31] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators. In Int. Conf. Comput. Vis., pages 15954–15964, 2023. 6, 13  
+[32] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyuan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Dax Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. HunyuanVideo: A Systematic Framework For Large Video Generative Models. arXiv preprint arXiv:2412.03603, 2024. 2  
+[33] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding. arXiv preprint arXiv:2405.08748, 2024. 2
+
+[34] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. MagicEdit: High-Fidelity and Temporally Coherent Video Editing. arXiv preprint arXiv:2308.14749, 2023. 3  
+[35] Lijie Liu, Tianxiang Ma, Bingchuan Li, Zhuowei Chen, Jiawei Liu, Qian He, and Xinglong Wu. Phantom: Subject-consistent video generation via cross-modal alignment. arXiv preprint arXiv:2502.11079, 2025. 2, 3  
+[36] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. arXiv preprint arXiv:2303.05499, 2023. 5  
+[37] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-P2P: Video Editing with Cross-attention Control. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8599–8608, 2024. 3  
+[38] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept Neurons in Diffusion Models for Customized Generation. In Int. Conf. Mach. Learn., 2023. 2  
+[39] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones 2: Customizable Image Synthesis with Multiple Subjects. In Adv. Neural Inform. Process. Syst., 2023. 2  
+[40] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Ying Shan, Xiu Li, and Qifeng Chen. Follow Your Pose: Pose-Guided Text-to-Video Generation using Pose-Free Videos. In Assoc. Adv. Artif. Intell., 2024. 6, 13  
+[41] Chaojie Mao, Jingfeng Zhang, Yulin Pan, Zeyinzi Jiang, Zhen Han, Yu Liu, and Jingren Zhou. ACE++: Instruction-Based Image Creation and Editing via Context-Aware Content Filling. arXiv preprint arXiv:2501.02487, 2025. 2, 3  
+[42] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. In Int. Conf. Learn. Represent., 2021. 2  
+[43] Midjourney. Midjourney, https://www.midjourney.com, 2023.2  
+[44] MiniMax. Hailuo AI Video, https://hailuoai.com/video, 2024.  
+[45] OpenAI. DALL·E 3, https://openai.com/dall-e-3, 2023.2  
+[46] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold. In ACM SIGGRAPH, 2023. 2  
+[47] Yulin Pan, Chaojie Mao, Zeyinzi Jiang, Zhen Han, and Jingfeng Zhang. Locate, Assign, Refine: Taming Customized Image Inpainting with Text-Subject Guidance. arXiv preprint arXiv:2403.19534, 2024. 2  
+[48] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers. In Int. Conf. Comput. Vis., pages 4195-4305, 2023. 2  
+[49] PiKa. PiKa, https://pika.art/, 2025.6, 7, 13  
+[50] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming
+
+Xiong, Silvio Savarese, Stefano Ermon, Yun Fu, and Ran Xu. UniControl: A Unified Diffusion Model for Controllable Visual Generation In the Wild. In Adv. Neural Inform. Process. Syst., 2023. 3  
+[51] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer. IEEE Trans. Pattern Anal. Mach. Intell., pages 1623-1637, 2022. 5  
+[52] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. SAM 2: Segment Anything in Images and Videos. In Int. Conf. Learn. Represent., 2025. 5  
+[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684-10695, 2022. 2  
+[54] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. Med. Image Comput. Computer-Assisted Interv., 2015. 2  
+[55] Runway. Gen-3, https://app.runwayml.com/video-tools, 2025.2  
+[56] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models. In Int. Conf. Learn. Represent., 2021. 2  
+[57] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. In Int. Conf. Learn. Represent., 2021. 2  
+[58] StabilityAI. Stable Diffusion v2-1 Model Card, https: //huggingface.co/stabilityai/stable-diffusion-2-1,2022.2  
+[59] StabilityAI. Stable Diffusion XL Model Card, https: //huggingface.co/stabilityai/stable-diffusion-x1-base-1.0,2022.2  
+[60] StabilityAI. CosXL Model Card, https://huggingface.co/stabilityai/cosxl, 2024.3  
+[61] Ya Sheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. ImageBrush: Learning Visual In-Context Instructions for Exemplar-Based Image Manipulation. In Adv. Neural Inform. Process. Syst., 2023. 3  
+[62] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-Robust Large Mask Inpainting With Fourier Convolutions. In IEEE Winter Conf. Appl. Comput. Vis., pages 2149-2159, 2022. 5  
+[63] Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. OminiControl: Minimal and Universal Control for Diffusion Transformer. arXiv preprint arXiv:2411.15098, 2024. 2, 3  
+[64] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 2, 6, 13
+
+[65] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. In Eur. Conf. Comput. Vis., pages 402-419, 2020. 5  
+[66] Vidu. Vidu, https://www.vidu.cn/, 2025. 6, 7, 13  
+[67] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. InstantID: Zero-shot Identity-Preserving Generation in Seconds. arXiv preprint arXiv:2401.07519, 2024. 2  
+[68] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional Video Synthesis with Motion Controllability. In Adv. Neural Inform. Process. Syst., 2023. 2, 6, 13  
+[69] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A Unified and Flexible Motion Controller for Video Generation. In ACM SIGGRAPH, pages 1-11, 2024. 3  
+[70] Yujie Wei, Shiwei Zhang, Hangjie Yuan, Xiang Wang, Haonan Qiu, Rui Zhao, Yutong Feng, Feng Liu, Zhizhong Huang, Jiaxin Ye, Yingya Zhang, and Hongming Shan. DreamVideo-2: Zero-Shot Subject-Driven Video Customization with Precise Motion Control. arXiv preprint arXiv:2410.13830, 2024. 2  
+[71] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. OmniGen: Unified Image Generation. arXiv preprint arXiv:2409.11340, 2024. 2, 3  
+[72] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective Whole-body Pose Estimation with Two-stages Distillation. In Int. Conf. Comput. Vis., pages 4210-4220, 2023. 5  
+[73] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer. In Int. Conf. Learn. Represent., 2025. 2, 6, 13  
+[74] Shenghai Yuan, Jinfa Huang, Xianyi He, Yunyuan Ge, Yu-jun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. Identity-Preserving Text-to-Video Generation by Frequency Decomposition. In IEEE Conf. Comput. Vis. Pattern Recog., 2025. 2  
+[75] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. MagicBrush: A Manually Annotated Dataset for Instruction-Guided Image Editing. In Adv. Neural Inform. Process. Syst., 2023. 2  
+[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In Int. Conf. Comput. Vis., pages 3836–3847, 2023. 2  
+[77] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2VGen-XL: High-Quality Image-to-Video Synthesis via Cascaded Diffusion Models. arXiv preprint arXiv:2311.04145, 2023. 2, 6, 13  
+[78] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo,
+
+Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize Anything: A Strong Image Tagging Model. arXiv preprint arXiv:2306.03514, 2023. 5  
+[79] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. ControlVideo: Training-free Controllable Text-to-Video Generation. In Int. Conf. Learn. Represent., 2024. 6, 13  
+[80] Yuechen Zhang, Yaoyang Liu, Bin Xia, Bohao Peng, Zexin Yan, Eric Lo, and Jiaya Jia. Magic Mirror: ID-Preserved Video Generation in Video Diffusion Transformers. arXiv preprint arXiv:2411.13503, 2025. 3  
+[81] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based Fine-Grained Image Editing at Scale. arXiv preprint arXiv:2407.05282v1, 2024. 2  
+[82] Shangchen Zhou, Chongyi Li, Kelvin C. K. Chan, and Chen Change Loy. ProPainter: Improving Propagation and Transformer for Video Inpainting. In Int. Conf. Comput. Vis., pages 10477-10486, 2023. 2, 6, 13

2025/VACE_ All-in-One Video Creation and Editing/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d0088dcd57ebce66515463fd6ba5126521846337b34eb2a20fe627ffa904654
+size 824666

2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/7812feed-641c-4918-a582-f934ca29c2a8_content_list.json

@@ -0,0 +1,1493 @@
+[
+    {
+        "type": "text",
+        "text": "VAFlow: Video-to-Audio Generation with Cross-Modality Flow Matching",
+        "text_level": 1,
+        "bbox": [
+            127,
+            130,
+            870,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Xihua Wang $^{1}$  Xin Cheng $^{1}$  Yuyue Wang $^{1}$  Ruihua Song $^{1*}$  Yunfeng Wang $^{2}$ $^{1}$ Gaoling School of Artificial Intelligence, Renmin University of China  \nxihuaw@ruc.edu.cn, songruihua.bloon@outlook.com",
+        "bbox": [
+            112,
+            178,
+            887,
+            233
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            248,
+            268,
+            326,
+            284
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Video-to-audio (V2A) generation aims to synthesize temporally aligned, realistic sounds for silent videos, a critical capability for immersive multimedia applications. Current V2A methods, predominantly based on diffusion or flow models, rely on suboptimal noise-to-audio paradigms that entangle cross-modal mappings with stochastic priors, resulting in inefficient training and convoluted transport paths. We propose VAFlow, a novel flow-based framework that directly models the video-to-audio transformation, eliminating reliance on noise priors. To address modality discrepancies, we employ an alignment variational autoencoder that compresses heterogeneous video features into audio-aligned latent spaces while preserving spatiotemporal semantics. By retaining cross-attention mechanisms between video features and flow blocks, our architecture enables classifier-free guidance within video source-driven generation. Without external data or complex training tricks, VAFlow achieves state-of-the-art performance on VGGSound benchmark, surpassing even textaugmented models in audio fidelity, diversity, and distribution alignment. This work establishes a new paradigm for V2A generation with a direct and effective video-to-audio transformation via flow matching.",
+        "bbox": [
+            86,
+            301,
+            485,
+            650
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            91,
+            681,
+            220,
+            698
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Video-to-audio (V2A) generation plays a vital role in multimedia content generation, including Foley for films and AI-generated silent videos. In recent years, this task has attracted increasing attention within the generative community. Early V2A approaches [15, 24, 26, 29, 34] adopted language modeling strategies by discretizing audio into tokens, and used parallel mask-prediction models (BERT-like) or token-by-token autoregressive models (GPT-like) borrowed from NLP community as shown in Figure 1. However, these methods, reliant on discrete tokens, incur inherent information losses by the discretization process.",
+        "bbox": [
+            89,
+            707,
+            483,
+            875
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/00b726f5ccb3264f0217ff01f4838780de47120879ff2588490bb01d049831e2.jpg",
+        "image_caption": [
+            "Figure 1. Four different V2A paradigms, along with state-of-the-art performance under each. The first row shows discrete token-based approaches (mask-prediction and autoregressive based generation), which are limited by information loss during tokenization. The second row presents methods generating continuous data. Previous diffusion and flow methods (bottom-left) model the noise-to-audio mapping conditioning on video. We find this approach suboptimal and demonstrate that directly modeling the video-to-audio transformation (bottom-right) yields better results."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            516,
+            266,
+            906,
+            473
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent advancements in Diffusion and Flow Matching (FM) models in image [4, 10, 33] and video [3, 31] generation have spurred similar approaches for V2A [9, 25, 41, 43, 45]. Generally, these methods model V2A as a conditional generation task, transforming Gaussian noise conditioned on video into audio, as shown in Figure 1 (bottom-left). The generation process typically involves two steps: (1) sampling a random noise from Gaussian prior and (2) using this noise and video condition as input to iteratively denoise the noise into audio. This approach avoids discrete audio tokens, directly fitting the continuous distribution of audio data, and currently holds state-of-the-art results.",
+        "bbox": [
+            511,
+            641,
+            908,
+            823
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, existing diffusion/FM-based V2A methods focus solely on the second step, i.e., denoising Gaussian noise into audio, assuming the first step is trivial. Even state-of-the-art methods [40, 43] begin with Gaussian noise, but we find that this approach is not optimal. Unlike text-to",
+        "bbox": [
+            511,
+            824,
+            908,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV 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": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author.",
+        "bbox": [
+            109,
+            887,
+            236,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "11777",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "image or text-to-video tasks, where textual conditions are abstract and highly defined by human, the video conditions in V2A are more complex and require both spatial and temporal understanding. The generation goal also demands a high degree of alignment between the generated audio and the video conditions, not just semantically, but also temporally. There are two inherent difficulties: (1) different conditions on the same noise require the model to generate different audio outputs, demanding advanced condition processing [8, 41, 46], and (2) noise variations under the same condition lead to unstable generation results, prompting research into methods for noise control [44], such as finding a \"golden noise\" for consistent results [47].",
+        "bbox": [
+            89,
+            90,
+            480,
+            287
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In contrast, we propose a new perspective by shifting focus to the first step of the generation process. We question: If there are better priors than Gaussian noise? Given that many diffusion and FM models dedicate significant effort to integrating video conditions into the denoising process for better generation performance, we further question: If V2A can work better when directly transforming the video distribution to audio rather than from noise? Fortunately, both answers are affirmative.",
+        "bbox": [
+            89,
+            290,
+            480,
+            425
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We introduce VAFlow, a novel framework for V2A generation, which directly denoises from the video space to the audio space. To implement this video-to-audio paradigm, we adopt the FM framework, as it theoretically supports any source distribution. There are three technical challenges for this implementation: (1) aligning source and target distributions: video and audio feature spaces typically have different time and spatial resolutions, yet FM requires the source and target distributions to be aligned (the same shape). (2) uncertainty of the feature type: identifying the optimal video representations/feature types as the source distribution. (3) conditioning mechanism: exploring additional video information integration ways beyond encoding it in the source distribution. In VAFlow, we first employ an alignment VAE to adjust video features to match the audio latent space. It then explores three different video feature types (semantic CLIP [32], spatiotemporal CAVP [25], reconstruction-focused VidTok [38]) to systematically investigate source video information. Experiments show that using video features both as a source distribution and as conditioning information yields the best results for VAFlow.",
+        "bbox": [
+            89,
+            429,
+            482,
+            746
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Moreover, the recent success of large language models (LLMs) [13, 28] has highlighted the effectiveness of scaling laws, and various generative works consequently explore similar scaling properties in each domain [30, 42]. However, existing V2A models have not demonstrated clear scaling trends, where increasing model size does not consistently achieve improvement [8, 43]. In contrast, our experiments show that VAFlow exhibits clear scaling property, with model size increases leading to consistent improvements in generation quality, further confirming its potential.",
+        "bbox": [
+            89,
+            750,
+            482,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Without elaborate training tricks or external data (e.g., text-audio pairs), VAFlow achieves state-of-the-art results on standard benchmarks (VGGSound), outperforming even text-augmented models in generation fidelity, diversity, and cross-modality alignment with input video. Playable audi-los generated with VAFlow can be accessed at the demo page: https://vaflow.github.io/demo.",
+        "bbox": [
+            511,
+            90,
+            903,
+            196
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            511,
+            212,
+            653,
+            228
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Flow Matching and Diffusion Models",
+        "text_level": 1,
+        "bbox": [
+            511,
+            238,
+            834,
+            253
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Diffusion models [17, 27, 35] reconstruct data distributions by iteratively denoising Gaussian-sampled latents, governmented by various designed schedulers achieving state-of-the-art performance in multimodal generation tasks. Unlike diffusions, flow-based models [1, 20, 22, 23] establish deterministic mappings between source (e.g., Gaussian) and target distributions through invertible neural transformations, with recent flow matching optimizing vector fields through optimal transport theory, enabling direct learning of straight transport paths between source and target distributions, thereby reducing training instability and inference steps. While both paradigms typically rely on Gaussian priors as their source distribution, existing video-to-audio systems on both paradigms model noise-to-audio generation conditioned on video features. In contrast, our approach directly aligns video latent spaces with target audio distributions via linear transformations under a flow matching framework, bypassing iterative noise modeling, preserving efficiency, and ensuring precise frame-to-frame alignment.",
+        "bbox": [
+            511,
+            260,
+            903,
+            547
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Video-to-Audio Generation",
+        "text_level": 1,
+        "bbox": [
+            511,
+            559,
+            758,
+            574
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Framework. Recent advances in generative models have driven significant progress in video-to-audio (V2A) synthesis. Current V2A methods fall into three main paradigms: (1) Autoregressive (AR) models [15, 26, 34] discretize audio into token (codec) sequences and utilize video features as prefixes to model unified video-audio sequences. While straightforward, their token-by-token generation process results in quadratic inference time scaling, limiting efficiency for long audio sequences. (2) Mask-based generation paradigms [24, 29, 37] also employ discrete tokens but decode audio in parallel by predicting masked tokens. Video features serve as the condition for mask-to-audio prediction. However, both AR and mask-based methods suffer from an inherent upper bound of fidelity due to lossy compression from audio discretization. (3) Flow/diffusion-based methods [7-9, 25, 40, 41, 43, 45, 46] leverage continuous representations and currently achieve state-of-the-art V2A performance [8, 40, 43]. These techniques iteratively transform noise into target audio distributions while conditioning on video features to steer the denoising trajectory.",
+        "bbox": [
+            511,
+            582,
+            903,
+            883
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Video Conditional Modeling. Unlike typical conditioned",
+        "bbox": [
+            511,
+            885,
+            903,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "11778",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "generation tasks like text-to-image [33], V2A involves fully aligned sequential correspondences between video and audio. Existing methods address this through various strategies: some AR-based methods investigate various attention mechanisms and masking schemes for temporally sequential alignment [26]; some mask-based approaches integrate hybrid ways to fuse video features into the mask-prediction process to synchronize audio generation [29]; and some flow/diffusion-based methods augment their pipelines with auxiliary modules (e.g., loudness or audio layout predictors [41, 46]) to enforce temporal consistency. In contrast, our approach builds on flow-based techniques by discarding the conventional noise-to-audio paradigm and complex video-conditioning designs. Instead, we propose a more intuitive method that directly transforms the video temporal space to the audio temporal space via a flow transport.",
+        "bbox": [
+            89,
+            90,
+            483,
+            333
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method:",
+        "text_level": 1,
+        "bbox": [
+            89,
+            345,
+            187,
+            362
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Preliminary: Flow Matching",
+        "text_level": 1,
+        "bbox": [
+            89,
+            369,
+            349,
+            386
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Let  $x_0 \\sim p(x_0)$  and  $x_1 \\sim p(x_1)$  denote samples from source and target distributions, respectively, with  $c$  representing an optional conditioning signal. Flow Matching (FM) [20, 23] establishes an inter-distribution transport process governed by the ordinary differential equation (ODE):",
+        "bbox": [
+            89,
+            392,
+            483,
+            470
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nd x (t) = v _ {\\theta} \\left(x _ {t}, t, c\\right) d t, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            209,
+            479,
+            482,
+            497
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $t \\in [0,1]$  denotes the continuous time step, and  $v_{\\theta}$  represents a neural network-parameterized velocity field. The condition  $c$  is incorporated for conditional flow implementations. Optimal transport flows enforce velocity alignment with the direction  $(x_{1} - x_{0})$  of the linear path pointing from  $x_{0}$  to  $x_{1}$  through regression:",
+        "bbox": [
+            89,
+            506,
+            483,
+            599
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} (\\theta) = \\mathbb {E} _ {t, x _ {0}, x _ {1}, c} \\| (x _ {1} - x _ {0}) - v _ {\\theta} (x _ {t}, t, c) \\| ^ {2}, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            133,
+            607,
+            482,
+            626
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $x_{t} = (1 - t)x_{0} + tx_{1}$  constitutes the linear interpolation between  $x_{0}$  and  $x_{1}$ . The learned velocity field enables target generation by solving the ODE in Equation 1.",
+        "bbox": [
+            89,
+            636,
+            482,
+            680
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Previous video-to-audio flow approaches [8, 43] (and related diffusion methods [25, 41]) use a Gaussian prior  $p(x_0) = \\mathcal{N}(0,1)$  for the source distribution, with audio as  $p(x_{1})$  and video as conditioning input  $c$ . During training, random noise  $x_0$  is paired with ground-truth audio  $x_{1}$ , prompting  $v_{\\theta}$  to model the paths from noise to audio under visual conditions.",
+        "bbox": [
+            89,
+            681,
+            482,
+            785
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This work challenges the Gaussian prior by proposing a direct video-to-audio transport. Theoretically, FM permits arbitrary source distributions, motivating our formulation:",
+        "bbox": [
+            89,
+            787,
+            483,
+            832
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nd x (t) = v _ {\\theta} \\left(x _ {t}, t\\right) d t, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            215,
+            843,
+            482,
+            861
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $x_{t} = (1 - t)x_{v} + tx_{a}$  linearly interpolates between paired video and audio samples  $x_{v},x_{a}\\sim p(v,a)$",
+        "bbox": [
+            89,
+            869,
+            483,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. VAFlow Architecture",
+        "text_level": 1,
+        "bbox": [
+            513,
+            90,
+            714,
+            104
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Model Overview. As illustrated in Figure 2, the input video features  $x_{v}$  are extracted via a visual encoder, while the target audio waveform can be obtained from latent representations  $x_{a}$  through WaveVAE. VAFlow attempts to establish FM between the video feature space  $x_{v}$  and audio latent space  $x_{a}$ , enabling direct audio generation from video features via ODE solving instead of Gaussian noise initialization. However, since FM requires consistent dimensional shapes between source and target distributions (see Equation 3), the inherent spatiotemporal resolution mismatch between  $x_{v}$  and  $x_{a}$  necessitates shape alignment. To address this, we employ a 1D VAE (namely alignment VAE) that compresses interpolated  $x_{v}$  into aligned features  $x_{v}^{\\mathrm{AE}}$  matching  $x_{a}$ 's dimensions.",
+        "bbox": [
+            511,
+            114,
+            906,
+            327
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "A diffusion transformer (DiT) subsequently works as a velocity field estimator to capture the transport dynamics between  $x_{v}^{\\mathrm{AE}}$  and  $x_{a}$ . This enables continuous trajectory sampling from  $x_{v}^{\\mathrm{AE}}$  to  $x_{a}$  through ODE integration. The sampled latent  $\\hat{x}_{a}$  is finally decoded by WaveVAE to produce the output audio. The following contents detail: the visual/audio representations  $(x_{v}, x_{v}^{\\mathrm{AE}}, x_{a})$ , and the velocity field estimator in VAFlow.",
+        "bbox": [
+            511,
+            330,
+            905,
+            450
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Video and Audio Representations. For an input video-audio pair  $(v, a)$ , the audio waveform  $a$  is compressed by a pretrained 1D WaveVAE's [11] encoder into  $x_{a} \\in \\mathbb{R}^{T_{a} \\times D_{a}}$ . This latent  $x_{a}$  serves as the target for flow generation and is later decoded to reconstruct the waveform.",
+        "bbox": [
+            511,
+            453,
+            906,
+            527
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For video representation, a pretrained visual encoder encodes  $v$  into  $x_v \\in \\mathbb{R}^{T_v \\times D_v}$  (e.g., CLIP [32], CAVP [25], or VidTok [38] as encoder). Given the discrepancies (e.g.,  $x_a \\in \\mathbb{R}^{215 \\times 64}$  versus  $x_v \\in \\mathbb{R}^{100 \\times 768}$  for a 10-second clip), we propose an alignment VAE  $\\epsilon_{\\phi}$  to further compress and align video latents with audio latents' space while preserving essential visual information. Specifically, we utilize a 1D convolutional encoder-decoder as  $\\epsilon_{\\phi}$  to interpolate and project  $x_v$  into a variational space  $x_v^{\\mathrm{AE}} \\in \\mathbb{R}^{T_a \\times D_a}$ , matching the shape of  $x_a$ . The encoder of  $\\epsilon_{\\phi}$  predicts the parameters  $\\mu_{x_v^{\\mathrm{AE}}}, \\sigma_{x_v^{\\mathrm{AE}}} = E_{\\epsilon_{\\phi}}(x_v)$  for a Gaussian distribution, from which  $x_v^{\\mathrm{AE}} \\sim \\mathcal{N}(\\mu_{x_v^{\\mathrm{AE}}}, \\sigma_{x_v^{\\mathrm{AE}}})$  is sampled. The decoder reconstructs  $x_v$  as  $\\hat{x}_v = D_{\\epsilon_{\\phi}}(x_v^{\\mathrm{AE}})$ . This process ensures dimensional alignment with audio latents while retaining critical spatiotemporal features of the original video representation.",
+        "bbox": [
+            511,
+            532,
+            906,
+            777
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Velocity Estimator. Built upon the diffusion transformer (DiT) architecture [11], the velocity estimator processes two key inputs: time embeddings of  $t$  and the interpolated state  $x_{t} = (1 - t)x_{v}^{\\mathrm{AE}} + tx_{a}$ , following the Equation 3. The time embedding is concatenated with the audio sequence to form an input tensor  $\\in \\mathbb{R}^{(T_a + 1)\\times D_a}$ , which is processed through stacked transformer blocks. The first  $T_{a}$  dimensions of the output sequence constitute the pre",
+        "bbox": [
+            511,
+            779,
+            908,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "11779",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/ff1ff64fb37921e45705859648878aee89608f3208a30432c279d825228be95d.jpg",
+        "image_caption": [
+            "Figure 2. VAFlow Architecture. The figure illustrates VAFow in the inference manner. A visual encoder initially extracts temporal features  $x_{v} \\in \\mathbb{R}^{T_{v} \\times D_{v}}$  from variable-frame-rate video inputs. Then, a 1D alignment VAE (Variational Autoencoder)  $\\epsilon_{\\phi}$  compresses interpolated  $x_{v}$  into audio-aligned latent features  $x_{v}^{\\mathrm{AE}}$  with matched dimensions  $(T_{v}^{\\mathrm{AE}} \\times D_{v}^{\\mathrm{AE}} = T_{a} \\times D_{a})$ . Next, a diffusion transformer  $(v_{\\theta})$  predicts the velocity field for the transport  $(x_{a} - x_{v}^{\\mathrm{AE}})$ , and by solving the corresponding ODE (see Equation 3), an estimated  $\\hat{x}_{a}$  is obtained. Finally, the WaveVAE decoder converts  $\\hat{x}_{a}$  into the waveform representing the predicted audio. Note that during diffusion transformer prediction,  $x_{v}$  can optionally be incorporated via cross-attention, detail discussion is conducted in the experiment Section 4.5."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            89,
+            898,
+            224
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "dicted velocity  $v_{\\theta}(x_t, t)$ , with 1D Rotary Position Embeddings (RoPE [36]) to preserve positional awareness in attention layers. Notably, while the interpolated state  $x_t$  implicitly encodes visual information through  $x_v^{\\mathrm{AE}}$ , this signal diminishes as  $t \\to 1$ . To address this limitation, we retain cross-attention layers in the transformer blocks that persistently integrate the original video features  $x_v$ . Further studies (Section 4.5) demonstrate that this architectural choice is critical to enhance visual perception throughout the flow trajectory and improve audio generation quality.",
+        "bbox": [
+            88,
+            349,
+            480,
+            501
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. VAFlow Training and Inference",
+        "text_level": 1,
+        "bbox": [
+            89,
+            510,
+            372,
+            526
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We conduct three-stage training: (1) pretrain the alignment VAE, (2) train the VAFlow estimator with frozen alignment VAE, and (3) jointly fine-tune both alignment VAE and estimator. We detail training and inference as follows:",
+        "bbox": [
+            89,
+            532,
+            482,
+            593
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Alignment VAE Training. Given encoded visual features  $x_{v} \\in \\mathbb{R}^{T_{v} \\times D_{v}}$ , the alignment VAE encoder predicts distribution parameters  $(\\mu_{x_{v}^{\\mathrm{AE}}}, \\sigma_{x_{v}^{\\mathrm{AE}}})$ , while the decoder reconstructs  $\\hat{x}_{v}$  from a sampled  $x_{v}^{\\mathrm{AE}} \\sim \\mathcal{N}(\\mu_{x_{v}^{\\mathrm{AE}}}, \\sigma_{x_{v}^{\\mathrm{AE}}})$ . The model optimizes three objectives to preserve visual semantics while regularizing  $x_{v}^{\\mathrm{AE}}$  within  $\\mathbb{R}^{T_{a} \\times D_{a}}$  Gaussian space:",
+        "bbox": [
+            89,
+            593,
+            482,
+            686
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "- KL Regularization: enforces Gaussian latent priors via:",
+        "bbox": [
+            89,
+            688,
+            470,
+            702
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K L}} = D _ {\\mathrm {K L}} \\left(\\mathcal {N} \\left(\\mu_ {x _ {0} ^ {\\mathrm {A E}}}, \\sigma_ {x _ {0} ^ {\\mathrm {A E}}}\\right) \\| \\mathcal {N} (0, I)\\right). \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            148,
+            714,
+            482,
+            732
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "- Reconstruction Loss: ensures feature-level consistency:",
+        "bbox": [
+            89,
+            742,
+            470,
+            757
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {r e c}} = \\left\\| x _ {v} - \\hat {x} _ {v} \\right\\| _ {2} ^ {2}. \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            223,
+            768,
+            482,
+            785
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "- Contrastive Alignment: strengthens feature preservation:",
+        "bbox": [
+            89,
+            797,
+            478,
+            811
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {c o n}} = - \\sum_ {s \\in \\mathcal {S}} \\log \\frac {\\exp \\left(f _ {s} \\left(\\hat {x} _ {v}\\right) \\cdot f _ {s} \\left(x _ {v}\\right) / \\tau\\right)}{\\sum_ {x ^ {\\prime} \\in \\mathcal {B}} \\exp \\left(f _ {s} \\left(\\hat {x} _ {v}\\right) \\cdot f _ {s} \\left(x ^ {\\prime}\\right) / \\tau\\right)}, \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            114,
+            821,
+            482,
+            858
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $S = \\{\\mathrm{frame},\\mathrm{clip},\\mathrm{global}\\}$  defines temporal granularities and  $\\tau = 0.7$  serves as a temperature parameter [6].",
+        "bbox": [
+            102,
+            869,
+            482,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For scale  $s$ ,  $f_{s}(\\cdot)$  extracts features via three ways: frame-level: single-frame sampling + mean pooling; clip-level:  $K = 8$ -adjacent-frame sampling + pooling; global-level: full temporal mean pooling at  $T_{v}$  dimension of feature  $x$ . Negative samples  $x'$  are drawn from training batch  $\\mathcal{B}$ .",
+        "bbox": [
+            524,
+            349,
+            906,
+            426
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The composite objective integrates these terms:",
+        "bbox": [
+            513,
+            428,
+            826,
+            443
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {V A E}} = \\lambda_ {\\mathrm {K L}} \\mathcal {L} _ {\\mathrm {K L}} + \\lambda_ {\\mathrm {r e c}} \\mathcal {L} _ {\\mathrm {r e c}} + \\lambda_ {\\mathrm {c o n}} \\mathcal {L} _ {\\mathrm {c o n}}, \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            453,
+            903,
+            468
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "with  $\\lambda_{\\mathrm{KL}} = 8\\times 10^{-4}$ $\\lambda_{\\mathrm{rec}} = 1.0$  , and  $\\lambda_{\\mathrm{con}} = 1.0$",
+        "bbox": [
+            511,
+            478,
+            839,
+            493
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Velocity Estimator Training Given encoded features  $x_{v}$ ,  $x_{v}^{\\mathrm{AE}}$ , and  $x_{a}$ , we train the velocity estimator  $v_{\\theta}$  to predict the linear transport direction  $x_{a} - x_{v}^{\\mathrm{AE}}$  through:",
+        "bbox": [
+            511,
+            494,
+            905,
+            539
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} (\\theta) = \\mathbb {E} _ {t \\sim \\mathcal {U} (0, 1)} \\left\\| \\left(x _ {a} - x _ {v} ^ {\\mathrm {A E}}\\right) - v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) \\right\\| _ {2} ^ {2}, \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            531,
+            547,
+            903,
+            570
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $x_{t} = (1 - t)x_{v}^{\\mathrm{AE}} + tx_{a}$ . For the conditioned version, visual features  $x_{v}$  are incorporated via cross-attention (with a  $10\\%$  probability of zeroing out  $x_{v}$  during training). The vanilla version omits this conditioning, simplifying the input of estimator in Equation 8 to  $v_{\\theta}(x_t,t)$ .",
+        "bbox": [
+            511,
+            578,
+            905,
+            654
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Joint Turning After separate training of the alignment VAE and the velocity estimator, a joint optimization stage is performed. In this stage, the VAE encoder and  $v_{\\theta}$  are co-trained (discarding the VAE decoder), while the WaveVAE and visual feature extractor remain frozen. For a pair  $(v, a)$ , the visual extractor and WaveVAE produce  $x_{v}$  and  $x_{a}$ . The VAE encoder compresses  $x_{v}$  to obtain  $x_{v}^{\\mathrm{AE}} \\sim q_{\\phi}(x_{v}^{\\mathrm{AE}}|x_{v})$  which together with  $x_{a}$  and a random  $t \\sim \\mathcal{U}(0,1)$  forms  $x_{t}$ . The joint objective integrates KL regularization in Equation 4 with velocity prediction in Equation 8:",
+        "bbox": [
+            511,
+            655,
+            906,
+            805
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} (\\theta , \\phi_ {\\text {E n c}}) = \\lambda_ {\\mathrm {K L}} \\mathcal {L} _ {\\mathrm {K L}} + \\mathbb {E} \\| \\left(x _ {a} - x _ {v} ^ {\\mathrm {A E}}\\right) - v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) \\| _ {2} ^ {2}. \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            513,
+            813,
+            906,
+            845
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Inference Upon training, VAFlow synthesizes audio  $a$  from an input video  $v$  through the following pipeline: The visual extractor first encodes  $v$  into a feature  $x_{v}$ , which is",
+        "bbox": [
+            511,
+            854,
+            906,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "11780",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "compressed by the alignment VAE encoder into distribution parameters  $(\\mu_{x_v^{\\mathrm{AE}}},\\sigma_{x_v^{\\mathrm{AE}}})$ . A latent sample  $x_{v}^{\\mathrm{AE}}\\sim \\mathcal{N}(\\mu_{x_v^{\\mathrm{AE}}},\\sigma_{x_v^{\\mathrm{AE}}}^2)$  is drawn as the initial state  $(x_0)$  for solving the ODE in Equation 1 with learned estimator. The predicted  $\\hat{x}_a(x_1)$  is then decoded into waveform via the frozen WaveVAE. During inference, the variational sampling of  $x_{v}^{\\mathrm{AE}}$  introduces stochasticity - even with fixed ODE integration paths, distinct samples yield diverse outputs.",
+        "bbox": [
+            89,
+            90,
+            480,
+            214
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For conditional velocity estimators, classifier-free guidance (CFG) [14] enhances synthesis quality by interpolating conditioned and unconditioned velocity predictions:",
+        "bbox": [
+            89,
+            215,
+            480,
+            260
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nv _ {\\theta} ^ {\\mathrm {C F G}} \\left(x _ {t}, t, x _ {v}\\right) = (1 + \\gamma) v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) - \\gamma v _ {\\theta} \\left(x _ {t}, t, \\emptyset\\right), \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            102,
+            268,
+            480,
+            301
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\gamma$  controls guidance strength and  $\\varnothing$  denotes null (zero) conditioning.",
+        "bbox": [
+            89,
+            311,
+            480,
+            343
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            356,
+            222,
+            372
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            89,
+            380,
+            279,
+            396
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Datasets. We employ two datasets: VGGSound [5] and AudioSet-V2A [25]. For VGGSound, we adhere to the original split, which comprises 186k training samples and 15k testing samples. AudioSet-V2A contains 390k videos that are exclusively used for training. During training, the alignment VAE is trained on a combined set of VGGSound training and AudioSet-V2A to yield more generalized visual feature compression. The subsequent velocity estimator training and joint tuning are conducted solely on the VGGSound training set. At inference, we evaluate on the VGGSound test set as in prior work.",
+        "bbox": [
+            89,
+            402,
+            480,
+            568
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. For the estimator's DiT architecture, we adapted the DiT implementation from StableAudio-Open-1.0 [11] (SDA1.0) by eliminating duration encoding, applying global time encoding solely to time steps, and restricting the cross-attention layer's key/value inputs to video features. For the visual extractor, we use three variants: CLIP [32] (a widely adopted visual semantic extractor in V2A tasks), CAVP [25] (which captures both visual semantic and temporal features), and VidTok [38] (a visual tokenizer for reconstruction). We train three alignment VAEs with comparable parameter scales—one for each visual extractor. In the velocity estimator training stage, the alignment VAE remains frozen; in the joint training stage, both the estimator and the alignment VAE are trained. In all stages, the audio WaveVAE and the visual extractor are kept frozen. Notably, we do not perform any further rectified iterations on the estimator; all VAFlow variants are trained for a single rectified iteration (Rectified-1, RF-1). During sampling, we employ ODE solvers of various orders—first-order Euler, second-order Midpoint, and fifth-order dopri5—with differing sampling steps and classifier-free guidance scales.",
+        "bbox": [
+            89,
+            569,
+            482,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Metrics. We employ both objective and subjective evaluations across three dimensions: generation quality, audiovisual relevance, and audiovisual synchronization. For objective evaluation, generation quality is measured using FD score [21], KL divergence [15], and IS score [21]. In previous work, each metric may be computed with different feature extraction models (e.g., Melception, PasST $^1$ ). Audiovisual relevance is quantified by computing the cosine similarity between the embeddings of the input video and the generated audio using the ImageBind model [12]. For synchronization, we use the synchronization accuracy (Acc.) [25] reflects the alignment between the input video and generated audio.",
+        "bbox": [
+            511,
+            90,
+            903,
+            286
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Baselines. Our method is comprehensively compared against all major generative paradigms, including autoregressive (AR), masked prediction (Mask), diffusion-based, and flow-based approaches, as detailed in Table 1. All baseline models were evaluated using either officially released code and models or directly downloaded audio outputs by their authors, with consistent experimental settings maintained across all evaluations on the same hardware platform. Implementation specifics include: (1) Im2Wav generates audio in 4s clips, truncated to 10s after three clips. (2) Diff-Foley here employs its double-guidance variant (demonstrating superior performance); (3) MMAudio-L here is the largest variant of MMAudio; and (4) Frieren here employs its final no-reflow configuration (achieving optimal Fréchet distance). For comprehensive evaluation, we also include text-augmented V2A models (denoted in gray rows) that leverage external textual data beyond benchmark audiovisual corpora during training or inference.",
+        "bbox": [
+            511,
+            287,
+            903,
+            559
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Diffusion vs. Flow vs. VAFlow Models",
+        "text_level": 1,
+        "bbox": [
+            511,
+            571,
+            841,
+            585
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We examine framework-specific performance by training same VAFlow-CLIP-142M models under three paradigms, maintaining identical configurations: CLIP visual features, architecture parameters, initialization weights, and other hyperparameters. The frameworks include: Diffusion: Standard DDPM [27, 33] with cross-attention video conditioning. Flow: Standard continuous-time flow (Equation 2) using Gaussian prior. VAFlow: Our framework employing aligned VAE-compressed video features as source distribution. All frameworks use their respective optimal inference settings (DDPM: DPMSolver; Flows: dopri5). Joint tuning in VAFlow begins at 100K steps.",
+        "bbox": [
+            511,
+            594,
+            903,
+            773
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Frechet Distance (FD) results in Figure 3 reveals: 1) Flow paradigms converge faster with lower FD than diffusion, demonstrating inherent optimization benefits. 2) VAFlow attains lower final FD than standard flow despite delayed early-stage convergence (<100K steps). This per",
+        "bbox": [
+            511,
+            776,
+            903,
+            851
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>We adopt PasST [19] model over PANN [18] for its improved robustness following the practice introduced in AudioLDM [21] evaluation repository: https://github.com/haoheliu/audioldm_eval.",
+        "bbox": [
+            511,
+            862,
+            903,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "11781",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/ddbe06928bea183e718f5bdb4312d42104ea62742c41fc44089c0837a7187e51.jpg",
+        "image_caption": [
+            "Figure 3. Comparison of different training frameworks. We examine the same model under different training paradigms by controlling training data, steps, and other hyperparameters. Results demonstrate that VAFlow converges to superior optima, validating the effectiveness of its direct video-to-audio transformation design and highlighting this paradigm advantages in V2A generation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            138,
+            90,
+            436,
+            239
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "formance attributes to VAFlow's explicit modeling of deterministic video-audio transport paths, avoiding the ambiguous noise-to-audio mappings in Gaussian-prior approaches. VAFlow learns optimal transport trajectories between video-audio distribution, bypassing error-prone intermediate transformations in Gaussian-noise-to-audio paths.",
+        "bbox": [
+            89,
+            363,
+            483,
+            455
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. VAFlow Scaling Analysis",
+        "text_level": 1,
+        "bbox": [
+            89,
+            465,
+            318,
+            482
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We investigate VAFlow's scaling capabilities by developing four model variants (142M, 264M, 527M, 1.05B parameters) through width/depth scaling of DiT blocks while maintaining fixed training protocols (data, steps, etc.). Evaluation metrics (FD, IS) versus parameter counts are visualized in Figure 4 (bubble sizes denote model scales).",
+        "bbox": [
+            89,
+            488,
+            482,
+            578
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "There are two key findings here: 1) Parameter Efficiency: VAFlow consistently outperforms flow-based baselines (Frieren, MMAudio) at comparable sizes (e.g., 142M vs. 159M, 1.05B vs. 1.03B) across both metrics. 2) Scaling Capability: Previous flow-based studies have struggled to show consistent scalability [8, 43], where larger models fail to achieve better metrics consistently. To the best of our knowledge, VAFlow is the first flow-based V2A framework to demonstrate stable scaling properties. As the parameter size increases, the models achieve better FD and IS scores. Interestingly, larger models tend to improve IS scores more significantly than FD scores, suggesting that as the VAFlow model scales, the quality of generated audio gradually saturates, and the model becomes more capable of deriving diverse audio from different video-sampled latents.",
+        "bbox": [
+            89,
+            579,
+            482,
+            806
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.4. Benchmark Results",
+        "text_level": 1,
+        "bbox": [
+            89,
+            816,
+            277,
+            832
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Quantitative evaluations are presented in Table 1, where VAFlow demonstrates state-of-the-art performance across key audio generation metrics:  $\\mathrm{FD}_{\\text {pas. }}$ ,  $\\mathrm{FD}_{\\text {mel. }}$ ,  $\\mathrm{IS}_{\\text {pas. }}$ ,  $\\mathrm{KL}_{\\text {pas. }}$ , and  $\\mathrm{KL}_{\\text {mel. }}$ . For audio-visual synchronization,",
+        "bbox": [
+            89,
+            839,
+            483,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/f09fb5f06231f54fb4c79ca7f5346440e949c46f53aa2c0415e37e982158f883.jpg",
+        "image_caption": [
+            "Figure 4. Performance of flow-based baselines and VAFlow variants in scale. VAFlow models (142M, 1.05B) consistently outperform comparable-scale baselines (Frieren-159M, MMAudio-1.03B), with stable scaling property indicated by progressive IS gains (diversity) and saturating FD improvements (quality)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            562,
+            90,
+            859,
+            261
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "VAFlow achieves  $96.3\\%$  accuracy (Acc.), comparable to Frieren  $(97.0\\%)$ . In semantic relevance, it attains an IB-VA score of 28.6, second only to TiVA's 30.3. Notably, VAFlow maintains a simple and straightforward model design without relying on complex video conditioning modules [8, 41, 46], carefully designed noise schedules [9], or additional guidance modules [25, 45]. Its performance across various metrics, achieved with a standard Flow Matching training approach and basic ODE solvers, underscores the simplicity and effectiveness of the direct video-to-audio mapping design. When compared to text-enhanced baselines (gray rows), VAFlow maintains advantages in  $\\mathrm{FD}_{\\text{pas.}}$ ,  $\\mathrm{FD}_{\\text{mel.}}$ ,  $\\mathrm{IS}_{\\text{pas.}}$ , and synchronization Acc., though text-augmented models exhibit superior audio semantic metrics (KL distances and IB-VA). This aligns with expectations since text conditioning explicitly constrains semantic space. We note that Seeing&Hearing's elevated IB-VA score (36.7) stems from classifier guidance using the same ImageBind model, which however compromises generation quality and synchronization performance.",
+        "bbox": [
+            511,
+            369,
+            906,
+            672
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We further conduct human evaluation with sampled 50 videos from the test set to compare our method with the flow-based baseline, Frieren. The generated results from both models were randomized, and 10 experts were invited to score the models based on three criteria: sound quality (Quality), audio-visual semantic consistency (Semantic) and audio-visual synchronization (Sync.). The scoring was done by selecting a winner or declaring a tie for each comparison. The results, summarized in Table 2, show that our model outperforms the baseline in all three dimensions.",
+        "bbox": [
+            511,
+            672,
+            908,
+            824
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.5. Ablative Analysis",
+        "text_level": 1,
+        "bbox": [
+            511,
+            832,
+            684,
+            849
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Visual Features. Three visual feature types were experimented with: semantic features (CLIP embeddings extracted per frame at 10 fps), temporally-enhanced semantic",
+        "bbox": [
+            511,
+            854,
+            906,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "11782",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/b5c13a95ac5535f0a7bfdf235fa260cbc27c52e8be3881895d050442f03d78b9.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Paradigm</td><td rowspan=\"2\">Method</td><td rowspan=\"2\">Vis.Feat.</td><td colspan=\"5\">Quality</td><td rowspan=\"2\">Sync. Acc. ↑</td><td rowspan=\"2\">Semantic IB-VA ↑</td></tr><tr><td>\\(FD_{pas.}\\)↓</td><td>\\(FD_{mel.}\\)↓</td><td>\\(IS_{mel.}\\)↑</td><td>\\(KL_{pas.}\\)↓</td><td>\\(KL_{mel.}\\)↓</td></tr><tr><td rowspan=\"2\">AR</td><td>SpecVQGAN [15]</td><td>RGB,Opti.</td><td>342.4</td><td>18.15</td><td>20.5</td><td>3.10</td><td>6.84</td><td>55.9</td><td>14.1</td></tr><tr><td>Im2Wav [34]</td><td>CLIP</td><td>252.2</td><td>13.03</td><td>35.8</td><td>2.26</td><td>5.41</td><td>78.9</td><td>19.6</td></tr><tr><td rowspan=\"2\">Mask</td><td>VATT [24]</td><td>eva-CLIP+</td><td>154.1</td><td>7.07</td><td>64.3</td><td>1.34</td><td>3.46</td><td>84.8</td><td>25.0</td></tr><tr><td>VAB [37]</td><td>eva-CLIP</td><td>193.1</td><td>12.21</td><td>34.4</td><td>2.31</td><td>5.09</td><td>82.4</td><td>25.9</td></tr><tr><td rowspan=\"7\">Diffusion</td><td>MultiFoley [7]</td><td>CAVP</td><td>241.6</td><td>13.00</td><td>67.7</td><td>1.59</td><td>3.34</td><td>85.2</td><td>27.0</td></tr><tr><td>Seeing&amp;Hearing [45]</td><td>ImageBind+</td><td>251.5</td><td>12.6</td><td>30.1</td><td>2.65</td><td>6.37</td><td>54.2</td><td>36.7</td></tr><tr><td>Diff-Foley [25]</td><td>CAVP</td><td>512.1</td><td>11.45</td><td>50.4</td><td>2.91</td><td>6.18</td><td>88.0</td><td>20.7</td></tr><tr><td>LoVA [9]</td><td>CLIP</td><td>149.1</td><td>6.90</td><td>58.7</td><td>2.10</td><td>5.12</td><td>87.0</td><td>26.3</td></tr><tr><td>FoleyCrafter [46]</td><td>CLIP</td><td>134.2</td><td>7.40</td><td>56.6</td><td>2.29</td><td>5.65</td><td>83.5</td><td>27.8</td></tr><tr><td>TiVA [41]</td><td>CLIP</td><td>111.9</td><td>4.31</td><td>67.0</td><td>2.02</td><td>5.21</td><td>87.5</td><td>30.3</td></tr><tr><td>V2A-Mapper [40]</td><td>CLIP</td><td>94.9</td><td>4.21</td><td>64.9</td><td>2.46</td><td>6.06</td><td>78.6</td><td>22.2</td></tr><tr><td rowspan=\"5\">Flow</td><td>MMAudio-L [8]</td><td>CLIP+</td><td>159.2</td><td>9.38</td><td>45.4</td><td>1.95</td><td>3.84</td><td>77.5</td><td>33.9</td></tr><tr><td>Frieren [43]</td><td>CAVP</td><td>259.9</td><td>6.70</td><td>52.4</td><td>2.92</td><td>5.64</td><td>97.0</td><td>23.0</td></tr><tr><td>VAFlow (ours)</td><td>CLIP</td><td>87.7</td><td>3.86</td><td>73.6</td><td>1.91</td><td>4.81</td><td>88.1</td><td>28.6</td></tr><tr><td>VAFlow (ours)</td><td>CAVP</td><td>91.8</td><td>4.73</td><td>64.8</td><td>2.41</td><td>5.96</td><td>96.3</td><td>25.1</td></tr><tr><td>VAFlow (ours)</td><td>VidTok</td><td>128.5</td><td>6.17</td><td>48.1</td><td>3.64</td><td>8.94</td><td>49.9</td><td>12.9</td></tr></table>",
+        "bbox": [
+            94,
+            88,
+            908,
+            380
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/cb0cc149a1b8270bb5b2fdc91aff8972184a97bd26578f2298e92afecff4b222.jpg",
+        "table_caption": [
+            "Table 1. Quantitative video-to-audio results on the VGGSound test set. “Vis. Feat.” indicates the employed visual features, with a “+” denoting the inclusion of auxiliary visual information (e.g., visual captions via LLama-2 [39] or Qwen-VL [2] in VATT and Seeing&Hearing, or SyncFormer [16] features in MMAudio). All results are obtained from official code or released audio outputs. The MultiFoley results here is an 8k subset filtered by ImageBind from the authors. Rows in gray denote methods that incorporate extra text data (e.g., video captions); for fairness, these results are provided for reference only and are not directly compared with standard video-to-audio models. The best score is bolded, and the second-best is underlined. Our three VAFlow variants, differing in visual features, achieve definitive gains in audio quality (FD, IS, KL) while delivering near state-of-the-art performance for synchronization (Acc.) and correlation (IB-VA)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Comparison</td><td>Quality</td><td>Win rate(%) Sync.</td><td>Semantic</td></tr><tr><td>Ours vs. Frieren</td><td>66.00(±4.18)</td><td>60.44(��5.72)</td><td>60.89(±4.48)</td></tr></table>",
+        "bbox": [
+            96,
+            516,
+            477,
+            571
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Human evaluation results, with comparison between our model and the flow-based baseline Frieren across three criteria: sound quality, audio-visual synchronization, and audio-visual semantic consistency. Ours outperforms the baseline in all criteria.",
+        "bbox": [
+            89,
+            582,
+            482,
+            638
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "features (CAVP features via video-audio contrastive learning at 4 fps), reconstruction features (VidTok tokenized latents at 15 fps). Utilizing the same model size and training process for these features, we developed three variants: VAFlow-CLIP, VAFlow-CAVP, and VAFlow-VidTok, with results presented in the last three rows of Table 1. The results indicate that the widely-used CLIP features overall perform the best, yielding superior audio quality and semantic relevance. The temporally enhanced CAVP features, on the other hand, demonstrate the best video-audio synchronization but slightly lag in other metrics compared to VAFlow-CLIP. Models based on reconstruction features exhibit the poorest performance across all evaluated aspects. These outcomes suggest two key insights: 1) The type of video features directly influences VAFlow's performance,",
+        "bbox": [
+            88,
+            672,
+            482,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "with different feature types leading to models with distinct performance preferences (either better semantic performance or synchronization); 2) Modality alignment matters more than information preservation. Although reconstruction features retain the most comprehensive video information among the three, their lack of modality alignment hampers the V2A model's ability to learn effective cross-modal generative capabilities.",
+        "bbox": [
+            511,
+            518,
+            906,
+            640
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Flow Matching Source Distributions. We evaluate VAFlow under two source distributions: video latent priors from alignment VAE and standard Gaussian noise. For video prior, we further test variants with/without joint training of the alignment VAE encoder (Method 3.3). All models share same training budgets (300K steps for non-joint training variants; 170K+130K for joint training).",
+        "bbox": [
+            511,
+            641,
+            908,
+            747
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Results in Table 3 demonstrate that video latent priors trained solely under alignment VAE objective in Equation 7 fail to surpass Gaussian noise baselines. Only when coupled with joint training—adapting the prior to flow-matching tasks—do video latent priors outperform Gaussian prior. This highlights the necessity of co-optimizing prior distributions for flow-based generation rather than relying on standalone pretraining on reconstruction task.",
+        "bbox": [
+            511,
+            748,
+            908,
+            869
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Conditioning Mechanisms and Guidance. Prior work highlights the critical role of condition and guidance mech",
+        "bbox": [
+            511,
+            869,
+            906,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "11783",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/72a285c0140b286cc05b89f5a361a88df831d6a8cdedaf0be2c027104301c080.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>V-Prior</td><td>Tuned</td><td>\\(FD_{pas.}\\)↓</td><td>\\(IS_{mel.}\\)↑</td><td>Acc.↑</td><td>IB-VA ↑</td></tr><tr><td>X</td><td>-</td><td>102.6</td><td>67.6</td><td>87.9</td><td>27.9</td></tr><tr><td>✓</td><td>X</td><td>111.8</td><td>57.4</td><td>82.3</td><td>25.4</td></tr><tr><td>✓</td><td>✓</td><td>87.7</td><td>73.6</td><td>88.1</td><td>28.6</td></tr></table>",
+        "bbox": [
+            107,
+            88,
+            468,
+            157
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/b3747903e3af065be5805bc844baf66fb78cff9f80dc36e605eea64dee29096e.jpg",
+        "table_caption": [
+            "Table 3. Ablation study on flow-matching source distributions. 'V-Prior' denotes sampling starts from video latent priors (vs. Gaussian priors). 'Tuned' indicates joint optimization of the alignment VAE encoder. Results show that jointly tuned video priors outperform Gaussian priors as source distributions for V2A flow models."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Cond.</td><td>CA</td><td>CFG</td><td>FDpas. ↓</td><td>ISmel. ↑</td><td>Acc. ↑</td><td>IB-VA ↑</td></tr><tr><td>X</td><td>X</td><td>-</td><td>135.9</td><td>41.4</td><td>31.7</td><td>8.5</td></tr><tr><td>X</td><td>✓</td><td>X</td><td>160.5</td><td>36.5</td><td>18.9</td><td>5.4</td></tr><tr><td>✓</td><td>✓</td><td>X</td><td>133.8</td><td>42.3</td><td>78.0</td><td>21.0</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>87.7</td><td>73.6</td><td>88.1</td><td>28.6</td></tr></table>",
+        "bbox": [
+            96,
+            268,
+            478,
+            349
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Ablation study on conditioning mechanisms and guidance. 'Cond.' denotes explicit video conditioning during inference; 'CA' indicates the presence of cross-attention layers of different variants; 'CFG' refers to classifier-free guidance. Results demonstrate that optimal performance requires both explicit conditioning and CFG when solving VAFlow ODE.",
+        "bbox": [
+            89,
+            361,
+            482,
+            445
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "anisms (e.g., classifier-free guidance, CFG) in enhancing model performance. We address two questions:",
+        "bbox": [
+            89,
+            489,
+            482,
+            521
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(1) Can VAFlow generate high-quality audio using only video priors, omitting explicit video conditioning during sampling? We experimented with two variants without explicit condition: No cross-attention (CA): Remove CA layers in DiT blocks and retrain the model (Table 4, Row 1). Zero-input CA: Retain CA layers but zero out video inputs during inference (Table 4, Row 2). 'No CA' variant exhibit comparable audio quality metrics (FD, IS) to the video-condition-visible variant (Table 4, Row 3: explicit video conditioning without CFG), yet suffer significant degradation in semantic relevance (IB-VA) and synchronization (Acc.). This indicates that while video priors partially preserve visual information, explicit video conditioning during sampling is essential for temporal and semantic alignment. We think that as the latent trajectory approaches the target audio distribution  $(t\\to 1)$ , initial visual cues in the prior degrade, necessitating direct video information.",
+            "(2) Does explicit video conditioning during sampling still require CFG for optimal performance? Adding CFG to the video-condition-visible variant (Row 4 in Table 4) further improves all metrics, demonstrating VAFlow's compatibility with established guidance techniques. This underscores the dual necessity of explicit video conditioning and CFG for optimal V2A synthesis with VAFlow framework."
+        ],
+        "bbox": [
+            89,
+            529,
+            483,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/541834df9a135b3f673f2135fc672b6335d888f01e8ee58e170b788fb154322c.jpg",
+        "image_caption": [
+            "A case video:",
+            "At a seaside with rolling waves, a group of people are swimming and playing. A woman walks toward the camera, points to the right, and says something.",
+            "Figure 5. Mel-spectrograms of audio generated by VAFlow and the baseline for a complex seaside scene, including background waves and varying voices, demonstrating VAFlow's superior content understanding and synchronization with the video."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            516,
+            118,
+            903,
+            449
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.6. Case Studies",
+        "text_level": 1,
+        "bbox": [
+            513,
+            544,
+            648,
+            556
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 5 illustrates the performance of VAFlow and the baselines on the complex scene video. The video depicts a seaside scene with various sounds include background waves and different voices associated with shifting characters (noisy laughter, woman speaking). We present the ground truth and the mel-spectrograms of the audio generated by different methods. Our results show that VAFlow not only accurately understands the video content and generates all the necessary sounds but also maintains synchronization with the visual timing.",
+        "bbox": [
+            511,
+            566,
+            906,
+            719
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion and Future Work",
+        "text_level": 1,
+        "bbox": [
+            511,
+            737,
+            782,
+            752
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we introduce VAFlow, a novel framework for V2A task, which directly denoises from the video space to the audio space via flow matching. demonstrating superior performance compared to previous methods. Future work will explore VAFlow's potential in more diverse audio domains, such as speech and music. Expanding the range of tasks and datasets will enable VAFlow's scaling properties to shine even further, laying a more general-purpose the foundation model for content generation community.",
+        "bbox": [
+            511,
+            763,
+            906,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "11784",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            91,
+            90,
+            250,
+            107
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work is supported by the National Natural Science Foundation of China (No. 62276268), Public Computing Cloud, Renmin University of China and ZHI-TECH GROUP.",
+        "bbox": [
+            89,
+            114,
+            483,
+            175
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            91,
+            186,
+            187,
+            203
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Michael S. Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In ICLR, 2023. 2",
+            "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 7",
+            "[3] 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. 1",
+            "[4] Li-Wen Chang, Wenlei Bao, Qi Hou, Chengquan Jiang, Ningxin Zheng, Yinmin Zhong, Xuanrun Zhang, Zuquan Song, Chengji Yao, Ziheng Jiang, et al. Flux: fast software-based communication overlap on gpus through kernel fusion. arXiv preprint arXiv:2406.06858, 2024. 1",
+            "[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, pages 721–725, 2020. 5",
+            "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 4",
+            "[7] Ziyang Chen, Prem Seetharaman, Bryan Russell, Oriol Nieto, David Bourgin, Andrew Owens, and Justin Salamon. Video-guided foley sound generation with multimodal controls. arXiv preprint arXiv:2411.17698, 2024. 2, 7",
+            "[8] Ho Kei Cheng, Masato Ishii, Akio Hayakawa, Takashi Shibuya, Alexander Schwing, and Yuki Mitsufuji. Taming multimodal joint training for high-quality video-to-audio synthesis. arXiv preprint arXiv:2412.15322, 2024. 2, 3, 6, 7",
+            "[9] Xin Cheng, Xihua Wang, Yihan Wu, Yuyue Wang, and Ruihua Song. Lova: Long-form video-to-audio generation. arXiv preprint arXiv:2409.15157, 2024. 1, 2, 6, 7",
+            "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633, 2024. 1",
+            "[11] Zach Evans, Julian D Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. Stable audio open. arXiv preprint arXiv:2407.14358, 2024. 3, 5",
+            "[12] 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 CVPR, pages 15180-15190, 2023. 5"
+        ],
+        "bbox": [
+            93,
+            213,
+            483,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[13] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2",
+            "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5",
+            "[15] Vladimir iashin and Esa Rahtu. Taming visually guided sound generation. In BMVC, 2021. 1, 2, 5, 7",
+            "[16] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Synchformer: Efficient synchronization from sparse cues. In ICASSP, pages 5325-5329, 2024. 7",
+            "[17] 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, 2022. 2",
+            "[18] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 5",
+            "[19] Khaled Koutini, Jan Schlüter, Hamid Eghbal-Zadeh, and Gerhard Widmer. Efficient training of audio transformers with patchout. arXiv preprint arXiv:2110.05069, 2021. 5",
+            "[20] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In ICLR, 2023. 2, 3",
+            "[21] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. In International Conference on Machine Learning, pages 21450-21474, 2023. 5",
+            "[22] Qihao Liu, Xi Yin, Alan Yuille, Andrew Brown, and Mannat Singh. Flowing from words to pixels: A noise-free framework for cross-modality evolution. In CVPR, pages 2755-2765, 2025. 2",
+            "[23] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 2, 3",
+            "[24] Xiulong Liu, Kun Su, and Eli Shlizerman. Tell what you hear from what you see - video to audio generation through text. In Advances in Neural Information Processing Systems, pages 101337-101366, 2024. 1, 2, 7",
+            "[25] Simian Luo, Chuanhao Yan, Chenxu Hu, and Hang Zhao. Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models. In Advances in Neural Information Processing Systems, pages 48855-48876, 2023. 1, 2, 3, 5, 6, 7",
+            "[26] Xinhao Mei, Varun Nagaraja, Gael Le Lan, Zhaoheng Ni, Ernie Chang, Yangyang Shi, and Vikas Chandra. Foleygen: Visually-guided audio generation. In 2024 IEEE 34th International Workshop on Machine Learning for Signal Processing, pages 1-6, 2024. 1, 2, 3",
+            "[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171, 2021. 2, 5"
+        ],
+        "bbox": [
+            516,
+            92,
+            905,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "11785",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[28] Shen Nie, Fengqi Zhu, Chao Du, Tianyu Pang, Qian Liu, Guangtao Zeng, Min Lin, and Chongxuan Li. Scaling up masked diffusion models on text. arXiv preprint arXiv:2410.18514, 2024. 2",
+            "[29] Santiago Pascual, Chunghsin Yeh, Ioannis Tsiamas, and Joan Serra. Masked generative video-to-audio transformers with enhanced synchronicity. In ECCV, pages 247-264, 2024. 1, 2, 3",
+            "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 2",
+            "[31] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2024. 1",
+            "[32] 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, 2021. 2, 3, 5",
+            "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 5",
+            "[34] Roy Sheffer and Yossi Adi. I hear your true colors: Image guided audio generation. In ICASSP, pages 1-5, 2023. 1, 2, 7",
+            "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2",
+            "[36] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4",
+            "[37] Kun Su, Xiulong Liu, and Eli Shlizerman. From vision to audio and beyond: A unified model for audio-visual representation and generation. In International Conference on Machine Learning, pages 46804–46822, 2024. 2, 7"
+        ],
+        "bbox": [
+            91,
+            90,
+            480,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[38] Anni Tang, Tianyu He, Junliang Guo, Xinle Cheng, Li Song, and Jiang Bian. Vidtok: A versatile and open-source video tokenizer. arXiv preprint arXiv:2412.13061, 2024. 2, 3, 5",
+            "[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7",
+            "[40] Heng Wang, Jianbo Ma, Santiago Pascual, Richard Cartwright, and Weidong Cai. V2a-mapper: A lightweight solution for vision-to-audio generation by connecting foundation models. In AAAI, pages 15492-15501, 2024. 1, 2, 7",
+            "[41] Xihua Wang, Yuyue Wang, Yihan Wu, Ruihua Song, Xu Tan, Zehua Chen, Hongteng Xu, and Guodong Sui. Tiva: Time-aligned video-to-audio generation. In ACM MM, pages 573-582, 2024. 1, 2, 3, 6, 7",
+            "[42] Xiang Wang, Shiwei Zhang, Hangjie Yuan, Zhiwu Qing, Biao Gong, Yingya Zhang, Yujun Shen, Changxin Gao, and Nong Sang. A recipe for scaling up text-to-video generation with text-free videos. In CVPR, pages 6572-6582, 2024. 2",
+            "[43] Yongqi Wang, Wenxiang Guo, Rongjie Huang, Jiawei Huang, Zehan Wang, Fuming You, Ruiqi Li, and Zhou Zhao. Frieren: Efficient video-to-audio generation network with rectified flow matching. In Advances in Neural Information Processing Systems, pages 128118-128138, 2024. 1, 2, 3, 6, 7",
+            "[44] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In ECCV, pages 378-394, 2024. 2",
+            "[45] Yazhou Xing, Yingqing He, Zeyue Tian, Xintao Wang, and Qifeng Chen. Seeing and hearing: Open-domain visual-audio generation with diffusion latent aligners. In CVPR, pages 7151-7161, 2024. 1, 2, 6, 7",
+            "[46] Yiming Zhang, Yicheng Gu, Yanhong Zeng, Zhening Xing, Yuancheng Wang, Zhizheng Wu, and Kai Chen. Foley-crafter: Bring silent videos to life with lifelike and synchronized sounds. arXiv preprint arXiv:2407.01494, 2024. 2, 3, 6, 7",
+            "[47]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.2"
+        ],
+        "bbox": [
+            516,
+            90,
+            903,
+            696
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "11786",
+        "bbox": [
+            480,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/7812feed-641c-4918-a582-f934ca29c2a8_model.json

@@ -0,0 +1,2024 @@
+[
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.107,
+                0.003,
+                0.182,
+                0.044
+            ],
+            "angle": 0,
+            "content": "CVF"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.239,
+                0.001,
+                0.808,
+                0.047
+            ],
+            "angle": 0,
+            "content": "This ICCV 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": [
+                0.128,
+                0.131,
+                0.872,
+                0.154
+            ],
+            "angle": 0,
+            "content": "VAFlow: Video-to-Audio Generation with Cross-Modality Flow Matching"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.179,
+                0.888,
+                0.234
+            ],
+            "angle": 0,
+            "content": "Xihua Wang\\(^{1}\\) Xin Cheng\\(^{1}\\) Yuyue Wang\\(^{1}\\) Ruihua Song\\(^{1*}\\) Yunfeng Wang\\(^{2}\\)  \n\\(^{1}\\)Gaoling School of Artificial Intelligence, Renmin University of China  \nxihuaw@ruc.edu.cn, songruihua.bloon@outlook.com"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.249,
+                0.269,
+                0.327,
+                0.285
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.088,
+                0.303,
+                0.486,
+                0.651
+            ],
+            "angle": 0,
+            "content": "Video-to-audio (V2A) generation aims to synthesize temporally aligned, realistic sounds for silent videos, a critical capability for immersive multimedia applications. Current V2A methods, predominantly based on diffusion or flow models, rely on suboptimal noise-to-audio paradigms that entangle cross-modal mappings with stochastic priors, resulting in inefficient training and convoluted transport paths. We propose VAFlow, a novel flow-based framework that directly models the video-to-audio transformation, eliminating reliance on noise priors. To address modality discrepancies, we employ an alignment variational autoencoder that compresses heterogeneous video features into audio-aligned latent spaces while preserving spatiotemporal semantics. By retaining cross-attention mechanisms between video features and flow blocks, our architecture enables classifier-free guidance within video source-driven generation. Without external data or complex training tricks, VAFlow achieves state-of-the-art performance on VGGSound benchmark, surpassing even textaugmented models in audio fidelity, diversity, and distribution alignment. This work establishes a new paradigm for V2A generation with a direct and effective video-to-audio transformation via flow matching."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.092,
+                0.682,
+                0.222,
+                0.699
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.708,
+                0.485,
+                0.875
+            ],
+            "angle": 0,
+            "content": "Video-to-audio (V2A) generation plays a vital role in multimedia content generation, including Foley for films and AI-generated silent videos. In recent years, this task has attracted increasing attention within the generative community. Early V2A approaches [15, 24, 26, 29, 34] adopted language modeling strategies by discretizing audio into tokens, and used parallel mask-prediction models (BERT-like) or token-by-token autoregressive models (GPT-like) borrowed from NLP community as shown in Figure 1. However, these methods, reliant on discrete tokens, incur inherent information losses by the discretization process."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.517,
+                0.267,
+                0.907,
+                0.474
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.512,
+                0.486,
+                0.909,
+                0.612
+            ],
+            "angle": 0,
+            "content": "Figure 1. Four different V2A paradigms, along with state-of-the-art performance under each. The first row shows discrete token-based approaches (mask-prediction and autoregressive based generation), which are limited by information loss during tokenization. The second row presents methods generating continuous data. Previous diffusion and flow methods (bottom-left) model the noise-to-audio mapping conditioning on video. We find this approach suboptimal and demonstrate that directly modeling the video-to-audio transformation (bottom-right) yields better results."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.642,
+                0.909,
+                0.824
+            ],
+            "angle": 0,
+            "content": "Recent advancements in Diffusion and Flow Matching (FM) models in image [4, 10, 33] and video [3, 31] generation have spurred similar approaches for V2A [9, 25, 41, 43, 45]. Generally, these methods model V2A as a conditional generation task, transforming Gaussian noise conditioned on video into audio, as shown in Figure 1 (bottom-left). The generation process typically involves two steps: (1) sampling a random noise from Gaussian prior and (2) using this noise and video condition as input to iteratively denoise the noise into audio. This approach avoids discrete audio tokens, directly fitting the continuous distribution of audio data, and currently holds state-of-the-art results."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.825,
+                0.909,
+                0.903
+            ],
+            "angle": 0,
+            "content": "However, existing diffusion/FM-based V2A methods focus solely on the second step, i.e., denoising Gaussian noise into audio, assuming the first step is trivial. Even state-of-the-art methods [40, 43] begin with Gaussian noise, but we find that this approach is not optimal. Unlike text-to"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.11,
+                0.888,
+                0.237,
+                0.901
+            ],
+            "angle": 0,
+            "content": "*Corresponding author."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "11777"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.092,
+                0.482,
+                0.288
+            ],
+            "angle": 0,
+            "content": "image or text-to-video tasks, where textual conditions are abstract and highly defined by human, the video conditions in V2A are more complex and require both spatial and temporal understanding. The generation goal also demands a high degree of alignment between the generated audio and the video conditions, not just semantically, but also temporally. There are two inherent difficulties: (1) different conditions on the same noise require the model to generate different audio outputs, demanding advanced condition processing [8, 41, 46], and (2) noise variations under the same condition lead to unstable generation results, prompting research into methods for noise control [44], such as finding a \"golden noise\" for consistent results [47]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.291,
+                0.482,
+                0.426
+            ],
+            "angle": 0,
+            "content": "In contrast, we propose a new perspective by shifting focus to the first step of the generation process. We question: If there are better priors than Gaussian noise? Given that many diffusion and FM models dedicate significant effort to integrating video conditions into the denoising process for better generation performance, we further question: If V2A can work better when directly transforming the video distribution to audio rather than from noise? Fortunately, both answers are affirmative."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.43,
+                0.483,
+                0.747
+            ],
+            "angle": 0,
+            "content": "We introduce VAFlow, a novel framework for V2A generation, which directly denoises from the video space to the audio space. To implement this video-to-audio paradigm, we adopt the FM framework, as it theoretically supports any source distribution. There are three technical challenges for this implementation: (1) aligning source and target distributions: video and audio feature spaces typically have different time and spatial resolutions, yet FM requires the source and target distributions to be aligned (the same shape). (2) uncertainty of the feature type: identifying the optimal video representations/feature types as the source distribution. (3) conditioning mechanism: exploring additional video information integration ways beyond encoding it in the source distribution. In VAFlow, we first employ an alignment VAE to adjust video features to match the audio latent space. It then explores three different video feature types (semantic CLIP [32], spatiotemporal CAVP [25], reconstruction-focused VidTok [38]) to systematically investigate source video information. Experiments show that using video features both as a source distribution and as conditioning information yields the best results for VAFlow."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.75,
+                0.483,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Moreover, the recent success of large language models (LLMs) [13, 28] has highlighted the effectiveness of scaling laws, and various generative works consequently explore similar scaling properties in each domain [30, 42]. However, existing V2A models have not demonstrated clear scaling trends, where increasing model size does not consistently achieve improvement [8, 43]. In contrast, our experiments show that VAFlow exhibits clear scaling property, with model size increases leading to consistent improvements in generation quality, further confirming its potential."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.092,
+                0.905,
+                0.198
+            ],
+            "angle": 0,
+            "content": "Without elaborate training tricks or external data (e.g., text-audio pairs), VAFlow achieves state-of-the-art results on standard benchmarks (VGGSound), outperforming even text-augmented models in generation fidelity, diversity, and cross-modality alignment with input video. Playable audi-los generated with VAFlow can be accessed at the demo page: https://vaflow.github.io/demo."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.213,
+                0.655,
+                0.229
+            ],
+            "angle": 0,
+            "content": "2. Related Work"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.239,
+                0.836,
+                0.255
+            ],
+            "angle": 0,
+            "content": "2.1. Flow Matching and Diffusion Models"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.261,
+                0.905,
+                0.549
+            ],
+            "angle": 0,
+            "content": "Diffusion models [17, 27, 35] reconstruct data distributions by iteratively denoising Gaussian-sampled latents, governmented by various designed schedulers achieving state-of-the-art performance in multimodal generation tasks. Unlike diffusions, flow-based models [1, 20, 22, 23] establish deterministic mappings between source (e.g., Gaussian) and target distributions through invertible neural transformations, with recent flow matching optimizing vector fields through optimal transport theory, enabling direct learning of straight transport paths between source and target distributions, thereby reducing training instability and inference steps. While both paradigms typically rely on Gaussian priors as their source distribution, existing video-to-audio systems on both paradigms model noise-to-audio generation conditioned on video features. In contrast, our approach directly aligns video latent spaces with target audio distributions via linear transformations under a flow matching framework, bypassing iterative noise modeling, preserving efficiency, and ensuring precise frame-to-frame alignment."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.56,
+                0.759,
+                0.575
+            ],
+            "angle": 0,
+            "content": "2.2. Video-to-Audio Generation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.583,
+                0.905,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Framework. Recent advances in generative models have driven significant progress in video-to-audio (V2A) synthesis. Current V2A methods fall into three main paradigms: (1) Autoregressive (AR) models [15, 26, 34] discretize audio into token (codec) sequences and utilize video features as prefixes to model unified video-audio sequences. While straightforward, their token-by-token generation process results in quadratic inference time scaling, limiting efficiency for long audio sequences. (2) Mask-based generation paradigms [24, 29, 37] also employ discrete tokens but decode audio in parallel by predicting masked tokens. Video features serve as the condition for mask-to-audio prediction. However, both AR and mask-based methods suffer from an inherent upper bound of fidelity due to lossy compression from audio discretization. (3) Flow/diffusion-based methods [7-9, 25, 40, 41, 43, 45, 46] leverage continuous representations and currently achieve state-of-the-art V2A performance [8, 40, 43]. These techniques iteratively transform noise into target audio distributions while conditioning on video features to steer the denoising trajectory."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.886,
+                0.905,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Video Conditional Modeling. Unlike typical conditioned"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.518,
+                0.956
+            ],
+            "angle": 0,
+            "content": "11778"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.092,
+                0.485,
+                0.334
+            ],
+            "angle": 0,
+            "content": "generation tasks like text-to-image [33], V2A involves fully aligned sequential correspondences between video and audio. Existing methods address this through various strategies: some AR-based methods investigate various attention mechanisms and masking schemes for temporally sequential alignment [26]; some mask-based approaches integrate hybrid ways to fuse video features into the mask-prediction process to synchronize audio generation [29]; and some flow/diffusion-based methods augment their pipelines with auxiliary modules (e.g., loudness or audio layout predictors [41, 46]) to enforce temporal consistency. In contrast, our approach builds on flow-based techniques by discarding the conventional noise-to-audio paradigm and complex video-conditioning designs. Instead, we propose a more intuitive method that directly transforms the video temporal space to the audio temporal space via a flow transport."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.346,
+                0.188,
+                0.363
+            ],
+            "angle": 0,
+            "content": "3. Method:"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.371,
+                0.35,
+                0.387
+            ],
+            "angle": 0,
+            "content": "3.1. Preliminary: Flow Matching"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.393,
+                0.484,
+                0.471
+            ],
+            "angle": 0,
+            "content": "Let \\( x_0 \\sim p(x_0) \\) and \\( x_1 \\sim p(x_1) \\) denote samples from source and target distributions, respectively, with \\( c \\) representing an optional conditioning signal. Flow Matching (FM) [20, 23] establishes an inter-distribution transport process governed by the ordinary differential equation (ODE):"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.21,
+                0.48,
+                0.483,
+                0.498
+            ],
+            "angle": 0,
+            "content": "\\[\nd x (t) = v _ {\\theta} \\left(x _ {t}, t, c\\right) d t, \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.507,
+                0.484,
+                0.6
+            ],
+            "angle": 0,
+            "content": "where \\( t \\in [0,1] \\) denotes the continuous time step, and \\( v_{\\theta} \\) represents a neural network-parameterized velocity field. The condition \\( c \\) is incorporated for conditional flow implementations. Optimal transport flows enforce velocity alignment with the direction \\( (x_{1} - x_{0}) \\) of the linear path pointing from \\( x_{0} \\) to \\( x_{1} \\) through regression:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.135,
+                0.608,
+                0.483,
+                0.627
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} (\\theta) = \\mathbb {E} _ {t, x _ {0}, x _ {1}, c} \\| (x _ {1} - x _ {0}) - v _ {\\theta} (x _ {t}, t, c) \\| ^ {2}, \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.637,
+                0.483,
+                0.681
+            ],
+            "angle": 0,
+            "content": "where \\( x_{t} = (1 - t)x_{0} + tx_{1} \\) constitutes the linear interpolation between \\( x_{0} \\) and \\( x_{1} \\). The learned velocity field enables target generation by solving the ODE in Equation 1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.682,
+                0.483,
+                0.786
+            ],
+            "angle": 0,
+            "content": "Previous video-to-audio flow approaches [8, 43] (and related diffusion methods [25, 41]) use a Gaussian prior \\( p(x_0) = \\mathcal{N}(0,1) \\) for the source distribution, with audio as \\( p(x_{1}) \\) and video as conditioning input \\( c \\). During training, random noise \\( x_0 \\) is paired with ground-truth audio \\( x_{1} \\), prompting \\( v_{\\theta} \\) to model the paths from noise to audio under visual conditions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.788,
+                0.484,
+                0.833
+            ],
+            "angle": 0,
+            "content": "This work challenges the Gaussian prior by proposing a direct video-to-audio transport. Theoretically, FM permits arbitrary source distributions, motivating our formulation:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.217,
+                0.844,
+                0.483,
+                0.862
+            ],
+            "angle": 0,
+            "content": "\\[\nd x (t) = v _ {\\theta} \\left(x _ {t}, t\\right) d t, \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.871,
+                0.484,
+                0.903
+            ],
+            "angle": 0,
+            "content": "where \\( x_{t} = (1 - t)x_{v} + tx_{a} \\) linearly interpolates between paired video and audio samples \\( x_{v},x_{a}\\sim p(v,a) \\)"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.091,
+                0.715,
+                0.106
+            ],
+            "angle": 0,
+            "content": "3.2. VAFlow Architecture"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.115,
+                0.907,
+                0.328
+            ],
+            "angle": 0,
+            "content": "Model Overview. As illustrated in Figure 2, the input video features \\( x_{v} \\) are extracted via a visual encoder, while the target audio waveform can be obtained from latent representations \\( x_{a} \\) through WaveVAE. VAFlow attempts to establish FM between the video feature space \\( x_{v} \\) and audio latent space \\( x_{a} \\), enabling direct audio generation from video features via ODE solving instead of Gaussian noise initialization. However, since FM requires consistent dimensional shapes between source and target distributions (see Equation 3), the inherent spatiotemporal resolution mismatch between \\( x_{v} \\) and \\( x_{a} \\) necessitates shape alignment. To address this, we employ a 1D VAE (namely alignment VAE) that compresses interpolated \\( x_{v} \\) into aligned features \\( x_{v}^{\\mathrm{AE}} \\) matching \\( x_{a} \\)'s dimensions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.331,
+                0.906,
+                0.451
+            ],
+            "angle": 0,
+            "content": "A diffusion transformer (DiT) subsequently works as a velocity field estimator to capture the transport dynamics between \\( x_{v}^{\\mathrm{AE}} \\) and \\( x_{a} \\). This enables continuous trajectory sampling from \\( x_{v}^{\\mathrm{AE}} \\) to \\( x_{a} \\) through ODE integration. The sampled latent \\( \\hat{x}_{a} \\) is finally decoded by WaveVAE to produce the output audio. The following contents detail: the visual/audio representations \\( (x_{v}, x_{v}^{\\mathrm{AE}}, x_{a}) \\), and the velocity field estimator in VAFlow."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.454,
+                0.907,
+                0.529
+            ],
+            "angle": 0,
+            "content": "Video and Audio Representations. For an input video-audio pair \\((v, a)\\), the audio waveform \\(a\\) is compressed by a pretrained 1D WaveVAE's [11] encoder into \\(x_{a} \\in \\mathbb{R}^{T_{a} \\times D_{a}}\\). This latent \\(x_{a}\\) serves as the target for flow generation and is later decoded to reconstruct the waveform."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.533,
+                0.907,
+                0.778
+            ],
+            "angle": 0,
+            "content": "For video representation, a pretrained visual encoder encodes \\( v \\) into \\( x_v \\in \\mathbb{R}^{T_v \\times D_v} \\) (e.g., CLIP [32], CAVP [25], or VidTok [38] as encoder). Given the discrepancies (e.g., \\( x_a \\in \\mathbb{R}^{215 \\times 64} \\) versus \\( x_v \\in \\mathbb{R}^{100 \\times 768} \\) for a 10-second clip), we propose an alignment VAE \\( \\epsilon_{\\phi} \\) to further compress and align video latents with audio latents' space while preserving essential visual information. Specifically, we utilize a 1D convolutional encoder-decoder as \\( \\epsilon_{\\phi} \\) to interpolate and project \\( x_v \\) into a variational space \\( x_v^{\\mathrm{AE}} \\in \\mathbb{R}^{T_a \\times D_a} \\), matching the shape of \\( x_a \\). The encoder of \\( \\epsilon_{\\phi} \\) predicts the parameters \\( \\mu_{x_v^{\\mathrm{AE}}}, \\sigma_{x_v^{\\mathrm{AE}}} = E_{\\epsilon_{\\phi}}(x_v) \\) for a Gaussian distribution, from which \\( x_v^{\\mathrm{AE}} \\sim \\mathcal{N}(\\mu_{x_v^{\\mathrm{AE}}}, \\sigma_{x_v^{\\mathrm{AE}}}) \\) is sampled. The decoder reconstructs \\( x_v \\) as \\( \\hat{x}_v = D_{\\epsilon_{\\phi}}(x_v^{\\mathrm{AE}}) \\). This process ensures dimensional alignment with audio latents while retaining critical spatiotemporal features of the original video representation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.78,
+                0.909,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Velocity Estimator. Built upon the diffusion transformer (DiT) architecture [11], the velocity estimator processes two key inputs: time embeddings of \\( t \\) and the interpolated state \\( x_{t} = (1 - t)x_{v}^{\\mathrm{AE}} + tx_{a} \\), following the Equation 3. The time embedding is concatenated with the audio sequence to form an input tensor \\( \\in \\mathbb{R}^{(T_a + 1)\\times D_a} \\), which is processed through stacked transformer blocks. The first \\( T_{a} \\) dimensions of the output sequence constitute the pre"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "11779"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.097,
+                0.09,
+                0.9,
+                0.226
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.09,
+                0.241,
+                0.907,
+                0.325
+            ],
+            "angle": 0,
+            "content": "Figure 2. VAFlow Architecture. The figure illustrates VAFow in the inference manner. A visual encoder initially extracts temporal features \\( x_{v} \\in \\mathbb{R}^{T_{v} \\times D_{v}} \\) from variable-frame-rate video inputs. Then, a 1D alignment VAE (Variational Autoencoder) \\( \\epsilon_{\\phi} \\) compresses interpolated \\( x_{v} \\) into audio-aligned latent features \\( x_{v}^{\\mathrm{AE}} \\) with matched dimensions \\( (T_{v}^{\\mathrm{AE}} \\times D_{v}^{\\mathrm{AE}} = T_{a} \\times D_{a}) \\). Next, a diffusion transformer \\( (v_{\\theta}) \\) predicts the velocity field for the transport \\( (x_{a} - x_{v}^{\\mathrm{AE}}) \\), and by solving the corresponding ODE (see Equation 3), an estimated \\( \\hat{x}_{a} \\) is obtained. Finally, the WaveVAE decoder converts \\( \\hat{x}_{a} \\) into the waveform representing the predicted audio. Note that during diffusion transformer prediction, \\( x_{v} \\) can optionally be incorporated via cross-attention, detail discussion is conducted in the experiment Section 4.5."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.351,
+                0.482,
+                0.502
+            ],
+            "angle": 0,
+            "content": "dicted velocity \\( v_{\\theta}(x_t, t) \\), with 1D Rotary Position Embeddings (RoPE [36]) to preserve positional awareness in attention layers. Notably, while the interpolated state \\( x_t \\) implicitly encodes visual information through \\( x_v^{\\mathrm{AE}} \\), this signal diminishes as \\( t \\to 1 \\). To address this limitation, we retain cross-attention layers in the transformer blocks that persistently integrate the original video features \\( x_v \\). Further studies (Section 4.5) demonstrate that this architectural choice is critical to enhance visual perception throughout the flow trajectory and improve audio generation quality."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.511,
+                0.373,
+                0.527
+            ],
+            "angle": 0,
+            "content": "3.3. VAFlow Training and Inference"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.533,
+                0.483,
+                0.594
+            ],
+            "angle": 0,
+            "content": "We conduct three-stage training: (1) pretrain the alignment VAE, (2) train the VAFlow estimator with frozen alignment VAE, and (3) jointly fine-tune both alignment VAE and estimator. We detail training and inference as follows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.594,
+                0.483,
+                0.687
+            ],
+            "angle": 0,
+            "content": "Alignment VAE Training. Given encoded visual features \\( x_{v} \\in \\mathbb{R}^{T_{v} \\times D_{v}} \\), the alignment VAE encoder predicts distribution parameters \\( (\\mu_{x_{v}^{\\mathrm{AE}}}, \\sigma_{x_{v}^{\\mathrm{AE}}}) \\), while the decoder reconstructs \\( \\hat{x}_{v} \\) from a sampled \\( x_{v}^{\\mathrm{AE}} \\sim \\mathcal{N}(\\mu_{x_{v}^{\\mathrm{AE}}}, \\sigma_{x_{v}^{\\mathrm{AE}}}) \\). The model optimizes three objectives to preserve visual semantics while regularizing \\( x_{v}^{\\mathrm{AE}} \\) within \\( \\mathbb{R}^{T_{a} \\times D_{a}} \\) Gaussian space:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.091,
+                0.689,
+                0.472,
+                0.703
+            ],
+            "angle": 0,
+            "content": "- KL Regularization: enforces Gaussian latent priors via:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.15,
+                0.715,
+                0.483,
+                0.733
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\mathrm {K L}} = D _ {\\mathrm {K L}} \\left(\\mathcal {N} \\left(\\mu_ {x _ {0} ^ {\\mathrm {A E}}}, \\sigma_ {x _ {0} ^ {\\mathrm {A E}}}\\right) \\| \\mathcal {N} (0, I)\\right). \\tag {4}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.091,
+                0.743,
+                0.472,
+                0.758
+            ],
+            "angle": 0,
+            "content": "- Reconstruction Loss: ensures feature-level consistency:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.225,
+                0.769,
+                0.483,
+                0.786
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\text {r e c}} = \\left\\| x _ {v} - \\hat {x} _ {v} \\right\\| _ {2} ^ {2}. \\tag {5}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.091,
+                0.798,
+                0.479,
+                0.813
+            ],
+            "angle": 0,
+            "content": "- Contrastive Alignment: strengthens feature preservation:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.115,
+                0.822,
+                0.483,
+                0.859
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\text {c o n}} = - \\sum_ {s \\in \\mathcal {S}} \\log \\frac {\\exp \\left(f _ {s} \\left(\\hat {x} _ {v}\\right) \\cdot f _ {s} \\left(x _ {v}\\right) / \\tau\\right)}{\\sum_ {x ^ {\\prime} \\in \\mathcal {B}} \\exp \\left(f _ {s} \\left(\\hat {x} _ {v}\\right) \\cdot f _ {s} \\left(x ^ {\\prime}\\right) / \\tau\\right)}, \\tag {6}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.104,
+                0.871,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "where \\(S = \\{\\mathrm{frame},\\mathrm{clip},\\mathrm{global}\\}\\) defines temporal granularities and \\(\\tau = 0.7\\) serves as a temperature parameter [6]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.526,
+                0.351,
+                0.907,
+                0.427
+            ],
+            "angle": 0,
+            "content": "For scale \\( s \\), \\( f_{s}(\\cdot) \\) extracts features via three ways: frame-level: single-frame sampling + mean pooling; clip-level: \\( K = 8 \\)-adjacent-frame sampling + pooling; global-level: full temporal mean pooling at \\( T_{v} \\) dimension of feature \\( x \\). Negative samples \\( x' \\) are drawn from training batch \\( \\mathcal{B} \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.429,
+                0.828,
+                0.444
+            ],
+            "angle": 0,
+            "content": "The composite objective integrates these terms:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.577,
+                0.454,
+                0.905,
+                0.469
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\mathrm {V A E}} = \\lambda_ {\\mathrm {K L}} \\mathcal {L} _ {\\mathrm {K L}} + \\lambda_ {\\mathrm {r e c}} \\mathcal {L} _ {\\mathrm {r e c}} + \\lambda_ {\\mathrm {c o n}} \\mathcal {L} _ {\\mathrm {c o n}}, \\tag {7}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.479,
+                0.84,
+                0.494
+            ],
+            "angle": 0,
+            "content": "with \\(\\lambda_{\\mathrm{KL}} = 8\\times 10^{-4}\\) \\(\\lambda_{\\mathrm{rec}} = 1.0\\) , and \\(\\lambda_{\\mathrm{con}} = 1.0\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.495,
+                0.906,
+                0.54
+            ],
+            "angle": 0,
+            "content": "Velocity Estimator Training Given encoded features \\( x_{v} \\), \\( x_{v}^{\\mathrm{AE}} \\), and \\( x_{a} \\), we train the velocity estimator \\( v_{\\theta} \\) to predict the linear transport direction \\( x_{a} - x_{v}^{\\mathrm{AE}} \\) through:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.532,
+                0.549,
+                0.905,
+                0.571
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} (\\theta) = \\mathbb {E} _ {t \\sim \\mathcal {U} (0, 1)} \\left\\| \\left(x _ {a} - x _ {v} ^ {\\mathrm {A E}}\\right) - v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) \\right\\| _ {2} ^ {2}, \\tag {8}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.579,
+                0.906,
+                0.655
+            ],
+            "angle": 0,
+            "content": "where \\( x_{t} = (1 - t)x_{v}^{\\mathrm{AE}} + tx_{a} \\). For the conditioned version, visual features \\( x_{v} \\) are incorporated via cross-attention (with a \\( 10\\% \\) probability of zeroing out \\( x_{v} \\) during training). The vanilla version omits this conditioning, simplifying the input of estimator in Equation 8 to \\( v_{\\theta}(x_t,t) \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.656,
+                0.907,
+                0.806
+            ],
+            "angle": 0,
+            "content": "Joint Turning After separate training of the alignment VAE and the velocity estimator, a joint optimization stage is performed. In this stage, the VAE encoder and \\( v_{\\theta} \\) are co-trained (discarding the VAE decoder), while the WaveVAE and visual feature extractor remain frozen. For a pair \\( (v, a) \\), the visual extractor and WaveVAE produce \\( x_{v} \\) and \\( x_{a} \\). The VAE encoder compresses \\( x_{v} \\) to obtain \\( x_{v}^{\\mathrm{AE}} \\sim q_{\\phi}(x_{v}^{\\mathrm{AE}}|x_{v}) \\) which together with \\( x_{a} \\) and a random \\( t \\sim \\mathcal{U}(0,1) \\) forms \\( x_{t} \\). The joint objective integrates KL regularization in Equation 4 with velocity prediction in Equation 8:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.514,
+                0.814,
+                0.907,
+                0.847
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} (\\theta , \\phi_ {\\text {E n c}}) = \\lambda_ {\\mathrm {K L}} \\mathcal {L} _ {\\mathrm {K L}} + \\mathbb {E} \\| \\left(x _ {a} - x _ {v} ^ {\\mathrm {A E}}\\right) - v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) \\| _ {2} ^ {2}. \\tag {9}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.856,
+                0.907,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Inference Upon training, VAFlow synthesizes audio \\(a\\) from an input video \\(v\\) through the following pipeline: The visual extractor first encodes \\(v\\) into a feature \\(x_{v}\\), which is"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "11780"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.092,
+                0.482,
+                0.215
+            ],
+            "angle": 0,
+            "content": "compressed by the alignment VAE encoder into distribution parameters \\((\\mu_{x_v^{\\mathrm{AE}}},\\sigma_{x_v^{\\mathrm{AE}}})\\). A latent sample \\(x_{v}^{\\mathrm{AE}}\\sim \\mathcal{N}(\\mu_{x_v^{\\mathrm{AE}}},\\sigma_{x_v^{\\mathrm{AE}}}^2)\\) is drawn as the initial state \\((x_0)\\) for solving the ODE in Equation 1 with learned estimator. The predicted \\(\\hat{x}_a(x_1)\\) is then decoded into waveform via the frozen WaveVAE. During inference, the variational sampling of \\(x_{v}^{\\mathrm{AE}}\\) introduces stochasticity - even with fixed ODE integration paths, distinct samples yield diverse outputs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.216,
+                0.482,
+                0.261
+            ],
+            "angle": 0,
+            "content": "For conditional velocity estimators, classifier-free guidance (CFG) [14] enhances synthesis quality by interpolating conditioned and unconditioned velocity predictions:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.104,
+                0.269,
+                0.482,
+                0.302
+            ],
+            "angle": 0,
+            "content": "\\[\nv _ {\\theta} ^ {\\mathrm {C F G}} \\left(x _ {t}, t, x _ {v}\\right) = (1 + \\gamma) v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) - \\gamma v _ {\\theta} \\left(x _ {t}, t, \\emptyset\\right), \\tag {10}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.313,
+                0.482,
+                0.344
+            ],
+            "angle": 0,
+            "content": "where \\(\\gamma\\) controls guidance strength and \\(\\varnothing\\) denotes null (zero) conditioning."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.09,
+                0.357,
+                0.223,
+                0.373
+            ],
+            "angle": 0,
+            "content": "4. Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.09,
+                0.381,
+                0.281,
+                0.397
+            ],
+            "angle": 0,
+            "content": "4.1. Experimental Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.403,
+                0.482,
+                0.569
+            ],
+            "angle": 0,
+            "content": "Datasets. We employ two datasets: VGGSound [5] and AudioSet-V2A [25]. For VGGSound, we adhere to the original split, which comprises 186k training samples and 15k testing samples. AudioSet-V2A contains 390k videos that are exclusively used for training. During training, the alignment VAE is trained on a combined set of VGGSound training and AudioSet-V2A to yield more generalized visual feature compression. The subsequent velocity estimator training and joint tuning are conducted solely on the VGGSound training set. At inference, we evaluate on the VGGSound test set as in prior work."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.57,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Implementation Details. For the estimator's DiT architecture, we adapted the DiT implementation from StableAudio-Open-1.0 [11] (SDA1.0) by eliminating duration encoding, applying global time encoding solely to time steps, and restricting the cross-attention layer's key/value inputs to video features. For the visual extractor, we use three variants: CLIP [32] (a widely adopted visual semantic extractor in V2A tasks), CAVP [25] (which captures both visual semantic and temporal features), and VidTok [38] (a visual tokenizer for reconstruction). We train three alignment VAEs with comparable parameter scales—one for each visual extractor. In the velocity estimator training stage, the alignment VAE remains frozen; in the joint training stage, both the estimator and the alignment VAE are trained. In all stages, the audio WaveVAE and the visual extractor are kept frozen. Notably, we do not perform any further rectified iterations on the estimator; all VAFlow variants are trained for a single rectified iteration (Rectified-1, RF-1). During sampling, we employ ODE solvers of various orders—first-order Euler, second-order Midpoint, and fifth-order dopri5—with differing sampling steps and classifier-free guidance scales."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.092,
+                0.905,
+                0.287
+            ],
+            "angle": 0,
+            "content": "Metrics. We employ both objective and subjective evaluations across three dimensions: generation quality, audiovisual relevance, and audiovisual synchronization. For objective evaluation, generation quality is measured using FD score [21], KL divergence [15], and IS score [21]. In previous work, each metric may be computed with different feature extraction models (e.g., Melception, PasST\\(^1\\)). Audiovisual relevance is quantified by computing the cosine similarity between the embeddings of the input video and the generated audio using the ImageBind model [12]. For synchronization, we use the synchronization accuracy (Acc.) [25] reflects the alignment between the input video and generated audio."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.289,
+                0.905,
+                0.56
+            ],
+            "angle": 0,
+            "content": "Baselines. Our method is comprehensively compared against all major generative paradigms, including autoregressive (AR), masked prediction (Mask), diffusion-based, and flow-based approaches, as detailed in Table 1. All baseline models were evaluated using either officially released code and models or directly downloaded audio outputs by their authors, with consistent experimental settings maintained across all evaluations on the same hardware platform. Implementation specifics include: (1) Im2Wav generates audio in 4s clips, truncated to 10s after three clips. (2) Diff-Foley here employs its double-guidance variant (demonstrating superior performance); (3) MMAudio-L here is the largest variant of MMAudio; and (4) Frieren here employs its final no-reflow configuration (achieving optimal Fréchet distance). For comprehensive evaluation, we also include text-augmented V2A models (denoted in gray rows) that leverage external textual data beyond benchmark audiovisual corpora during training or inference."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.572,
+                0.843,
+                0.587
+            ],
+            "angle": 0,
+            "content": "4.2. Diffusion vs. Flow vs. VAFlow Models"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.595,
+                0.905,
+                0.775
+            ],
+            "angle": 0,
+            "content": "We examine framework-specific performance by training same VAFlow-CLIP-142M models under three paradigms, maintaining identical configurations: CLIP visual features, architecture parameters, initialization weights, and other hyperparameters. The frameworks include: Diffusion: Standard DDPM [27, 33] with cross-attention video conditioning. Flow: Standard continuous-time flow (Equation 2) using Gaussian prior. VAFlow: Our framework employing aligned VAE-compressed video features as source distribution. All frameworks use their respective optimal inference settings (DDPM: DPMSolver; Flows: dopri5). Joint tuning in VAFlow begins at 100K steps."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.777,
+                0.905,
+                0.852
+            ],
+            "angle": 0,
+            "content": "Frechet Distance (FD) results in Figure 3 reveals: 1) Flow paradigms converge faster with lower FD than diffusion, demonstrating inherent optimization benefits. 2) VAFlow attains lower final FD than standard flow despite delayed early-stage convergence (<100K steps). This per"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.512,
+                0.863,
+                0.905,
+                0.901
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>We adopt PasST [19] model over PANN [18] for its improved robustness following the practice introduced in AudioLDM [21] evaluation repository: https://github.com/haoheliu/audioldm_eval."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.518,
+                0.957
+            ],
+            "angle": 0,
+            "content": "11781"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.14,
+                0.091,
+                0.437,
+                0.24
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.09,
+                0.253,
+                0.485,
+                0.337
+            ],
+            "angle": 0,
+            "content": "Figure 3. Comparison of different training frameworks. We examine the same model under different training paradigms by controlling training data, steps, and other hyperparameters. Results demonstrate that VAFlow converges to superior optima, validating the effectiveness of its direct video-to-audio transformation design and highlighting this paradigm advantages in V2A generation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.364,
+                0.484,
+                0.456
+            ],
+            "angle": 0,
+            "content": "formance attributes to VAFlow's explicit modeling of deterministic video-audio transport paths, avoiding the ambiguous noise-to-audio mappings in Gaussian-prior approaches. VAFlow learns optimal transport trajectories between video-audio distribution, bypassing error-prone intermediate transformations in Gaussian-noise-to-audio paths."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.466,
+                0.32,
+                0.483
+            ],
+            "angle": 0,
+            "content": "4.3. VAFlow Scaling Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.489,
+                0.483,
+                0.579
+            ],
+            "angle": 0,
+            "content": "We investigate VAFlow's scaling capabilities by developing four model variants (142M, 264M, 527M, 1.05B parameters) through width/depth scaling of DiT blocks while maintaining fixed training protocols (data, steps, etc.). Evaluation metrics (FD, IS) versus parameter counts are visualized in Figure 4 (bubble sizes denote model scales)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.58,
+                0.483,
+                0.807
+            ],
+            "angle": 0,
+            "content": "There are two key findings here: 1) Parameter Efficiency: VAFlow consistently outperforms flow-based baselines (Frieren, MMAudio) at comparable sizes (e.g., 142M vs. 159M, 1.05B vs. 1.03B) across both metrics. 2) Scaling Capability: Previous flow-based studies have struggled to show consistent scalability [8, 43], where larger models fail to achieve better metrics consistently. To the best of our knowledge, VAFlow is the first flow-based V2A framework to demonstrate stable scaling properties. As the parameter size increases, the models achieve better FD and IS scores. Interestingly, larger models tend to improve IS scores more significantly than FD scores, suggesting that as the VAFlow model scales, the quality of generated audio gradually saturates, and the model becomes more capable of deriving diverse audio from different video-sampled latents."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.091,
+                0.818,
+                0.279,
+                0.833
+            ],
+            "angle": 0,
+            "content": "4.4. Benchmark Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.84,
+                0.484,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Quantitative evaluations are presented in Table 1, where VAFlow demonstrates state-of-the-art performance across key audio generation metrics: \\(\\mathrm{FD}_{\\text {pas. }}\\), \\(\\mathrm{FD}_{\\text {mel. }}\\), \\(\\mathrm{IS}_{\\text {pas. }}\\), \\(\\mathrm{KL}_{\\text {pas. }}\\), and \\(\\mathrm{KL}_{\\text {mel. }}\\). For audio-visual synchronization,"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.563,
+                0.091,
+                0.86,
+                0.262
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.513,
+                0.274,
+                0.908,
+                0.346
+            ],
+            "angle": 0,
+            "content": "Figure 4. Performance of flow-based baselines and VAFlow variants in scale. VAFlow models (142M, 1.05B) consistently outperform comparable-scale baselines (Frieren-159M, MMAudio-1.03B), with stable scaling property indicated by progressive IS gains (diversity) and saturating FD improvements (quality)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.37,
+                0.907,
+                0.673
+            ],
+            "angle": 0,
+            "content": "VAFlow achieves \\(96.3\\%\\) accuracy (Acc.), comparable to Frieren \\((97.0\\%)\\). In semantic relevance, it attains an IB-VA score of 28.6, second only to TiVA's 30.3. Notably, VAFlow maintains a simple and straightforward model design without relying on complex video conditioning modules [8, 41, 46], carefully designed noise schedules [9], or additional guidance modules [25, 45]. Its performance across various metrics, achieved with a standard Flow Matching training approach and basic ODE solvers, underscores the simplicity and effectiveness of the direct video-to-audio mapping design. When compared to text-enhanced baselines (gray rows), VAFlow maintains advantages in \\(\\mathrm{FD}_{\\text{pas.}}\\), \\(\\mathrm{FD}_{\\text{mel.}}\\), \\(\\mathrm{IS}_{\\text{pas.}}\\), and synchronization Acc., though text-augmented models exhibit superior audio semantic metrics (KL distances and IB-VA). This aligns with expectations since text conditioning explicitly constrains semantic space. We note that Seeing&Hearing's elevated IB-VA score (36.7) stems from classifier guidance using the same ImageBind model, which however compromises generation quality and synchronization performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.673,
+                0.909,
+                0.825
+            ],
+            "angle": 0,
+            "content": "We further conduct human evaluation with sampled 50 videos from the test set to compare our method with the flow-based baseline, Frieren. The generated results from both models were randomized, and 10 experts were invited to score the models based on three criteria: sound quality (Quality), audio-visual semantic consistency (Semantic) and audio-visual synchronization (Sync.). The scoring was done by selecting a winner or declaring a tie for each comparison. The results, summarized in Table 2, show that our model outperforms the baseline in all three dimensions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.833,
+                0.686,
+                0.85
+            ],
+            "angle": 0,
+            "content": "4.5. Ablative Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.856,
+                0.908,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Visual Features. Three visual feature types were experimented with: semantic features (CLIP embeddings extracted per frame at 10 fps), temporally-enhanced semantic"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.52,
+                0.957
+            ],
+            "angle": 0,
+            "content": "11782"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.095,
+                0.089,
+                0.91,
+                0.381
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Paradigm</td><td rowspan=\"2\">Method</td><td rowspan=\"2\">Vis.Feat.</td><td colspan=\"5\">Quality</td><td rowspan=\"2\">Sync. Acc. ↑</td><td rowspan=\"2\">Semantic IB-VA ↑</td></tr><tr><td>\\(FD_{pas.}\\)↓</td><td>\\(FD_{mel.}\\)↓</td><td>\\(IS_{mel.}\\)↑</td><td>\\(KL_{pas.}\\)↓</td><td>\\(KL_{mel.}\\)↓</td></tr><tr><td rowspan=\"2\">AR</td><td>SpecVQGAN [15]</td><td>RGB,Opti.</td><td>342.4</td><td>18.15</td><td>20.5</td><td>3.10</td><td>6.84</td><td>55.9</td><td>14.1</td></tr><tr><td>Im2Wav [34]</td><td>CLIP</td><td>252.2</td><td>13.03</td><td>35.8</td><td>2.26</td><td>5.41</td><td>78.9</td><td>19.6</td></tr><tr><td rowspan=\"2\">Mask</td><td>VATT [24]</td><td>eva-CLIP+</td><td>154.1</td><td>7.07</td><td>64.3</td><td>1.34</td><td>3.46</td><td>84.8</td><td>25.0</td></tr><tr><td>VAB [37]</td><td>eva-CLIP</td><td>193.1</td><td>12.21</td><td>34.4</td><td>2.31</td><td>5.09</td><td>82.4</td><td>25.9</td></tr><tr><td rowspan=\"7\">Diffusion</td><td>MultiFoley [7]</td><td>CAVP</td><td>241.6</td><td>13.00</td><td>67.7</td><td>1.59</td><td>3.34</td><td>85.2</td><td>27.0</td></tr><tr><td>Seeing&amp;Hearing [45]</td><td>ImageBind+</td><td>251.5</td><td>12.6</td><td>30.1</td><td>2.65</td><td>6.37</td><td>54.2</td><td>36.7</td></tr><tr><td>Diff-Foley [25]</td><td>CAVP</td><td>512.1</td><td>11.45</td><td>50.4</td><td>2.91</td><td>6.18</td><td>88.0</td><td>20.7</td></tr><tr><td>LoVA [9]</td><td>CLIP</td><td>149.1</td><td>6.90</td><td>58.7</td><td>2.10</td><td>5.12</td><td>87.0</td><td>26.3</td></tr><tr><td>FoleyCrafter [46]</td><td>CLIP</td><td>134.2</td><td>7.40</td><td>56.6</td><td>2.29</td><td>5.65</td><td>83.5</td><td>27.8</td></tr><tr><td>TiVA [41]</td><td>CLIP</td><td>111.9</td><td>4.31</td><td>67.0</td><td>2.02</td><td>5.21</td><td>87.5</td><td>30.3</td></tr><tr><td>V2A-Mapper [40]</td><td>CLIP</td><td>94.9</td><td>4.21</td><td>64.9</td><td>2.46</td><td>6.06</td><td>78.6</td><td>22.2</td></tr><tr><td rowspan=\"5\">Flow</td><td>MMAudio-L [8]</td><td>CLIP+</td><td>159.2</td><td>9.38</td><td>45.4</td><td>1.95</td><td>3.84</td><td>77.5</td><td>33.9</td></tr><tr><td>Frieren [43]</td><td>CAVP</td><td>259.9</td><td>6.70</td><td>52.4</td><td>2.92</td><td>5.64</td><td>97.0</td><td>23.0</td></tr><tr><td>VAFlow (ours)</td><td>CLIP</td><td>87.7</td><td>3.86</td><td>73.6</td><td>1.91</td><td>4.81</td><td>88.1</td><td>28.6</td></tr><tr><td>VAFlow (ours)</td><td>CAVP</td><td>91.8</td><td>4.73</td><td>64.8</td><td>2.41</td><td>5.96</td><td>96.3</td><td>25.1</td></tr><tr><td>VAFlow (ours)</td><td>VidTok</td><td>128.5</td><td>6.17</td><td>48.1</td><td>3.64</td><td>8.94</td><td>49.9</td><td>12.9</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.09,
+                0.391,
+                0.907,
+                0.493
+            ],
+            "angle": 0,
+            "content": "Table 1. Quantitative video-to-audio results on the VGGSound test set. “Vis. Feat.” indicates the employed visual features, with a “+” denoting the inclusion of auxiliary visual information (e.g., visual captions via LLama-2 [39] or Qwen-VL [2] in VATT and Seeing&Hearing, or SyncFormer [16] features in MMAudio). All results are obtained from official code or released audio outputs. The MultiFoley results here is an 8k subset filtered by ImageBind from the authors. Rows in gray denote methods that incorporate extra text data (e.g., video captions); for fairness, these results are provided for reference only and are not directly compared with standard video-to-audio models. The best score is bolded, and the second-best is underlined. Our three VAFlow variants, differing in visual features, achieve definitive gains in audio quality (FD, IS, KL) while delivering near state-of-the-art performance for synchronization (Acc.) and correlation (IB-VA)."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.098,
+                0.517,
+                0.478,
+                0.572
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Comparison</td><td>Quality</td><td>Win rate(%) Sync.</td><td>Semantic</td></tr><tr><td>Ours vs. Frieren</td><td>66.00(±4.18)</td><td>60.44(±5.72)</td><td>60.89(±4.48)</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.09,
+                0.583,
+                0.483,
+                0.639
+            ],
+            "angle": 0,
+            "content": "Table 2. Human evaluation results, with comparison between our model and the flow-based baseline Frieren across three criteria: sound quality, audio-visual synchronization, and audio-visual semantic consistency. Ours outperforms the baseline in all criteria."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.089,
+                0.674,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "features (CAVP features via video-audio contrastive learning at 4 fps), reconstruction features (VidTok tokenized latents at 15 fps). Utilizing the same model size and training process for these features, we developed three variants: VAFlow-CLIP, VAFlow-CAVP, and VAFlow-VidTok, with results presented in the last three rows of Table 1. The results indicate that the widely-used CLIP features overall perform the best, yielding superior audio quality and semantic relevance. The temporally enhanced CAVP features, on the other hand, demonstrate the best video-audio synchronization but slightly lag in other metrics compared to VAFlow-CLIP. Models based on reconstruction features exhibit the poorest performance across all evaluated aspects. These outcomes suggest two key insights: 1) The type of video features directly influences VAFlow's performance,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.52,
+                0.907,
+                0.641
+            ],
+            "angle": 0,
+            "content": "with different feature types leading to models with distinct performance preferences (either better semantic performance or synchronization); 2) Modality alignment matters more than information preservation. Although reconstruction features retain the most comprehensive video information among the three, their lack of modality alignment hampers the V2A model's ability to learn effective cross-modal generative capabilities."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.642,
+                0.909,
+                0.748
+            ],
+            "angle": 0,
+            "content": "Flow Matching Source Distributions. We evaluate VAFlow under two source distributions: video latent priors from alignment VAE and standard Gaussian noise. For video prior, we further test variants with/without joint training of the alignment VAE encoder (Method 3.3). All models share same training budgets (300K steps for non-joint training variants; 170K+130K for joint training)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.749,
+                0.909,
+                0.87
+            ],
+            "angle": 0,
+            "content": "Results in Table 3 demonstrate that video latent priors trained solely under alignment VAE objective in Equation 7 fail to surpass Gaussian noise baselines. Only when coupled with joint training—adapting the prior to flow-matching tasks—do video latent priors outperform Gaussian prior. This highlights the necessity of co-optimizing prior distributions for flow-based generation rather than relying on standalone pretraining on reconstruction task."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.871,
+                0.908,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Conditioning Mechanisms and Guidance. Prior work highlights the critical role of condition and guidance mech"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.52,
+                0.957
+            ],
+            "angle": 0,
+            "content": "11783"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.108,
+                0.089,
+                0.47,
+                0.158
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>V-Prior</td><td>Tuned</td><td>\\(FD_{pas.}\\)↓</td><td>\\(IS_{mel.}\\)↑</td><td>Acc.↑</td><td>IB-VA ↑</td></tr><tr><td>X</td><td>-</td><td>102.6</td><td>67.6</td><td>87.9</td><td>27.9</td></tr><tr><td>✓</td><td>X</td><td>111.8</td><td>57.4</td><td>82.3</td><td>25.4</td></tr><tr><td>✓</td><td>✓</td><td>87.7</td><td>73.6</td><td>88.1</td><td>28.6</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.091,
+                0.168,
+                0.483,
+                0.239
+            ],
+            "angle": 0,
+            "content": "Table 3. Ablation study on flow-matching source distributions. 'V-Prior' denotes sampling starts from video latent priors (vs. Gaussian priors). 'Tuned' indicates joint optimization of the alignment VAE encoder. Results show that jointly tuned video priors outperform Gaussian priors as source distributions for V2A flow models."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.097,
+                0.269,
+                0.48,
+                0.351
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Cond.</td><td>CA</td><td>CFG</td><td>FDpas. ↓</td><td>ISmel. ↑</td><td>Acc. ↑</td><td>IB-VA ↑</td></tr><tr><td>X</td><td>X</td><td>-</td><td>135.9</td><td>41.4</td><td>31.7</td><td>8.5</td></tr><tr><td>X</td><td>✓</td><td>X</td><td>160.5</td><td>36.5</td><td>18.9</td><td>5.4</td></tr><tr><td>✓</td><td>✓</td><td>X</td><td>133.8</td><td>42.3</td><td>78.0</td><td>21.0</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>87.7</td><td>73.6</td><td>88.1</td><td>28.6</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.09,
+                0.362,
+                0.483,
+                0.446
+            ],
+            "angle": 0,
+            "content": "Table 4. Ablation study on conditioning mechanisms and guidance. 'Cond.' denotes explicit video conditioning during inference; 'CA' indicates the presence of cross-attention layers of different variants; 'CFG' refers to classifier-free guidance. Results demonstrate that optimal performance requires both explicit conditioning and CFG when solving VAFlow ODE."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.49,
+                0.483,
+                0.522
+            ],
+            "angle": 0,
+            "content": "anisms (e.g., classifier-free guidance, CFG) in enhancing model performance. We address two questions:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.53,
+                0.484,
+                0.786
+            ],
+            "angle": 0,
+            "content": "(1) Can VAFlow generate high-quality audio using only video priors, omitting explicit video conditioning during sampling? We experimented with two variants without explicit condition: No cross-attention (CA): Remove CA layers in DiT blocks and retrain the model (Table 4, Row 1). Zero-input CA: Retain CA layers but zero out video inputs during inference (Table 4, Row 2). 'No CA' variant exhibit comparable audio quality metrics (FD, IS) to the video-condition-visible variant (Table 4, Row 3: explicit video conditioning without CFG), yet suffer significant degradation in semantic relevance (IB-VA) and synchronization (Acc.). This indicates that while video priors partially preserve visual information, explicit video conditioning during sampling is essential for temporal and semantic alignment. We think that as the latent trajectory approaches the target audio distribution \\((t\\to 1)\\), initial visual cues in the prior degrade, necessitating direct video information."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.09,
+                0.795,
+                0.484,
+                0.902
+            ],
+            "angle": 0,
+            "content": "(2) Does explicit video conditioning during sampling still require CFG for optimal performance? Adding CFG to the video-condition-visible variant (Row 4 in Table 4) further improves all metrics, demonstrating VAFlow's compatibility with established guidance techniques. This underscores the dual necessity of explicit video conditioning and CFG for optimal V2A synthesis with VAFlow framework."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.09,
+                0.53,
+                0.484,
+                0.902
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.529,
+                0.089,
+                0.596,
+                0.099
+            ],
+            "angle": 0,
+            "content": "A case video:"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.53,
+                0.099,
+                0.888,
+                0.118
+            ],
+            "angle": 0,
+            "content": "At a seaside with rolling waves, a group of people are swimming and playing. A woman walks toward the camera, points to the right, and says something."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.517,
+                0.119,
+                0.905,
+                0.45
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.513,
+                0.459,
+                0.908,
+                0.515
+            ],
+            "angle": 0,
+            "content": "Figure 5. Mel-spectrograms of audio generated by VAFlow and the baseline for a complex seaside scene, including background waves and varying voices, demonstrating VAFlow's superior content understanding and synchronization with the video."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.545,
+                0.649,
+                0.558
+            ],
+            "angle": 0,
+            "content": "4.6. Case Studies"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.568,
+                0.907,
+                0.72
+            ],
+            "angle": 0,
+            "content": "Figure 5 illustrates the performance of VAFlow and the baselines on the complex scene video. The video depicts a seaside scene with various sounds include background waves and different voices associated with shifting characters (noisy laughter, woman speaking). We present the ground truth and the mel-spectrograms of the audio generated by different methods. Our results show that VAFlow not only accurately understands the video content and generates all the necessary sounds but also maintains synchronization with the visual timing."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.738,
+                0.783,
+                0.753
+            ],
+            "angle": 0,
+            "content": "5. Conclusion and Future Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.765,
+                0.907,
+                0.903
+            ],
+            "angle": 0,
+            "content": "In this paper, we introduce VAFlow, a novel framework for V2A task, which directly denoises from the video space to the audio space via flow matching. demonstrating superior performance compared to previous methods. Future work will explore VAFlow's potential in more diverse audio domains, such as speech and music. Expanding the range of tasks and datasets will enable VAFlow's scaling properties to shine even further, laying a more general-purpose the foundation model for content generation community."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "11784"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.092,
+                0.091,
+                0.251,
+                0.108
+            ],
+            "angle": 0,
+            "content": "Acknowledgement"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.091,
+                0.115,
+                0.484,
+                0.176
+            ],
+            "angle": 0,
+            "content": "This work is supported by the National Natural Science Foundation of China (No. 62276268), Public Computing Cloud, Renmin University of China and ZHI-TECH GROUP."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.092,
+                0.188,
+                0.188,
+                0.204
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.101,
+                0.214,
+                0.484,
+                0.255
+            ],
+            "angle": 0,
+            "content": "[1] Michael S. Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In ICLR, 2023. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.1,
+                0.256,
+                0.484,
+                0.326
+            ],
+            "angle": 0,
+            "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.101,
+                0.327,
+                0.483,
+                0.395
+            ],
+            "angle": 0,
+            "content": "[3] 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. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.101,
+                0.396,
+                0.483,
+                0.466
+            ],
+            "angle": 0,
+            "content": "[4] Li-Wen Chang, Wenlei Bao, Qi Hou, Chengquan Jiang, Ningxin Zheng, Yinmin Zhong, Xuanrun Zhang, Zuquan Song, Chengji Yao, Ziheng Jiang, et al. Flux: fast software-based communication overlap on gpus through kernel fusion. arXiv preprint arXiv:2406.06858, 2024. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.467,
+                0.483,
+                0.508
+            ],
+            "angle": 0,
+            "content": "[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, pages 721–725, 2020. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.509,
+                0.483,
+                0.564
+            ],
+            "angle": 0,
+            "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.565,
+                0.483,
+                0.62
+            ],
+            "angle": 0,
+            "content": "[7] Ziyang Chen, Prem Seetharaman, Bryan Russell, Oriol Nieto, David Bourgin, Andrew Owens, and Justin Salamon. Video-guided foley sound generation with multimodal controls. arXiv preprint arXiv:2411.17698, 2024. 2, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.621,
+                0.483,
+                0.677
+            ],
+            "angle": 0,
+            "content": "[8] Ho Kei Cheng, Masato Ishii, Akio Hayakawa, Takashi Shibuya, Alexander Schwing, and Yuki Mitsufuji. Taming multimodal joint training for high-quality video-to-audio synthesis. arXiv preprint arXiv:2412.15322, 2024. 2, 3, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.102,
+                0.678,
+                0.483,
+                0.718
+            ],
+            "angle": 0,
+            "content": "[9] Xin Cheng, Xihua Wang, Yihan Wu, Yuyue Wang, and Ruihua Song. Lova: Long-form video-to-audio generation. arXiv preprint arXiv:2409.15157, 2024. 1, 2, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.72,
+                0.483,
+                0.801
+            ],
+            "angle": 0,
+            "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633, 2024. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.803,
+                0.483,
+                0.844
+            ],
+            "angle": 0,
+            "content": "[11] Zach Evans, Julian D Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. Stable audio open. arXiv preprint arXiv:2407.14358, 2024. 3, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.846,
+                0.483,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[12] 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 CVPR, pages 15180-15190, 2023. 5"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.094,
+                0.214,
+                0.484,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.093,
+                0.906,
+                0.161
+            ],
+            "angle": 0,
+            "content": "[13] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.164,
+                0.905,
+                0.19
+            ],
+            "angle": 0,
+            "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.192,
+                0.905,
+                0.221
+            ],
+            "angle": 0,
+            "content": "[15] Vladimir iashin and Esa Rahtu. Taming visually guided sound generation. In BMVC, 2021. 1, 2, 5, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.222,
+                0.905,
+                0.263
+            ],
+            "angle": 0,
+            "content": "[16] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Synchformer: Efficient synchronization from sparse cues. In ICASSP, pages 5325-5329, 2024. 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.264,
+                0.905,
+                0.32
+            ],
+            "angle": 0,
+            "content": "[17] 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, 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.321,
+                0.905,
+                0.39
+            ],
+            "angle": 0,
+            "content": "[18] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.392,
+                0.905,
+                0.434
+            ],
+            "angle": 0,
+            "content": "[19] Khaled Koutini, Jan Schlüter, Hamid Eghbal-Zadeh, and Gerhard Widmer. Efficient training of audio transformers with patchout. arXiv preprint arXiv:2110.05069, 2021. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.435,
+                0.905,
+                0.476
+            ],
+            "angle": 0,
+            "content": "[20] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In ICLR, 2023. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.478,
+                0.905,
+                0.546
+            ],
+            "angle": 0,
+            "content": "[21] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. In International Conference on Machine Learning, pages 21450-21474, 2023. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.548,
+                0.905,
+                0.603
+            ],
+            "angle": 0,
+            "content": "[22] Qihao Liu, Xi Yin, Alan Yuille, Andrew Brown, and Mannat Singh. Flowing from words to pixels: A noise-free framework for cross-modality evolution. In CVPR, pages 2755-2765, 2025. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.605,
+                0.905,
+                0.646
+            ],
+            "angle": 0,
+            "content": "[23] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.647,
+                0.905,
+                0.703
+            ],
+            "angle": 0,
+            "content": "[24] Xiulong Liu, Kun Su, and Eli Shlizerman. Tell what you hear from what you see - video to audio generation through text. In Advances in Neural Information Processing Systems, pages 101337-101366, 2024. 1, 2, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.704,
+                0.905,
+                0.773
+            ],
+            "angle": 0,
+            "content": "[25] Simian Luo, Chuanhao Yan, Chenxu Hu, and Hang Zhao. Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models. In Advances in Neural Information Processing Systems, pages 48855-48876, 2023. 1, 2, 3, 5, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.775,
+                0.905,
+                0.844
+            ],
+            "angle": 0,
+            "content": "[26] Xinhao Mei, Varun Nagaraja, Gael Le Lan, Zhaoheng Ni, Ernie Chang, Yangyang Shi, and Vikas Chandra. Foleygen: Visually-guided audio generation. In 2024 IEEE 34th International Workshop on Machine Learning for Signal Processing, pages 1-6, 2024. 1, 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.845,
+                0.905,
+                0.9
+            ],
+            "angle": 0,
+            "content": "[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171, 2021. 2, 5"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.517,
+                0.093,
+                0.906,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "11785"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.092,
+                0.482,
+                0.147
+            ],
+            "angle": 0,
+            "content": "[28] Shen Nie, Fengqi Zhu, Chao Du, Tianyu Pang, Qian Liu, Guangtao Zeng, Min Lin, and Chongxuan Li. Scaling up masked diffusion models on text. arXiv preprint arXiv:2410.18514, 2024. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.15,
+                0.482,
+                0.205
+            ],
+            "angle": 0,
+            "content": "[29] Santiago Pascual, Chunghsin Yeh, Ioannis Tsiamas, and Joan Serra. Masked generative video-to-audio transformers with enhanced synchronicity. In ECCV, pages 247-264, 2024. 1, 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.208,
+                0.482,
+                0.235
+            ],
+            "angle": 0,
+            "content": "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.238,
+                0.482,
+                0.568
+            ],
+            "angle": 0,
+            "content": "[31] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2024. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.572,
+                0.482,
+                0.655
+            ],
+            "angle": 0,
+            "content": "[32] 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, 2021. 2, 3, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.657,
+                0.482,
+                0.711
+            ],
+            "angle": 0,
+            "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.093,
+                0.715,
+                0.482,
+                0.754
+            ],
+            "angle": 0,
+            "content": "[34] Roy Sheffer and Yossi Adi. I hear your true colors: Image guided audio generation. In ICASSP, pages 1-5, 2023. 1, 2, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.758,
+                0.482,
+                0.786
+            ],
+            "angle": 0,
+            "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.789,
+                0.482,
+                0.842
+            ],
+            "angle": 0,
+            "content": "[36] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.094,
+                0.846,
+                0.482,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[37] Kun Su, Xiulong Liu, and Eli Shlizerman. From vision to audio and beyond: A unified model for audio-visual representation and generation. In International Conference on Machine Learning, pages 46804–46822, 2024. 2, 7"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.093,
+                0.092,
+                0.482,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.092,
+                0.905,
+                0.134
+            ],
+            "angle": 0,
+            "content": "[38] Anni Tang, Tianyu He, Junliang Guo, Xinle Cheng, Li Song, and Jiang Bian. Vidtok: A versatile and open-source video tokenizer. arXiv preprint arXiv:2412.13061, 2024. 2, 3, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.136,
+                0.905,
+                0.205
+            ],
+            "angle": 0,
+            "content": "[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.518,
+                0.207,
+                0.905,
+                0.272
+            ],
+            "angle": 0,
+            "content": "[40] Heng Wang, Jianbo Ma, Santiago Pascual, Richard Cartwright, and Weidong Cai. V2a-mapper: A lightweight solution for vision-to-audio generation by connecting foundation models. In AAAI, pages 15492-15501, 2024. 1, 2, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.277,
+                0.905,
+                0.33
+            ],
+            "angle": 0,
+            "content": "[41] Xihua Wang, Yuyue Wang, Yihan Wu, Ruihua Song, Xu Tan, Zehua Chen, Hongteng Xu, and Guodong Sui. Tiva: Time-aligned video-to-audio generation. In ACM MM, pages 573-582, 2024. 1, 2, 3, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.333,
+                0.905,
+                0.388
+            ],
+            "angle": 0,
+            "content": "[42] Xiang Wang, Shiwei Zhang, Hangjie Yuan, Zhiwu Qing, Biao Gong, Yingya Zhang, Yujun Shen, Changxin Gao, and Nong Sang. A recipe for scaling up text-to-video generation with text-free videos. In CVPR, pages 6572-6582, 2024. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.39,
+                0.905,
+                0.47
+            ],
+            "angle": 0,
+            "content": "[43] Yongqi Wang, Wenxiang Guo, Rongjie Huang, Jiawei Huang, Zehan Wang, Fuming You, Ruiqi Li, and Zhou Zhao. Frieren: Efficient video-to-audio generation network with rectified flow matching. In Advances in Neural Information Processing Systems, pages 128118-128138, 2024. 1, 2, 3, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.474,
+                0.905,
+                0.515
+            ],
+            "angle": 0,
+            "content": "[44] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In ECCV, pages 378-394, 2024. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.517,
+                0.905,
+                0.572
+            ],
+            "angle": 0,
+            "content": "[45] Yazhou Xing, Yingqing He, Zeyue Tian, Xintao Wang, and Qifeng Chen. Seeing and hearing: Open-domain visual-audio generation with diffusion latent aligners. In CVPR, pages 7151-7161, 2024. 1, 2, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.574,
+                0.905,
+                0.641
+            ],
+            "angle": 0,
+            "content": "[46] Yiming Zhang, Yicheng Gu, Yanhong Zeng, Zhening Xing, Yuancheng Wang, Zhizheng Wu, and Kai Chen. Foley-crafter: Bring silent videos to life with lifelike and synchronized sounds. arXiv preprint arXiv:2407.01494, 2024. 2, 3, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.517,
+                0.644,
+                0.905,
+                0.698
+            ],
+            "angle": 0,
+            "content": "[47]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.2"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.517,
+                0.092,
+                0.905,
+                0.698
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.946,
+                0.52,
+                0.957
+            ],
+            "angle": 0,
+            "content": "11786"
+        }
+    ]
+]

2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/7812feed-641c-4918-a582-f934ca29c2a8_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68daea7515223e0114cf193f42366c9da7d466916cb1d8394e2a99954eb6c279
+size 3628600

2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/full.md

@@ -0,0 +1,295 @@
+# VAFlow: Video-to-Audio Generation with Cross-Modality Flow Matching
+
+Xihua Wang $^{1}$  Xin Cheng $^{1}$  Yuyue Wang $^{1}$  Ruihua Song $^{1*}$  Yunfeng Wang $^{2}$ $^{1}$ Gaoling School of Artificial Intelligence, Renmin University of China  
+xihuaw@ruc.edu.cn, songruihua.bloon@outlook.com
+
+# Abstract
+
+Video-to-audio (V2A) generation aims to synthesize temporally aligned, realistic sounds for silent videos, a critical capability for immersive multimedia applications. Current V2A methods, predominantly based on diffusion or flow models, rely on suboptimal noise-to-audio paradigms that entangle cross-modal mappings with stochastic priors, resulting in inefficient training and convoluted transport paths. We propose VAFlow, a novel flow-based framework that directly models the video-to-audio transformation, eliminating reliance on noise priors. To address modality discrepancies, we employ an alignment variational autoencoder that compresses heterogeneous video features into audio-aligned latent spaces while preserving spatiotemporal semantics. By retaining cross-attention mechanisms between video features and flow blocks, our architecture enables classifier-free guidance within video source-driven generation. Without external data or complex training tricks, VAFlow achieves state-of-the-art performance on VGGSound benchmark, surpassing even textaugmented models in audio fidelity, diversity, and distribution alignment. This work establishes a new paradigm for V2A generation with a direct and effective video-to-audio transformation via flow matching.
+
+# 1. Introduction
+
+Video-to-audio (V2A) generation plays a vital role in multimedia content generation, including Foley for films and AI-generated silent videos. In recent years, this task has attracted increasing attention within the generative community. Early V2A approaches [15, 24, 26, 29, 34] adopted language modeling strategies by discretizing audio into tokens, and used parallel mask-prediction models (BERT-like) or token-by-token autoregressive models (GPT-like) borrowed from NLP community as shown in Figure 1. However, these methods, reliant on discrete tokens, incur inherent information losses by the discretization process.
+
+![](images/00b726f5ccb3264f0217ff01f4838780de47120879ff2588490bb01d049831e2.jpg)  
+Figure 1. Four different V2A paradigms, along with state-of-the-art performance under each. The first row shows discrete token-based approaches (mask-prediction and autoregressive based generation), which are limited by information loss during tokenization. The second row presents methods generating continuous data. Previous diffusion and flow methods (bottom-left) model the noise-to-audio mapping conditioning on video. We find this approach suboptimal and demonstrate that directly modeling the video-to-audio transformation (bottom-right) yields better results.
+
+Recent advancements in Diffusion and Flow Matching (FM) models in image [4, 10, 33] and video [3, 31] generation have spurred similar approaches for V2A [9, 25, 41, 43, 45]. Generally, these methods model V2A as a conditional generation task, transforming Gaussian noise conditioned on video into audio, as shown in Figure 1 (bottom-left). The generation process typically involves two steps: (1) sampling a random noise from Gaussian prior and (2) using this noise and video condition as input to iteratively denoise the noise into audio. This approach avoids discrete audio tokens, directly fitting the continuous distribution of audio data, and currently holds state-of-the-art results.
+
+However, existing diffusion/FM-based V2A methods focus solely on the second step, i.e., denoising Gaussian noise into audio, assuming the first step is trivial. Even state-of-the-art methods [40, 43] begin with Gaussian noise, but we find that this approach is not optimal. Unlike text-to
+
+image or text-to-video tasks, where textual conditions are abstract and highly defined by human, the video conditions in V2A are more complex and require both spatial and temporal understanding. The generation goal also demands a high degree of alignment between the generated audio and the video conditions, not just semantically, but also temporally. There are two inherent difficulties: (1) different conditions on the same noise require the model to generate different audio outputs, demanding advanced condition processing [8, 41, 46], and (2) noise variations under the same condition lead to unstable generation results, prompting research into methods for noise control [44], such as finding a "golden noise" for consistent results [47].
+
+In contrast, we propose a new perspective by shifting focus to the first step of the generation process. We question: If there are better priors than Gaussian noise? Given that many diffusion and FM models dedicate significant effort to integrating video conditions into the denoising process for better generation performance, we further question: If V2A can work better when directly transforming the video distribution to audio rather than from noise? Fortunately, both answers are affirmative.
+
+We introduce VAFlow, a novel framework for V2A generation, which directly denoises from the video space to the audio space. To implement this video-to-audio paradigm, we adopt the FM framework, as it theoretically supports any source distribution. There are three technical challenges for this implementation: (1) aligning source and target distributions: video and audio feature spaces typically have different time and spatial resolutions, yet FM requires the source and target distributions to be aligned (the same shape). (2) uncertainty of the feature type: identifying the optimal video representations/feature types as the source distribution. (3) conditioning mechanism: exploring additional video information integration ways beyond encoding it in the source distribution. In VAFlow, we first employ an alignment VAE to adjust video features to match the audio latent space. It then explores three different video feature types (semantic CLIP [32], spatiotemporal CAVP [25], reconstruction-focused VidTok [38]) to systematically investigate source video information. Experiments show that using video features both as a source distribution and as conditioning information yields the best results for VAFlow.
+
+Moreover, the recent success of large language models (LLMs) [13, 28] has highlighted the effectiveness of scaling laws, and various generative works consequently explore similar scaling properties in each domain [30, 42]. However, existing V2A models have not demonstrated clear scaling trends, where increasing model size does not consistently achieve improvement [8, 43]. In contrast, our experiments show that VAFlow exhibits clear scaling property, with model size increases leading to consistent improvements in generation quality, further confirming its potential.
+
+Without elaborate training tricks or external data (e.g., text-audio pairs), VAFlow achieves state-of-the-art results on standard benchmarks (VGGSound), outperforming even text-augmented models in generation fidelity, diversity, and cross-modality alignment with input video. Playable audi-los generated with VAFlow can be accessed at the demo page: https://vaflow.github.io/demo.
+
+# 2. Related Work
+
+# 2.1. Flow Matching and Diffusion Models
+
+Diffusion models [17, 27, 35] reconstruct data distributions by iteratively denoising Gaussian-sampled latents, governmented by various designed schedulers achieving state-of-the-art performance in multimodal generation tasks. Unlike diffusions, flow-based models [1, 20, 22, 23] establish deterministic mappings between source (e.g., Gaussian) and target distributions through invertible neural transformations, with recent flow matching optimizing vector fields through optimal transport theory, enabling direct learning of straight transport paths between source and target distributions, thereby reducing training instability and inference steps. While both paradigms typically rely on Gaussian priors as their source distribution, existing video-to-audio systems on both paradigms model noise-to-audio generation conditioned on video features. In contrast, our approach directly aligns video latent spaces with target audio distributions via linear transformations under a flow matching framework, bypassing iterative noise modeling, preserving efficiency, and ensuring precise frame-to-frame alignment.
+
+# 2.2. Video-to-Audio Generation
+
+Framework. Recent advances in generative models have driven significant progress in video-to-audio (V2A) synthesis. Current V2A methods fall into three main paradigms: (1) Autoregressive (AR) models [15, 26, 34] discretize audio into token (codec) sequences and utilize video features as prefixes to model unified video-audio sequences. While straightforward, their token-by-token generation process results in quadratic inference time scaling, limiting efficiency for long audio sequences. (2) Mask-based generation paradigms [24, 29, 37] also employ discrete tokens but decode audio in parallel by predicting masked tokens. Video features serve as the condition for mask-to-audio prediction. However, both AR and mask-based methods suffer from an inherent upper bound of fidelity due to lossy compression from audio discretization. (3) Flow/diffusion-based methods [7-9, 25, 40, 41, 43, 45, 46] leverage continuous representations and currently achieve state-of-the-art V2A performance [8, 40, 43]. These techniques iteratively transform noise into target audio distributions while conditioning on video features to steer the denoising trajectory.
+
+Video Conditional Modeling. Unlike typical conditioned
+
+generation tasks like text-to-image [33], V2A involves fully aligned sequential correspondences between video and audio. Existing methods address this through various strategies: some AR-based methods investigate various attention mechanisms and masking schemes for temporally sequential alignment [26]; some mask-based approaches integrate hybrid ways to fuse video features into the mask-prediction process to synchronize audio generation [29]; and some flow/diffusion-based methods augment their pipelines with auxiliary modules (e.g., loudness or audio layout predictors [41, 46]) to enforce temporal consistency. In contrast, our approach builds on flow-based techniques by discarding the conventional noise-to-audio paradigm and complex video-conditioning designs. Instead, we propose a more intuitive method that directly transforms the video temporal space to the audio temporal space via a flow transport.
+
+# 3. Method:
+
+# 3.1. Preliminary: Flow Matching
+
+Let  $x_0 \sim p(x_0)$  and  $x_1 \sim p(x_1)$  denote samples from source and target distributions, respectively, with  $c$  representing an optional conditioning signal. Flow Matching (FM) [20, 23] establishes an inter-distribution transport process governed by the ordinary differential equation (ODE):
+
+$$
+d x (t) = v _ {\theta} \left(x _ {t}, t, c\right) d t, \tag {1}
+$$
+
+where  $t \in [0,1]$  denotes the continuous time step, and  $v_{\theta}$  represents a neural network-parameterized velocity field. The condition  $c$  is incorporated for conditional flow implementations. Optimal transport flows enforce velocity alignment with the direction  $(x_{1} - x_{0})$  of the linear path pointing from  $x_{0}$  to  $x_{1}$  through regression:
+
+$$
+\mathcal {L} (\theta) = \mathbb {E} _ {t, x _ {0}, x _ {1}, c} \| (x _ {1} - x _ {0}) - v _ {\theta} (x _ {t}, t, c) \| ^ {2}, \tag {2}
+$$
+
+where  $x_{t} = (1 - t)x_{0} + tx_{1}$  constitutes the linear interpolation between  $x_{0}$  and  $x_{1}$ . The learned velocity field enables target generation by solving the ODE in Equation 1.
+
+Previous video-to-audio flow approaches [8, 43] (and related diffusion methods [25, 41]) use a Gaussian prior  $p(x_0) = \mathcal{N}(0,1)$  for the source distribution, with audio as  $p(x_{1})$  and video as conditioning input  $c$ . During training, random noise  $x_0$  is paired with ground-truth audio  $x_{1}$ , prompting  $v_{\theta}$  to model the paths from noise to audio under visual conditions.
+
+This work challenges the Gaussian prior by proposing a direct video-to-audio transport. Theoretically, FM permits arbitrary source distributions, motivating our formulation:
+
+$$
+d x (t) = v _ {\theta} \left(x _ {t}, t\right) d t, \tag {3}
+$$
+
+where  $x_{t} = (1 - t)x_{v} + tx_{a}$  linearly interpolates between paired video and audio samples  $x_{v},x_{a}\sim p(v,a)$
+
+# 3.2. VAFlow Architecture
+
+Model Overview. As illustrated in Figure 2, the input video features  $x_{v}$  are extracted via a visual encoder, while the target audio waveform can be obtained from latent representations  $x_{a}$  through WaveVAE. VAFlow attempts to establish FM between the video feature space  $x_{v}$  and audio latent space  $x_{a}$ , enabling direct audio generation from video features via ODE solving instead of Gaussian noise initialization. However, since FM requires consistent dimensional shapes between source and target distributions (see Equation 3), the inherent spatiotemporal resolution mismatch between  $x_{v}$  and  $x_{a}$  necessitates shape alignment. To address this, we employ a 1D VAE (namely alignment VAE) that compresses interpolated  $x_{v}$  into aligned features  $x_{v}^{\mathrm{AE}}$  matching  $x_{a}$ 's dimensions.
+
+A diffusion transformer (DiT) subsequently works as a velocity field estimator to capture the transport dynamics between  $x_{v}^{\mathrm{AE}}$  and  $x_{a}$ . This enables continuous trajectory sampling from  $x_{v}^{\mathrm{AE}}$  to  $x_{a}$  through ODE integration. The sampled latent  $\hat{x}_{a}$  is finally decoded by WaveVAE to produce the output audio. The following contents detail: the visual/audio representations  $(x_{v}, x_{v}^{\mathrm{AE}}, x_{a})$ , and the velocity field estimator in VAFlow.
+
+Video and Audio Representations. For an input video-audio pair  $(v, a)$ , the audio waveform  $a$  is compressed by a pretrained 1D WaveVAE's [11] encoder into  $x_{a} \in \mathbb{R}^{T_{a} \times D_{a}}$ . This latent  $x_{a}$  serves as the target for flow generation and is later decoded to reconstruct the waveform.
+
+For video representation, a pretrained visual encoder encodes  $v$  into  $x_v \in \mathbb{R}^{T_v \times D_v}$  (e.g., CLIP [32], CAVP [25], or VidTok [38] as encoder). Given the discrepancies (e.g.,  $x_a \in \mathbb{R}^{215 \times 64}$  versus  $x_v \in \mathbb{R}^{100 \times 768}$  for a 10-second clip), we propose an alignment VAE  $\epsilon_{\phi}$  to further compress and align video latents with audio latents' space while preserving essential visual information. Specifically, we utilize a 1D convolutional encoder-decoder as  $\epsilon_{\phi}$  to interpolate and project  $x_v$  into a variational space  $x_v^{\mathrm{AE}} \in \mathbb{R}^{T_a \times D_a}$ , matching the shape of  $x_a$ . The encoder of  $\epsilon_{\phi}$  predicts the parameters  $\mu_{x_v^{\mathrm{AE}}}, \sigma_{x_v^{\mathrm{AE}}} = E_{\epsilon_{\phi}}(x_v)$  for a Gaussian distribution, from which  $x_v^{\mathrm{AE}} \sim \mathcal{N}(\mu_{x_v^{\mathrm{AE}}}, \sigma_{x_v^{\mathrm{AE}}})$  is sampled. The decoder reconstructs  $x_v$  as  $\hat{x}_v = D_{\epsilon_{\phi}}(x_v^{\mathrm{AE}})$ . This process ensures dimensional alignment with audio latents while retaining critical spatiotemporal features of the original video representation.
+
+Velocity Estimator. Built upon the diffusion transformer (DiT) architecture [11], the velocity estimator processes two key inputs: time embeddings of  $t$  and the interpolated state  $x_{t} = (1 - t)x_{v}^{\mathrm{AE}} + tx_{a}$ , following the Equation 3. The time embedding is concatenated with the audio sequence to form an input tensor  $\in \mathbb{R}^{(T_a + 1)\times D_a}$ , which is processed through stacked transformer blocks. The first  $T_{a}$  dimensions of the output sequence constitute the pre
+
+![](images/ff1ff64fb37921e45705859648878aee89608f3208a30432c279d825228be95d.jpg)  
+Figure 2. VAFlow Architecture. The figure illustrates VAFow in the inference manner. A visual encoder initially extracts temporal features  $x_{v} \in \mathbb{R}^{T_{v} \times D_{v}}$  from variable-frame-rate video inputs. Then, a 1D alignment VAE (Variational Autoencoder)  $\epsilon_{\phi}$  compresses interpolated  $x_{v}$  into audio-aligned latent features  $x_{v}^{\mathrm{AE}}$  with matched dimensions  $(T_{v}^{\mathrm{AE}} \times D_{v}^{\mathrm{AE}} = T_{a} \times D_{a})$ . Next, a diffusion transformer  $(v_{\theta})$  predicts the velocity field for the transport  $(x_{a} - x_{v}^{\mathrm{AE}})$ , and by solving the corresponding ODE (see Equation 3), an estimated  $\hat{x}_{a}$  is obtained. Finally, the WaveVAE decoder converts  $\hat{x}_{a}$  into the waveform representing the predicted audio. Note that during diffusion transformer prediction,  $x_{v}$  can optionally be incorporated via cross-attention, detail discussion is conducted in the experiment Section 4.5.
+
+dicted velocity  $v_{\theta}(x_t, t)$ , with 1D Rotary Position Embeddings (RoPE [36]) to preserve positional awareness in attention layers. Notably, while the interpolated state  $x_t$  implicitly encodes visual information through  $x_v^{\mathrm{AE}}$ , this signal diminishes as  $t \to 1$ . To address this limitation, we retain cross-attention layers in the transformer blocks that persistently integrate the original video features  $x_v$ . Further studies (Section 4.5) demonstrate that this architectural choice is critical to enhance visual perception throughout the flow trajectory and improve audio generation quality.
+
+# 3.3. VAFlow Training and Inference
+
+We conduct three-stage training: (1) pretrain the alignment VAE, (2) train the VAFlow estimator with frozen alignment VAE, and (3) jointly fine-tune both alignment VAE and estimator. We detail training and inference as follows:
+
+Alignment VAE Training. Given encoded visual features  $x_{v} \in \mathbb{R}^{T_{v} \times D_{v}}$ , the alignment VAE encoder predicts distribution parameters  $(\mu_{x_{v}^{\mathrm{AE}}}, \sigma_{x_{v}^{\mathrm{AE}}})$ , while the decoder reconstructs  $\hat{x}_{v}$  from a sampled  $x_{v}^{\mathrm{AE}} \sim \mathcal{N}(\mu_{x_{v}^{\mathrm{AE}}}, \sigma_{x_{v}^{\mathrm{AE}}})$ . The model optimizes three objectives to preserve visual semantics while regularizing  $x_{v}^{\mathrm{AE}}$  within  $\mathbb{R}^{T_{a} \times D_{a}}$  Gaussian space:
+
+- KL Regularization: enforces Gaussian latent priors via:
+
+$$
+\mathcal {L} _ {\mathrm {K L}} = D _ {\mathrm {K L}} \left(\mathcal {N} \left(\mu_ {x _ {0} ^ {\mathrm {A E}}}, \sigma_ {x _ {0} ^ {\mathrm {A E}}}\right) \| \mathcal {N} (0, I)\right). \tag {4}
+$$
+
+- Reconstruction Loss: ensures feature-level consistency:
+
+$$
+\mathcal {L} _ {\text {r e c}} = \left\| x _ {v} - \hat {x} _ {v} \right\| _ {2} ^ {2}. \tag {5}
+$$
+
+- Contrastive Alignment: strengthens feature preservation:
+
+$$
+\mathcal {L} _ {\text {c o n}} = - \sum_ {s \in \mathcal {S}} \log \frac {\exp \left(f _ {s} \left(\hat {x} _ {v}\right) \cdot f _ {s} \left(x _ {v}\right) / \tau\right)}{\sum_ {x ^ {\prime} \in \mathcal {B}} \exp \left(f _ {s} \left(\hat {x} _ {v}\right) \cdot f _ {s} \left(x ^ {\prime}\right) / \tau\right)}, \tag {6}
+$$
+
+where  $S = \{\mathrm{frame},\mathrm{clip},\mathrm{global}\}$  defines temporal granularities and  $\tau = 0.7$  serves as a temperature parameter [6].
+
+For scale  $s$ ,  $f_{s}(\cdot)$  extracts features via three ways: frame-level: single-frame sampling + mean pooling; clip-level:  $K = 8$ -adjacent-frame sampling + pooling; global-level: full temporal mean pooling at  $T_{v}$  dimension of feature  $x$ . Negative samples  $x'$  are drawn from training batch  $\mathcal{B}$ .
+
+The composite objective integrates these terms:
+
+$$
+\mathcal {L} _ {\mathrm {V A E}} = \lambda_ {\mathrm {K L}} \mathcal {L} _ {\mathrm {K L}} + \lambda_ {\mathrm {r e c}} \mathcal {L} _ {\mathrm {r e c}} + \lambda_ {\mathrm {c o n}} \mathcal {L} _ {\mathrm {c o n}}, \tag {7}
+$$
+
+with  $\lambda_{\mathrm{KL}} = 8\times 10^{-4}$ $\lambda_{\mathrm{rec}} = 1.0$  , and  $\lambda_{\mathrm{con}} = 1.0$
+
+Velocity Estimator Training Given encoded features  $x_{v}$ ,  $x_{v}^{\mathrm{AE}}$ , and  $x_{a}$ , we train the velocity estimator  $v_{\theta}$  to predict the linear transport direction  $x_{a} - x_{v}^{\mathrm{AE}}$  through:
+
+$$
+\mathcal {L} (\theta) = \mathbb {E} _ {t \sim \mathcal {U} (0, 1)} \left\| \left(x _ {a} - x _ {v} ^ {\mathrm {A E}}\right) - v _ {\theta} \left(x _ {t}, t, x _ {v}\right) \right\| _ {2} ^ {2}, \tag {8}
+$$
+
+where  $x_{t} = (1 - t)x_{v}^{\mathrm{AE}} + tx_{a}$ . For the conditioned version, visual features  $x_{v}$  are incorporated via cross-attention (with a  $10\%$  probability of zeroing out  $x_{v}$  during training). The vanilla version omits this conditioning, simplifying the input of estimator in Equation 8 to  $v_{\theta}(x_t,t)$ .
+
+Joint Turning After separate training of the alignment VAE and the velocity estimator, a joint optimization stage is performed. In this stage, the VAE encoder and  $v_{\theta}$  are co-trained (discarding the VAE decoder), while the WaveVAE and visual feature extractor remain frozen. For a pair  $(v, a)$ , the visual extractor and WaveVAE produce  $x_{v}$  and  $x_{a}$ . The VAE encoder compresses  $x_{v}$  to obtain  $x_{v}^{\mathrm{AE}} \sim q_{\phi}(x_{v}^{\mathrm{AE}}|x_{v})$  which together with  $x_{a}$  and a random  $t \sim \mathcal{U}(0,1)$  forms  $x_{t}$ . The joint objective integrates KL regularization in Equation 4 with velocity prediction in Equation 8:
+
+$$
+\mathcal {L} (\theta , \phi_ {\text {E n c}}) = \lambda_ {\mathrm {K L}} \mathcal {L} _ {\mathrm {K L}} + \mathbb {E} \| \left(x _ {a} - x _ {v} ^ {\mathrm {A E}}\right) - v _ {\theta} \left(x _ {t}, t, x _ {v}\right) \| _ {2} ^ {2}. \tag {9}
+$$
+
+Inference Upon training, VAFlow synthesizes audio  $a$  from an input video  $v$  through the following pipeline: The visual extractor first encodes  $v$  into a feature  $x_{v}$ , which is
+
+compressed by the alignment VAE encoder into distribution parameters  $(\mu_{x_v^{\mathrm{AE}}},\sigma_{x_v^{\mathrm{AE}}})$ . A latent sample  $x_{v}^{\mathrm{AE}}\sim \mathcal{N}(\mu_{x_v^{\mathrm{AE}}},\sigma_{x_v^{\mathrm{AE}}}^2)$  is drawn as the initial state  $(x_0)$  for solving the ODE in Equation 1 with learned estimator. The predicted  $\hat{x}_a(x_1)$  is then decoded into waveform via the frozen WaveVAE. During inference, the variational sampling of  $x_{v}^{\mathrm{AE}}$  introduces stochasticity - even with fixed ODE integration paths, distinct samples yield diverse outputs.
+
+For conditional velocity estimators, classifier-free guidance (CFG) [14] enhances synthesis quality by interpolating conditioned and unconditioned velocity predictions:
+
+$$
+v _ {\theta} ^ {\mathrm {C F G}} \left(x _ {t}, t, x _ {v}\right) = (1 + \gamma) v _ {\theta} \left(x _ {t}, t, x _ {v}\right) - \gamma v _ {\theta} \left(x _ {t}, t, \emptyset\right), \tag {10}
+$$
+
+where  $\gamma$  controls guidance strength and  $\varnothing$  denotes null (zero) conditioning.
+
+# 4. Experiments
+
+# 4.1. Experimental Setup
+
+Datasets. We employ two datasets: VGGSound [5] and AudioSet-V2A [25]. For VGGSound, we adhere to the original split, which comprises 186k training samples and 15k testing samples. AudioSet-V2A contains 390k videos that are exclusively used for training. During training, the alignment VAE is trained on a combined set of VGGSound training and AudioSet-V2A to yield more generalized visual feature compression. The subsequent velocity estimator training and joint tuning are conducted solely on the VGGSound training set. At inference, we evaluate on the VGGSound test set as in prior work.
+
+Implementation Details. For the estimator's DiT architecture, we adapted the DiT implementation from StableAudio-Open-1.0 [11] (SDA1.0) by eliminating duration encoding, applying global time encoding solely to time steps, and restricting the cross-attention layer's key/value inputs to video features. For the visual extractor, we use three variants: CLIP [32] (a widely adopted visual semantic extractor in V2A tasks), CAVP [25] (which captures both visual semantic and temporal features), and VidTok [38] (a visual tokenizer for reconstruction). We train three alignment VAEs with comparable parameter scales—one for each visual extractor. In the velocity estimator training stage, the alignment VAE remains frozen; in the joint training stage, both the estimator and the alignment VAE are trained. In all stages, the audio WaveVAE and the visual extractor are kept frozen. Notably, we do not perform any further rectified iterations on the estimator; all VAFlow variants are trained for a single rectified iteration (Rectified-1, RF-1). During sampling, we employ ODE solvers of various orders—first-order Euler, second-order Midpoint, and fifth-order dopri5—with differing sampling steps and classifier-free guidance scales.
+
+Metrics. We employ both objective and subjective evaluations across three dimensions: generation quality, audiovisual relevance, and audiovisual synchronization. For objective evaluation, generation quality is measured using FD score [21], KL divergence [15], and IS score [21]. In previous work, each metric may be computed with different feature extraction models (e.g., Melception, PasST $^1$ ). Audiovisual relevance is quantified by computing the cosine similarity between the embeddings of the input video and the generated audio using the ImageBind model [12]. For synchronization, we use the synchronization accuracy (Acc.) [25] reflects the alignment between the input video and generated audio.
+
+Baselines. Our method is comprehensively compared against all major generative paradigms, including autoregressive (AR), masked prediction (Mask), diffusion-based, and flow-based approaches, as detailed in Table 1. All baseline models were evaluated using either officially released code and models or directly downloaded audio outputs by their authors, with consistent experimental settings maintained across all evaluations on the same hardware platform. Implementation specifics include: (1) Im2Wav generates audio in 4s clips, truncated to 10s after three clips. (2) Diff-Foley here employs its double-guidance variant (demonstrating superior performance); (3) MMAudio-L here is the largest variant of MMAudio; and (4) Frieren here employs its final no-reflow configuration (achieving optimal Fréchet distance). For comprehensive evaluation, we also include text-augmented V2A models (denoted in gray rows) that leverage external textual data beyond benchmark audiovisual corpora during training or inference.
+
+# 4.2. Diffusion vs. Flow vs. VAFlow Models
+
+We examine framework-specific performance by training same VAFlow-CLIP-142M models under three paradigms, maintaining identical configurations: CLIP visual features, architecture parameters, initialization weights, and other hyperparameters. The frameworks include: Diffusion: Standard DDPM [27, 33] with cross-attention video conditioning. Flow: Standard continuous-time flow (Equation 2) using Gaussian prior. VAFlow: Our framework employing aligned VAE-compressed video features as source distribution. All frameworks use their respective optimal inference settings (DDPM: DPMSolver; Flows: dopri5). Joint tuning in VAFlow begins at 100K steps.
+
+Frechet Distance (FD) results in Figure 3 reveals: 1) Flow paradigms converge faster with lower FD than diffusion, demonstrating inherent optimization benefits. 2) VAFlow attains lower final FD than standard flow despite delayed early-stage convergence (<100K steps). This per
+
+![](images/ddbe06928bea183e718f5bdb4312d42104ea62742c41fc44089c0837a7187e51.jpg)  
+Figure 3. Comparison of different training frameworks. We examine the same model under different training paradigms by controlling training data, steps, and other hyperparameters. Results demonstrate that VAFlow converges to superior optima, validating the effectiveness of its direct video-to-audio transformation design and highlighting this paradigm advantages in V2A generation.
+
+formance attributes to VAFlow's explicit modeling of deterministic video-audio transport paths, avoiding the ambiguous noise-to-audio mappings in Gaussian-prior approaches. VAFlow learns optimal transport trajectories between video-audio distribution, bypassing error-prone intermediate transformations in Gaussian-noise-to-audio paths.
+
+# 4.3. VAFlow Scaling Analysis
+
+We investigate VAFlow's scaling capabilities by developing four model variants (142M, 264M, 527M, 1.05B parameters) through width/depth scaling of DiT blocks while maintaining fixed training protocols (data, steps, etc.). Evaluation metrics (FD, IS) versus parameter counts are visualized in Figure 4 (bubble sizes denote model scales).
+
+There are two key findings here: 1) Parameter Efficiency: VAFlow consistently outperforms flow-based baselines (Frieren, MMAudio) at comparable sizes (e.g., 142M vs. 159M, 1.05B vs. 1.03B) across both metrics. 2) Scaling Capability: Previous flow-based studies have struggled to show consistent scalability [8, 43], where larger models fail to achieve better metrics consistently. To the best of our knowledge, VAFlow is the first flow-based V2A framework to demonstrate stable scaling properties. As the parameter size increases, the models achieve better FD and IS scores. Interestingly, larger models tend to improve IS scores more significantly than FD scores, suggesting that as the VAFlow model scales, the quality of generated audio gradually saturates, and the model becomes more capable of deriving diverse audio from different video-sampled latents.
+
+# 4.4. Benchmark Results
+
+Quantitative evaluations are presented in Table 1, where VAFlow demonstrates state-of-the-art performance across key audio generation metrics:  $\mathrm{FD}_{\text {pas. }}$ ,  $\mathrm{FD}_{\text {mel. }}$ ,  $\mathrm{IS}_{\text {pas. }}$ ,  $\mathrm{KL}_{\text {pas. }}$ , and  $\mathrm{KL}_{\text {mel. }}$ . For audio-visual synchronization,
+
+![](images/f09fb5f06231f54fb4c79ca7f5346440e949c46f53aa2c0415e37e982158f883.jpg)  
+Figure 4. Performance of flow-based baselines and VAFlow variants in scale. VAFlow models (142M, 1.05B) consistently outperform comparable-scale baselines (Frieren-159M, MMAudio-1.03B), with stable scaling property indicated by progressive IS gains (diversity) and saturating FD improvements (quality).
+
+VAFlow achieves  $96.3\%$  accuracy (Acc.), comparable to Frieren  $(97.0\%)$ . In semantic relevance, it attains an IB-VA score of 28.6, second only to TiVA's 30.3. Notably, VAFlow maintains a simple and straightforward model design without relying on complex video conditioning modules [8, 41, 46], carefully designed noise schedules [9], or additional guidance modules [25, 45]. Its performance across various metrics, achieved with a standard Flow Matching training approach and basic ODE solvers, underscores the simplicity and effectiveness of the direct video-to-audio mapping design. When compared to text-enhanced baselines (gray rows), VAFlow maintains advantages in  $\mathrm{FD}_{\text{pas.}}$ ,  $\mathrm{FD}_{\text{mel.}}$ ,  $\mathrm{IS}_{\text{pas.}}$ , and synchronization Acc., though text-augmented models exhibit superior audio semantic metrics (KL distances and IB-VA). This aligns with expectations since text conditioning explicitly constrains semantic space. We note that Seeing&Hearing's elevated IB-VA score (36.7) stems from classifier guidance using the same ImageBind model, which however compromises generation quality and synchronization performance.
+
+We further conduct human evaluation with sampled 50 videos from the test set to compare our method with the flow-based baseline, Frieren. The generated results from both models were randomized, and 10 experts were invited to score the models based on three criteria: sound quality (Quality), audio-visual semantic consistency (Semantic) and audio-visual synchronization (Sync.). The scoring was done by selecting a winner or declaring a tie for each comparison. The results, summarized in Table 2, show that our model outperforms the baseline in all three dimensions.
+
+# 4.5. Ablative Analysis
+
+Visual Features. Three visual feature types were experimented with: semantic features (CLIP embeddings extracted per frame at 10 fps), temporally-enhanced semantic
+
+<table><tr><td rowspan="2">Paradigm</td><td rowspan="2">Method</td><td rowspan="2">Vis.Feat.</td><td colspan="5">Quality</td><td rowspan="2">Sync. Acc. ↑</td><td rowspan="2">Semantic IB-VA ↑</td></tr><tr><td>\(FD_{pas.}\)↓</td><td>\(FD_{mel.}\)↓</td><td>\(IS_{mel.}\)↑</td><td>\(KL_{pas.}\)↓</td><td>\(KL_{mel.}\)↓</td></tr><tr><td rowspan="2">AR</td><td>SpecVQGAN [15]</td><td>RGB,Opti.</td><td>342.4</td><td>18.15</td><td>20.5</td><td>3.10</td><td>6.84</td><td>55.9</td><td>14.1</td></tr><tr><td>Im2Wav [34]</td><td>CLIP</td><td>252.2</td><td>13.03</td><td>35.8</td><td>2.26</td><td>5.41</td><td>78.9</td><td>19.6</td></tr><tr><td rowspan="2">Mask</td><td>VATT [24]</td><td>eva-CLIP+</td><td>154.1</td><td>7.07</td><td>64.3</td><td>1.34</td><td>3.46</td><td>84.8</td><td>25.0</td></tr><tr><td>VAB [37]</td><td>eva-CLIP</td><td>193.1</td><td>12.21</td><td>34.4</td><td>2.31</td><td>5.09</td><td>82.4</td><td>25.9</td></tr><tr><td rowspan="7">Diffusion</td><td>MultiFoley [7]</td><td>CAVP</td><td>241.6</td><td>13.00</td><td>67.7</td><td>1.59</td><td>3.34</td><td>85.2</td><td>27.0</td></tr><tr><td>Seeing&amp;Hearing [45]</td><td>ImageBind+</td><td>251.5</td><td>12.6</td><td>30.1</td><td>2.65</td><td>6.37</td><td>54.2</td><td>36.7</td></tr><tr><td>Diff-Foley [25]</td><td>CAVP</td><td>512.1</td><td>11.45</td><td>50.4</td><td>2.91</td><td>6.18</td><td>88.0</td><td>20.7</td></tr><tr><td>LoVA [9]</td><td>CLIP</td><td>149.1</td><td>6.90</td><td>58.7</td><td>2.10</td><td>5.12</td><td>87.0</td><td>26.3</td></tr><tr><td>FoleyCrafter [46]</td><td>CLIP</td><td>134.2</td><td>7.40</td><td>56.6</td><td>2.29</td><td>5.65</td><td>83.5</td><td>27.8</td></tr><tr><td>TiVA [41]</td><td>CLIP</td><td>111.9</td><td>4.31</td><td>67.0</td><td>2.02</td><td>5.21</td><td>87.5</td><td>30.3</td></tr><tr><td>V2A-Mapper [40]</td><td>CLIP</td><td>94.9</td><td>4.21</td><td>64.9</td><td>2.46</td><td>6.06</td><td>78.6</td><td>22.2</td></tr><tr><td rowspan="5">Flow</td><td>MMAudio-L [8]</td><td>CLIP+</td><td>159.2</td><td>9.38</td><td>45.4</td><td>1.95</td><td>3.84</td><td>77.5</td><td>33.9</td></tr><tr><td>Frieren [43]</td><td>CAVP</td><td>259.9</td><td>6.70</td><td>52.4</td><td>2.92</td><td>5.64</td><td>97.0</td><td>23.0</td></tr><tr><td>VAFlow (ours)</td><td>CLIP</td><td>87.7</td><td>3.86</td><td>73.6</td><td>1.91</td><td>4.81</td><td>88.1</td><td>28.6</td></tr><tr><td>VAFlow (ours)</td><td>CAVP</td><td>91.8</td><td>4.73</td><td>64.8</td><td>2.41</td><td>5.96</td><td>96.3</td><td>25.1</td></tr><tr><td>VAFlow (ours)</td><td>VidTok</td><td>128.5</td><td>6.17</td><td>48.1</td><td>3.64</td><td>8.94</td><td>49.9</td><td>12.9</td></tr></table>
+
+Table 1. Quantitative video-to-audio results on the VGGSound test set. “Vis. Feat.” indicates the employed visual features, with a “+” denoting the inclusion of auxiliary visual information (e.g., visual captions via LLama-2 [39] or Qwen-VL [2] in VATT and Seeing&Hearing, or SyncFormer [16] features in MMAudio). All results are obtained from official code or released audio outputs. The MultiFoley results here is an 8k subset filtered by ImageBind from the authors. Rows in gray denote methods that incorporate extra text data (e.g., video captions); for fairness, these results are provided for reference only and are not directly compared with standard video-to-audio models. The best score is bolded, and the second-best is underlined. Our three VAFlow variants, differing in visual features, achieve definitive gains in audio quality (FD, IS, KL) while delivering near state-of-the-art performance for synchronization (Acc.) and correlation (IB-VA).  
+
+<table><tr><td>Comparison</td><td>Quality</td><td>Win rate(%) Sync.</td><td>Semantic</td></tr><tr><td>Ours vs. Frieren</td><td>66.00(±4.18)</td><td>60.44(±5.72)</td><td>60.89(±4.48)</td></tr></table>
+
+Table 2. Human evaluation results, with comparison between our model and the flow-based baseline Frieren across three criteria: sound quality, audio-visual synchronization, and audio-visual semantic consistency. Ours outperforms the baseline in all criteria.
+
+features (CAVP features via video-audio contrastive learning at 4 fps), reconstruction features (VidTok tokenized latents at 15 fps). Utilizing the same model size and training process for these features, we developed three variants: VAFlow-CLIP, VAFlow-CAVP, and VAFlow-VidTok, with results presented in the last three rows of Table 1. The results indicate that the widely-used CLIP features overall perform the best, yielding superior audio quality and semantic relevance. The temporally enhanced CAVP features, on the other hand, demonstrate the best video-audio synchronization but slightly lag in other metrics compared to VAFlow-CLIP. Models based on reconstruction features exhibit the poorest performance across all evaluated aspects. These outcomes suggest two key insights: 1) The type of video features directly influences VAFlow's performance,
+
+with different feature types leading to models with distinct performance preferences (either better semantic performance or synchronization); 2) Modality alignment matters more than information preservation. Although reconstruction features retain the most comprehensive video information among the three, their lack of modality alignment hampers the V2A model's ability to learn effective cross-modal generative capabilities.
+
+Flow Matching Source Distributions. We evaluate VAFlow under two source distributions: video latent priors from alignment VAE and standard Gaussian noise. For video prior, we further test variants with/without joint training of the alignment VAE encoder (Method 3.3). All models share same training budgets (300K steps for non-joint training variants; 170K+130K for joint training).
+
+Results in Table 3 demonstrate that video latent priors trained solely under alignment VAE objective in Equation 7 fail to surpass Gaussian noise baselines. Only when coupled with joint training—adapting the prior to flow-matching tasks—do video latent priors outperform Gaussian prior. This highlights the necessity of co-optimizing prior distributions for flow-based generation rather than relying on standalone pretraining on reconstruction task.
+
+Conditioning Mechanisms and Guidance. Prior work highlights the critical role of condition and guidance mech
+
+<table><tr><td>V-Prior</td><td>Tuned</td><td>\(FD_{pas.}\)↓</td><td>\(IS_{mel.}\)↑</td><td>Acc.↑</td><td>IB-VA ↑</td></tr><tr><td>X</td><td>-</td><td>102.6</td><td>67.6</td><td>87.9</td><td>27.9</td></tr><tr><td>✓</td><td>X</td><td>111.8</td><td>57.4</td><td>82.3</td><td>25.4</td></tr><tr><td>✓</td><td>✓</td><td>87.7</td><td>73.6</td><td>88.1</td><td>28.6</td></tr></table>
+
+Table 3. Ablation study on flow-matching source distributions. 'V-Prior' denotes sampling starts from video latent priors (vs. Gaussian priors). 'Tuned' indicates joint optimization of the alignment VAE encoder. Results show that jointly tuned video priors outperform Gaussian priors as source distributions for V2A flow models.  
+
+<table><tr><td>Cond.</td><td>CA</td><td>CFG</td><td>FDpas. ↓</td><td>ISmel. ↑</td><td>Acc. ↑</td><td>IB-VA ↑</td></tr><tr><td>X</td><td>X</td><td>-</td><td>135.9</td><td>41.4</td><td>31.7</td><td>8.5</td></tr><tr><td>X</td><td>✓</td><td>X</td><td>160.5</td><td>36.5</td><td>18.9</td><td>5.4</td></tr><tr><td>✓</td><td>✓</td><td>X</td><td>133.8</td><td>42.3</td><td>78.0</td><td>21.0</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>87.7</td><td>73.6</td><td>88.1</td><td>28.6</td></tr></table>
+
+Table 4. Ablation study on conditioning mechanisms and guidance. 'Cond.' denotes explicit video conditioning during inference; 'CA' indicates the presence of cross-attention layers of different variants; 'CFG' refers to classifier-free guidance. Results demonstrate that optimal performance requires both explicit conditioning and CFG when solving VAFlow ODE.
+
+anisms (e.g., classifier-free guidance, CFG) in enhancing model performance. We address two questions:
+
+(1) Can VAFlow generate high-quality audio using only video priors, omitting explicit video conditioning during sampling? We experimented with two variants without explicit condition: No cross-attention (CA): Remove CA layers in DiT blocks and retrain the model (Table 4, Row 1). Zero-input CA: Retain CA layers but zero out video inputs during inference (Table 4, Row 2). 'No CA' variant exhibit comparable audio quality metrics (FD, IS) to the video-condition-visible variant (Table 4, Row 3: explicit video conditioning without CFG), yet suffer significant degradation in semantic relevance (IB-VA) and synchronization (Acc.). This indicates that while video priors partially preserve visual information, explicit video conditioning during sampling is essential for temporal and semantic alignment. We think that as the latent trajectory approaches the target audio distribution  $(t\to 1)$ , initial visual cues in the prior degrade, necessitating direct video information.  
+(2) Does explicit video conditioning during sampling still require CFG for optimal performance? Adding CFG to the video-condition-visible variant (Row 4 in Table 4) further improves all metrics, demonstrating VAFlow's compatibility with established guidance techniques. This underscores the dual necessity of explicit video conditioning and CFG for optimal V2A synthesis with VAFlow framework.
+
+![](images/541834df9a135b3f673f2135fc672b6335d888f01e8ee58e170b788fb154322c.jpg)  
+A case video:  
+At a seaside with rolling waves, a group of people are swimming and playing. A woman walks toward the camera, points to the right, and says something.  
+Figure 5. Mel-spectrograms of audio generated by VAFlow and the baseline for a complex seaside scene, including background waves and varying voices, demonstrating VAFlow's superior content understanding and synchronization with the video.
+
+# 4.6. Case Studies
+
+Figure 5 illustrates the performance of VAFlow and the baselines on the complex scene video. The video depicts a seaside scene with various sounds include background waves and different voices associated with shifting characters (noisy laughter, woman speaking). We present the ground truth and the mel-spectrograms of the audio generated by different methods. Our results show that VAFlow not only accurately understands the video content and generates all the necessary sounds but also maintains synchronization with the visual timing.
+
+# 5. Conclusion and Future Work
+
+In this paper, we introduce VAFlow, a novel framework for V2A task, which directly denoises from the video space to the audio space via flow matching. demonstrating superior performance compared to previous methods. Future work will explore VAFlow's potential in more diverse audio domains, such as speech and music. Expanding the range of tasks and datasets will enable VAFlow's scaling properties to shine even further, laying a more general-purpose the foundation model for content generation community.
+
+# Acknowledgement
+
+This work is supported by the National Natural Science Foundation of China (No. 62276268), Public Computing Cloud, Renmin University of China and ZHI-TECH GROUP.
+
+# References
+
+[1] Michael S. Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In ICLR, 2023. 2  
+[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 7  
+[3] 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. 1  
+[4] Li-Wen Chang, Wenlei Bao, Qi Hou, Chengquan Jiang, Ningxin Zheng, Yinmin Zhong, Xuanrun Zhang, Zuquan Song, Chengji Yao, Ziheng Jiang, et al. Flux: fast software-based communication overlap on gpus through kernel fusion. arXiv preprint arXiv:2406.06858, 2024. 1  
+[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, pages 721–725, 2020. 5  
+[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 4  
+[7] Ziyang Chen, Prem Seetharaman, Bryan Russell, Oriol Nieto, David Bourgin, Andrew Owens, and Justin Salamon. Video-guided foley sound generation with multimodal controls. arXiv preprint arXiv:2411.17698, 2024. 2, 7  
+[8] Ho Kei Cheng, Masato Ishii, Akio Hayakawa, Takashi Shibuya, Alexander Schwing, and Yuki Mitsufuji. Taming multimodal joint training for high-quality video-to-audio synthesis. arXiv preprint arXiv:2412.15322, 2024. 2, 3, 6, 7  
+[9] Xin Cheng, Xihua Wang, Yihan Wu, Yuyue Wang, and Ruihua Song. Lova: Long-form video-to-audio generation. arXiv preprint arXiv:2409.15157, 2024. 1, 2, 6, 7  
+[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633, 2024. 1  
+[11] Zach Evans, Julian D Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. Stable audio open. arXiv preprint arXiv:2407.14358, 2024. 3, 5  
+[12] 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 CVPR, pages 15180-15190, 2023. 5
+
+[13] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2  
+[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5  
+[15] Vladimir iashin and Esa Rahtu. Taming visually guided sound generation. In BMVC, 2021. 1, 2, 5, 7  
+[16] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Synchformer: Efficient synchronization from sparse cues. In ICASSP, pages 5325-5329, 2024. 7  
+[17] 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, 2022. 2  
+[18] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 5  
+[19] Khaled Koutini, Jan Schlüter, Hamid Eghbal-Zadeh, and Gerhard Widmer. Efficient training of audio transformers with patchout. arXiv preprint arXiv:2110.05069, 2021. 5  
+[20] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In ICLR, 2023. 2, 3  
+[21] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. In International Conference on Machine Learning, pages 21450-21474, 2023. 5  
+[22] Qihao Liu, Xi Yin, Alan Yuille, Andrew Brown, and Mannat Singh. Flowing from words to pixels: A noise-free framework for cross-modality evolution. In CVPR, pages 2755-2765, 2025. 2  
+[23] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 2, 3  
+[24] Xiulong Liu, Kun Su, and Eli Shlizerman. Tell what you hear from what you see - video to audio generation through text. In Advances in Neural Information Processing Systems, pages 101337-101366, 2024. 1, 2, 7  
+[25] Simian Luo, Chuanhao Yan, Chenxu Hu, and Hang Zhao. Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models. In Advances in Neural Information Processing Systems, pages 48855-48876, 2023. 1, 2, 3, 5, 6, 7  
+[26] Xinhao Mei, Varun Nagaraja, Gael Le Lan, Zhaoheng Ni, Ernie Chang, Yangyang Shi, and Vikas Chandra. Foleygen: Visually-guided audio generation. In 2024 IEEE 34th International Workshop on Machine Learning for Signal Processing, pages 1-6, 2024. 1, 2, 3  
+[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171, 2021. 2, 5
+
+[28] Shen Nie, Fengqi Zhu, Chao Du, Tianyu Pang, Qian Liu, Guangtao Zeng, Min Lin, and Chongxuan Li. Scaling up masked diffusion models on text. arXiv preprint arXiv:2410.18514, 2024. 2  
+[29] Santiago Pascual, Chunghsin Yeh, Ioannis Tsiamas, and Joan Serra. Masked generative video-to-audio transformers with enhanced synchronicity. In ECCV, pages 247-264, 2024. 1, 2, 3  
+[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 2  
+[31] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2024. 1  
+[32] 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, 2021. 2, 3, 5  
+[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 5  
+[34] Roy Sheffer and Yossi Adi. I hear your true colors: Image guided audio generation. In ICASSP, pages 1-5, 2023. 1, 2, 7  
+[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2  
+[36] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4  
+[37] Kun Su, Xiulong Liu, and Eli Shlizerman. From vision to audio and beyond: A unified model for audio-visual representation and generation. In International Conference on Machine Learning, pages 46804–46822, 2024. 2, 7
+
+[38] Anni Tang, Tianyu He, Junliang Guo, Xinle Cheng, Li Song, and Jiang Bian. Vidtok: A versatile and open-source video tokenizer. arXiv preprint arXiv:2412.13061, 2024. 2, 3, 5  
+[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7  
+[40] Heng Wang, Jianbo Ma, Santiago Pascual, Richard Cartwright, and Weidong Cai. V2a-mapper: A lightweight solution for vision-to-audio generation by connecting foundation models. In AAAI, pages 15492-15501, 2024. 1, 2, 7  
+[41] Xihua Wang, Yuyue Wang, Yihan Wu, Ruihua Song, Xu Tan, Zehua Chen, Hongteng Xu, and Guodong Sui. Tiva: Time-aligned video-to-audio generation. In ACM MM, pages 573-582, 2024. 1, 2, 3, 6, 7  
+[42] Xiang Wang, Shiwei Zhang, Hangjie Yuan, Zhiwu Qing, Biao Gong, Yingya Zhang, Yujun Shen, Changxin Gao, and Nong Sang. A recipe for scaling up text-to-video generation with text-free videos. In CVPR, pages 6572-6582, 2024. 2  
+[43] Yongqi Wang, Wenxiang Guo, Rongjie Huang, Jiawei Huang, Zehan Wang, Fuming You, Ruiqi Li, and Zhou Zhao. Frieren: Efficient video-to-audio generation network with rectified flow matching. In Advances in Neural Information Processing Systems, pages 128118-128138, 2024. 1, 2, 3, 6, 7  
+[44] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In ECCV, pages 378-394, 2024. 2  
+[45] Yazhou Xing, Yingqing He, Zeyue Tian, Xintao Wang, and Qifeng Chen. Seeing and hearing: Open-domain visual-audio generation with diffusion latent aligners. In CVPR, pages 7151-7161, 2024. 1, 2, 6, 7  
+[46] Yiming Zhang, Yicheng Gu, Yanhong Zeng, Zhening Xing, Yuancheng Wang, Zhizheng Wu, and Kai Chen. Foley-crafter: Bring silent videos to life with lifelike and synchronized sounds. arXiv preprint arXiv:2407.01494, 2024. 2, 3, 6, 7  
+[47]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.2

2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:03bd79028fb9b67a086f910a95625ac525c7b72fdb4cece8ed2e54def3d9f3e5
+size 415117

2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/321d7adc-1768-447e-aad1-8d77c13702f1_content_list.json

@@ -0,0 +1,1640 @@
+[
+    {
+        "type": "text",
+        "text": "VAGUE: Visual Contexts Clarify Ambiguous Expressions",
+        "text_level": 1,
+        "bbox": [
+            207,
+            130,
+            789,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Heejeong  $\\mathrm{Nam}^{*1}$ , Jinwoo Ahn $^{*2}$ , Keummin  $\\mathrm{Ka}^3$ , Jiwan Chung $^3$ , and Youngjae  $\\mathrm{Yu}^{\\dagger 3}$",
+        "bbox": [
+            169,
+            191,
+            826,
+            212
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Brown University",
+        "bbox": [
+            421,
+            223,
+            575,
+            241
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ UC Berkeley",
+        "bbox": [
+            441,
+            241,
+            557,
+            258
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "3Yonsei University",
+        "bbox": [
+            423,
+            258,
+            573,
+            277
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            248,
+            321,
+            326,
+            338
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Human communication often relies on visual cues to resolve ambiguity. While humans can intuitively integrate these cues, AI systems often find it challenging to engage in sophisticated multimodal reasoning. We introduce VAGUE, a benchmark evaluating multimodal AI systems' ability to integrate visual context for intent disambiguation. VAGUE consists of 1.6K ambiguous textual expressions, each paired with an image and multiple-choice interpretations, where the correct answer is only apparent with visual context. The dataset spans both staged, complex (Visual Commonsense Reasoning) and natural, personal (Ego4D) scenes, ensuring diversity. Our experiments reveal that existing multimodal AI models struggle to infer the speaker's true intent. While performance consistently improves from the introduction of more visual cues, the overall accuracy remains far below human performance, highlighting a critical gap in multimodal reasoning. Analysis of failure cases demonstrates that current models fail to distinguish true intent from superficial correlations in the visual scene, indicating that they perceive images but do not effectively reason with them. We release our code and data at https://hazel-heejeong-nam.github.io/vague/.",
+        "bbox": [
+            88,
+            354,
+            485,
+            672
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            91,
+            700,
+            220,
+            717
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Human communication is inherently contextual; for example, exclaiming \"Hey, this is a disaster!\" upon seeing a cluttered room conveys frustration or exaggeration rather than referring to an actual catastrophe. Without surrounding cues, textual dialogues can be ambiguous, making it difficult for models to accurately capture intent and nuance.",
+        "bbox": [
+            89,
+            727,
+            483,
+            816
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We consider the case of visual contextual cues. Consider Fig. 1, which depicts a speaker making a remark in a certain situation. Without specifying contexts introduced",
+        "bbox": [
+            89,
+            818,
+            483,
+            864
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/e2a8dbca01ca7f9cf2de59e57ff233135cba5e760ade85d2b736c88db6c48b0f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            568,
+            329,
+            599,
+            354
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Hey person1, looks like you're planning a party for the neighborhood squirrels.",
+        "bbox": [
+            604,
+            330,
+            846,
+            354
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/bb8ba85a6b859c036abd1d76f868a67d75e2753b51dd3ebe0919d8588c4362f9.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            562,
+            368,
+            580,
+            377
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/9b146a4622b5effe44329b07299bd76f8a88078e788e652b22affa5b4f4e0d2a.jpg",
+        "image_caption": [
+            "Visual Context 1"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            380,
+            705,
+            474
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/509ea48f33bcd7e74ab1e7eabd010e041181ef09744a26e395c2c99caa9b002f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            751,
+            368,
+            769,
+            377
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/1bdd4bc388a4d7802e8453cf79b21f4526a541f221b7420bec508b25359cb457.jpg",
+        "image_caption": [
+            "Visual Context 2"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            712,
+            378,
+            900,
+            476
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/87898e9bddf2979562ffd934554fc2f25bb756e79aba17f08f153bf2b21ff6c4.jpg",
+        "image_caption": [
+            "The speaker wants person1 to..."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            602,
+            481,
+            627,
+            502
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/f39f695adddd5796b67d5a45b0ac7a1636c83143ef23955c8f89b99045f1c29a.jpg",
+        "image_caption": [
+            "close the cereal box",
+            "hand out the envelop to the squirrels",
+            "Figure 1. A motivating example demonstrating the importance of visual context in understanding intention. Without a predetermined context, a single expression can convey multiple different intentions. The textual expression and Context 1 are from our dataset, while Context 2 is generated using DALL-E 3 [2] to help understanding."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            699,
+            518,
+            718,
+            532
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "from visual cues, the speaker's intention can vary, thus remaining ambiguous. This implies that the visual contexts play important roles in communication, raising the question: can AI systems integrate visual cues with ambiguous dialogue to infer the speaker's intent?",
+        "bbox": [
+            511,
+            656,
+            906,
+            733
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We introduce Visual Contexts ClArify ambiGUous Expressions (VAGUE), a benchmark consisting of 1.6K ambiguous textual expressions, each paired with a single image. VAGUE aims to model diverse and natural human-to-human interactions by setting each image as the speaker's viewpoint, where the speaker implicitly requests a certain action from a person within their field of view. We define the problem addressed through this setup as Multimodal Intention Disambiguation (MID), which involves reasoning about the most plausible request conditioned on visual context. Each sample in VAGUE is annotated with four multiple-choice candi",
+        "bbox": [
+            509,
+            734,
+            908,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV 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": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*These authors contributed equally.",
+        "bbox": [
+            107,
+            875,
+            297,
+            887
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "† Corresponding author.",
+        "bbox": [
+            109,
+            887,
+            235,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "1537",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "dates, ensuring clarity and preventing multiple valid answers caused by paraphrasing or hierarchical inclusion of meaning. The dataset is meticulously curated to ensure visual dependency; the ground-truth candidate is only preferable when considering the visual context. VAGUE includes visual scenes from both artificial sources (Visual Commonsense Reasoning [46]) and real-world scenarios (Ego4D [41]), capturing a broad spectrum of scene complexity and naturalness. The textual expressions in VAGUE are initially generated by GPT-4o [31] following instructions, then reviewed through extensive human rating and filtering to ensure naturalness and alignment with the corresponding images.",
+        "bbox": [
+            89,
+            90,
+            485,
+            273
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Experiments on VAGUE demonstrate that existing multimodal AI models struggle to infer a speaker's true intent in a multimodal setting. First, although models can leverage visual context—as seen by a performance progression from text-only Language Models (LMs) to pipelined Socratic Models (SMs) [47] and ultimately to end-to-end Visual Language Models (VLMs)—their overall accuracy remains significantly lower than that of humans, indicating a failure to capture the true intent. A closer analysis of failure cases reveals that the primary source of error is the models' inability to distinguish the true intent from a superficial understanding of the visual context. In other words, even though these multimodal systems can perceive the image content, they cannot effectively use this information to reason about the speaker's true intent.",
+        "bbox": [
+            89,
+            285,
+            485,
+            512
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In conclusion, we introduce a benchmark that exposes the limitations of current models in integrating visual cues with intent comprehension and identifies their primary failure mode. We anticipate that VAGUE will serve as a testing ground for the development of future multimodal conversational or embodied agents—systems that combine robust visual perception with nuanced conversational reasoning to effectively respond to user requests in complex scenes.",
+        "bbox": [
+            89,
+            523,
+            483,
+            647
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions are threefold:",
+        "bbox": [
+            109,
+            657,
+            321,
+            672
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- VAGUE: a novel benchmark for evaluating multimodal intention disambiguation. Validated through extensive human filtering, VAGUE is designed for robust quantitative assessment by ensuring both the ambiguity of queries and the visual (in)dependency to answer candidates.",
+            "- Carefully curated 1,677 scene images sourced from VCR [46] and Ego4D [41], capturing a wide range of scene complexity, diversity, and naturalness to ensure VAGUE's generalizability across various contexts.",
+            "- Experimental results highlighting a critical challenge in multimodal intention disambiguation: while existing models can perceive visual cues, they fail to effectively integrate this information into reasoning to deduce the speaker's true intent."
+        ],
+        "bbox": [
+            91,
+            686,
+            482,
+            897
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            513,
+            89,
+            653,
+            104
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Multimodal Theory of Mind",
+        "text_level": 1,
+        "bbox": [
+            513,
+            114,
+            767,
+            131
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Theory of Mind (ToM) refers to the ability to infer and reason about the intentions of others based on available information [34], where recent language models still struggle with relevant tasks [7] highlighting the need for dedicated research in this area. Initially, various methods and benchmarks have been proposed in unimodal settings, relying on text-based approaches [11, 36]. However, these methods often fail to capture the richness of real-world interactions, which often require integrating both linguistic and visual cues.",
+        "bbox": [
+            511,
+            138,
+            906,
+            273
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Moving beyond text-only contexts, recent work has incorporated visual information. MMToM [17] introduces a benchmark where models must process both visual and textual cues to solve question-answering tasks related to ToM. The BOSS dataset [9] is a multimodal dataset collected in situations where nonverbal communication is required. It is used to evaluate whether human beliefs can be inferred based on nonverbal cues during social interactions. Similarly, Chen et al. (2024) [5] propose a Video ToM model that leverages key video frames and transcripts, demonstrating improved reasoning on the Social-IQ 2.0 dataset [45]. MuMA-ToM [37] further extends this direction by assessing ToM reasoning in multi-agent interactions, evaluating a model's ability to infer human beliefs and goals based on video and text inputs. MToMnet [3] introduces a ToM-based neural network that integrates contextual cues, such as scene videos and object locations, with person-specific cues, to predict human beliefs in specific scenarios.",
+        "bbox": [
+            511,
+            275,
+            908,
+            547
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "However, progress in multimodal ToM remains constrained not only by the scarcity of high-quality datasets [5] but also by the lack of explicit consideration for the ambiguity and indirectness inherent in human communication.",
+        "bbox": [
+            511,
+            547,
+            908,
+            608
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Multimodal Implicature Understanding",
+        "text_level": 1,
+        "bbox": [
+            513,
+            619,
+            856,
+            636
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Implicature and the ambiguity that arises from it naturally emerge in everyday human conversation, requiring pragmatic understanding [38]. Early research on implicature understanding has primarily been conducted in text-only settings [27, 30, 40], with some studies specifically focusing on figurative language and metaphor [4, 21, 39]. However, since the ambiguity of standalone text is inherently limited, recent studies have extended to multiple modalities. One example is multimodal sarcasm understanding (MSU). WITS [19] and MOSES [20] are benchmarks for sarcasm explanation, both providing the speaker's emotion and voice tone as cues. DocMSU [8] is a document-level benchmark for sarcasm localization and detection. To improve MSU, EDGE [32], a graph-based approach, achieved strong performance. UR-FUNNY [13] is a benchmark for multimodal humor comprehension, incorporating facial expressions and voice tones as in MSU [19, 20]. Hessel et al. (2023) [14] intro",
+        "bbox": [
+            511,
+            643,
+            910,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "1538",
+        "bbox": [
+            483,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Q. Select the option that best explains the underlying intention of the utterance based on the given image.",
+        "bbox": [
+            106,
+            99,
+            465,
+            127
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/5e3b0a1b1a27751d011a6dcbe11cb25d073beac26e2f82c50585b0bb63611535.jpg",
+        "image_caption": [
+            "Figure 2. Description of the Multimodal Intention Disambiguation (MID) task in a Multiple-Choice Question format: Given an input image  $(I)$  and an indirect expression  $(p_i)$ , the goal is to infer the speaker's hidden intent  $(T)$  and select the most likely answer."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            106,
+            143,
+            151,
+            175
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Hey person1, spot the difference, this parking's a bit too special isn't it?",
+        "bbox": [
+            163,
+            146,
+            421,
+            172
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/d314ec8014afa1b1655e660cbf27f3b14d3fc17bd409690d07d9a53e5c01bf22.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            98,
+            183,
+            478,
+            311
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "a) The speaker wants person1 to admire the unusually decorated motorcycle in the parking lot.",
+            "b) The speaker wants person1 to enjoy playing a puzzle game and spot the difference.",
+            "c) The speaker wants person1 to move the sedan because it's in a handicapped parking spot.",
+            "d) The speaker wants person1 to move the sedan because it's parked in front of a fire hydrant."
+        ],
+        "bbox": [
+            98,
+            316,
+            475,
+            450
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "duced a benchmark derived from a Cartoon Caption Contest, exploring humor identification and explanation. Baluja et al. (2024) [1] demonstrated that models benefit from multimodal cues in humor understanding. Memes also involve implicature, with multimodal datasets such as MemeCap [15] and MultiBully-Ex [16].",
+        "bbox": [
+            89,
+            551,
+            483,
+            642
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "However, the cues used in multimodal implicature understanding remain simple, primarily appearing in images with a single main object or person, overlooking the importance of interactions between multiple objects and people in real-world scenarios. These limitations underscore the need for more complex cues, as addressed in VAGUE.",
+        "bbox": [
+            89,
+            643,
+            483,
+            733
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Multimodal Intention Disambiguation",
+        "text_level": 1,
+        "bbox": [
+            89,
+            750,
+            433,
+            766
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we outline the structure and rationale behind the format of our primary task, which we term Multimodal Intention Disambiguation (MID). Then, we further specify the necessary components that form the basis of the task.",
+        "bbox": [
+            89,
+            775,
+            482,
+            835
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Problem Setting",
+        "text_level": 1,
+        "bbox": [
+            89,
+            847,
+            253,
+            864
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Each MID problem comprises an input image  $I$ , a direct text expression  $p_d$ , and an indirect text expression  $p_i$ . Here,",
+        "bbox": [
+            89,
+            869,
+            483,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "the direct expression  $p_d$  clearly shows the underlying intention of the corresponding  $p_i$  and serves as an essential intermediate step of generating  $p_i$ .",
+        "bbox": [
+            511,
+            90,
+            906,
+            136
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To clarify our problem, we assume that all reasoning is confined to the depicted scene and that each expression is spoken by a human who intends for the listener to take a particular action based on the situation. The ultimate objective of the task is to interpret the hidden intention  $T$  effectively by leveraging the contextual cues within the image.",
+        "bbox": [
+            511,
+            137,
+            906,
+            228
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To evaluate how well models capture such intentions, we adopt a multiple-choice (MCQ) format as the primary setup, as shown in Fig. 2. This decision reflects the fact that certain prompts can lead to multiple plausible outcomes, driven by hierarchical relations (e.g., pick up the chips - snack - food) or by the inherent uncertainty of what action best satisfies the speaker's goal (e.g., an indirect prompt complaining about darkness could be addressed by either turning on a light or opening curtains). Exploring all possible valid interpretations is labor-intensive and often infeasible. Consequently, each MID instance is presented as four distinct options, one correct and three intentionally designed to be incorrect for different reasons (see Sec. 4.2.3), challenging models in both linguistic and visual reasoning.",
+        "bbox": [
+            511,
+            229,
+            908,
+            441
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Formally, let  $C$  be the set of all multiple-choice options  $c_{n}$ . Given an image  $I$  and an indirect prompt  $p_i$ , the task is to select the most valid interpretation of  $p_i$  from the predefined options, conditioned on the visual context in  $I$ . We define this task as follows:",
+        "bbox": [
+            511,
+            441,
+            905,
+            516
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nT (I, p _ {i}) := \\operatorname {a r g m a x} _ {c _ {n} \\in C} \\Pr \\left(c _ {n} \\mid I, p _ {i}\\right). \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            573,
+            537,
+            906,
+            555
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Direct and Indirect Expressions",
+        "text_level": 1,
+        "bbox": [
+            511,
+            566,
+            794,
+            582
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "By the design of our task, curating effective input prompts  $p_d$  and  $p_i$  is crucial for ensuring accurate interpretation. What makes a good prompt, though? In this section, we define and explain the criteria that both direct and indirect expressions must satisfy. For more details on good and bad examples for each criterion, please refer to our Appendix A.",
+        "bbox": [
+            511,
+            589,
+            905,
+            679
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2.1. Directness: Relevance and Solvability",
+        "text_level": 1,
+        "bbox": [
+            511,
+            691,
+            821,
+            705
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Relevance The direct prompt is an utterance from the speaker that explicitly conveys its intended meaning without ambiguity. However, it is equally important that this intention aligns with the visual context of the scene. For example, a direct prompt  $p_d$  such as \"Hey person1, I want you to stop the fireworks\" clearly expresses its intended action. However, if the corresponding image, as shown in Fig. 2, contains no elements related to fireworks, the prompt is misaligned with the scene. Thus, a direct prompt must not only reveal its intention but also maintain relevance to the image. In the context of our task, relevance is determined by whether a human can reasonably establish a connection between the prompt and the depicted scene.",
+        "bbox": [
+            511,
+            710,
+            906,
+            907
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "1539",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Solvability Relevance alone does not guarantee that a prompt is useful. As outlined in Sec. 3.1, the prompt  $p_d$  must explicitly request an action that the listener can reasonably perform. This introduces solvability, which requires that the prompt present a clear and actionable problem. A solvable prompt defines a specific issue that can be addressed independently, ensuring that the listener is not left with multiple competing actions to choose from.",
+        "bbox": [
+            89,
+            90,
+            480,
+            210
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.2. Indirectness: Consistency and Ambiguity",
+        "text_level": 1,
+        "bbox": [
+            89,
+            224,
+            424,
+            242
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Consistency Indirect prompts are designed to obscure their true intention, but they must still convey the same underlying intention as their direct counterpart—essentially requesting the same solution. Since indirect prompts are derived from direct prompts, consistency serves as a key criterion. We define a direct prompt  $p_d$  and an indirect prompt  $p_i$  as consistent if their interpretations could potentially align in intention. The term \"potentially\" is used because indirect prompts, by nature, may have multiple valid interpretations. For instance, in the earlier \"this is a disaster!\" example, it conveys distress but allows for multiple reasonable responses, such as cleaning the room or providing reassurance.",
+        "bbox": [
+            89,
+            244,
+            482,
+            428
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Ambiguity If an indirect prompt is entirely consistent with its direct counterpart without introducing any additional complexity, it becomes indistinguishable from a direct prompt. Therefore, an indirect prompt should conceal its underlying intention, which we define as ambiguity. The key principle behind this criterion is that neither the specific action required nor the key entity involved should be explicitly or implicitly mentioned within the prompt. Once these two elements are concealed, further refinements—such as adjusting the tone to be more indirect, sarcastic, or humorous—can enhance the overall nuance and difficulty of interpretation.",
+        "bbox": [
+            89,
+            446,
+            482,
+            612
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4. VAGUE Benchmark Construction",
+        "text_level": 1,
+        "bbox": [
+            89,
+            631,
+            401,
+            647
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "VAGUE is a novel benchmark that extends single-modal or simple multimodal ambiguity to more realistic domains and evaluates whether concurrent vision-language models can perform human-like reasoning with complex visual contexts. It comprises 1,677 images, with 1,144 sourced from the VCR [46] dataset and 533 from Ego4D [41], covering diverse contextual scenarios as well as real-world human interactions. On average, VAGUE contains seven objects and four people per image. Each image is paired with a direct expression  $p_d$ , an indirect expression  $p_i$ , and four multiple-choice answers, along with relevant meta-information. All textual components are generated using GPT-4o, then refined through extensive human rating, selection, and filtering, ensuring a carefully curated benchmark dataset for testing and advancing multimodal reasoning. We provide detailed benchmark statistics and a diversity analysis in Appendix B.2.",
+        "bbox": [
+            89,
+            657,
+            482,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1. Visual Data Curation",
+        "text_level": 1,
+        "bbox": [
+            513,
+            90,
+            714,
+            104
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1.1. Sampling",
+        "text_level": 1,
+        "bbox": [
+            513,
+            113,
+            627,
+            128
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "VCR [46] The VCR dataset consists of 110K movie scenes sourced from the Large Scale Movie Description Challenge [35] and YouTube clips. These images are curated based on an \"interestingness\" criterion [46], ensuring the presence of at least two people, which promotes interactive scenarios. To prevent redundancy in our dataset, we sample 10K images while carefully avoiding neighboring frames, as adjacent frames exhibit minimal variation. This selection process preserves the contextual diversity of the dataset while maintaining its focus on complex, multi-entity interactions.",
+        "bbox": [
+            511,
+            132,
+            906,
+            285
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Ego4D [41] While VCR provides a wide range of contextual diversity, it often includes artificially composed settings that may not fully capture real-world interactions. To address this, we integrate frames from the Ego4D dataset, which offers a more naturalistic depiction of human interactions. We specifically leverage the AV (Audio-Visual), which indicate conversational exchanges between individuals, to ensure the presence of people in the selected frames. Similar to VCR, we avoid neighboring frames to maintain diversity and filter out heavily blurry images to enhance data quality. This process results in 888 candidate images from 94 videos, which serve as the basis for further text processing.",
+        "bbox": [
+            511,
+            297,
+            908,
+            479
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1.2. Object Extraction",
+        "text_level": 1,
+        "bbox": [
+            513,
+            488,
+            684,
+            503
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To ensure the complexity of the visual information, we extract a list of physical objects present in each image using a tagging model, RAM [49]. This step allows us to easily identify scenes with sufficient visual detail. In the case of VCR, many scenes are relatively simple, often containing only a few objects. Therefore, we sort VCR images by the number of detected objects and retained the top 4,000 as candidates for text processing, ensuring that our benchmark primarily consists of rich visual cues in contextually diverse scenes. Please refer to Appendix B.3 for more details.",
+        "bbox": [
+            511,
+            508,
+            908,
+            659
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1.3. Person Indicator",
+        "text_level": 1,
+        "bbox": [
+            513,
+            669,
+            676,
+            683
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In our task, we assume that the speaker is outside the scene, viewing the image and talking to a person in it. However, identifying the addressee is not always straightforward, as images mostly contain multiple individuals. While grounding the specific person referenced in the utterance could introduce additional complexity, it is not the primary focus of our evaluation. To clarify the listener, we assign an indicator tag to each person in the image as shown in Fig. 3. For VCR, we use their existing annotations [46], while for Ego4D, where bounding boxes are not fully available, we employ YOLOv11 [18] to detect and annotate humans. While this provides a straightforward method for grounding the target person, it requires models to perform basic Optical Character Recognition (OCR). Therefore, we conduct exper",
+        "bbox": [
+            511,
+            688,
+            908,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "1540",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/9e1c2cc12c6b85ca62f868102b42281b9ec99a2fd7df1e4755d09e2cd23bc692.jpg",
+        "image_caption": [
+            "Figure 3. Overview of the data generation process. Based on human-defined criteria and instructions, GPT-4o [31] generates initial data, which are then rated and filtered by humans to ensure quality. Since generating high-quality indirect expressions from raw images is challenging, the process follows these steps: generating a direct expression and intention from the image, creating an indirect expression using information from the previous step, and producing answer candidates based on all the information gathered so far. In the answer candidates, FS stands for Fake Scene Understanding, SU stands for Superficial Understanding, and NE stands for Nonexistent Entity. FS evaluates global hallucination, where the model misinterprets visual context from an entirely different image (e.g., an outdoor camping scene with a candle), while NE addresses local hallucination, where only a specific object is replaced with a fabricated one (e.g., curtains)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            92,
+            898,
+            292
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "iments to evaluate the OCR capabilities of the models used in our study. See Appendix B.4 for details.",
+        "bbox": [
+            89,
+            430,
+            482,
+            460
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Multimodal Expression Synthesis",
+        "text_level": 1,
+        "bbox": [
+            89,
+            470,
+            382,
+            486
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As shown in Fig. 3, VAGUE's textual expressions are generated first by the model and undergo an extensive process of human rating and filtering. We use GPT-4o [31] for all text processing. The instructions used for generating direct  $(p_d)$  and indirect  $(p_i)$  expressions are provided in Appendix B.5, while the instructions for generating answer candidates for multiple-choice questions are detailed in Appendix B.7.",
+        "bbox": [
+            89,
+            494,
+            483,
+            599
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2.1. Direct Expressions",
+        "text_level": 1,
+        "bbox": [
+            89,
+            608,
+            266,
+            623
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The direct expression  $p_d$  serves as a crucial foundation for crafting the indirect one  $p_i$ , as both share the same underlying intention. To ensure that  $p_d$  adhere to the principles of relevance and solvability discussed in Sec. 3.2.1, we generate  $p_d$  conditioned on the input image  $I$  and a task prompt that explicitly defines these criteria. During the generation process, we also instruct the model to output a solution triplet in the format: (subject, action, object). Since direct expressions explicitly state their intentions, extracting each component of the solution triplet is straightforward. To maintain consistency with the visual context, the \"object\" in the triplet is restricted to physical objects we extracted in Sec. 4.1.2.",
+        "bbox": [
+            89,
+            628,
+            483,
+            808
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After generating  $p_d$  for all candidate images, human raters evaluate each prompt based on relevance and solvability, assigning scores from 1 to 5. Only those prompts that receive a rating of 4 or 5 are retained for further use. The detailed rating criteria for human verification and an example of the rating process are provided in Fig. J13.",
+        "bbox": [
+            89,
+            810,
+            483,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2.2. Indirect Expressions",
+        "text_level": 1,
+        "bbox": [
+            513,
+            430,
+            702,
+            444
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To ensure that the indirect expressions  $p_i$  maintain both ambiguity and fluency while aligning with the true intent  $T$ , we adopt a two-stage process: proposal and selection.",
+        "bbox": [
+            511,
+            449,
+            906,
+            494
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the initial step, the model is prompted to generate three distinct candidate options. Each candidate follows the criteria outlined in Section 3.2.2, but with explicit instructions to incorporate different linguistic strategies: sarcasm, humor, and meme/idiomatic expressions, respectively. This approach ensures diversity in the generated responses. In the second step, human annotators evaluate the three candidates and select the one that best aligns with the intended indirectness. The selected prompt is then rated on a scale from 1 to 5 based on more specific criteria. Only those prompts that receive a score of 3 or higher are retained for use in the dataset. The detailed rating criteria for human verification and an example of rating process are provided in Fig. J14",
+        "bbox": [
+            511,
+            496,
+            908,
+            691
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2.3. Counterfactual Choices",
+        "text_level": 1,
+        "bbox": [
+            511,
+            702,
+            723,
+            715
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Generating high-quality counterfactual choices is crucial. To enable detailed analysis of model weaknesses in multimodal intent disambiguation, we design interpretable counterfactual choices that provide more plausible alternatives.",
+        "bbox": [
+            511,
+            720,
+            906,
+            781
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Fake Scene Understanding The first counterfactual choice is an interpretation that could arise when the model largely misinterprets the image. This process is conducted in two steps. In the first step, a fake caption is generated by assuming an imaginary scene that can be aligned with the indirect expression but is inconsistent with the true intent. The caption of fake scene is then combined with the speaker's",
+        "bbox": [
+            511,
+            794,
+            908,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "1541",
+        "bbox": [
+            483,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "indirect statement to derive the most likely interpretation.",
+        "bbox": [
+            89,
+            90,
+            468,
+            107
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Superficial Understanding The subsequent choice corresponds to an interpretation generated when the model fails to deeply reason about the implicit intent of the sentence and instead relies on surface-level meaning. We enforce the model to focus only on the literal wording, without considering any implied or deeper meaning of the indirect sentence. These answer choices are generated alongside the indirect expression  $p_i$ . During the indirect selection phase, the corresponding superficially understood choice is selected together, maintaining coherence between them.",
+        "bbox": [
+            89,
+            117,
+            485,
+            268
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Nonexistent Entity The last choice arises when the model interprets the text correctly but fails to adequately consider the details of the image, resulting in a plausible yet incorrect choice. This resembles the correct answer in structure, but replaces the key object in the solution with one that does not exist in the image. To prevent the task from becoming too easy by generating highly irrelevant objects, we constrain the substituted object to one that, while absent from the image, is highly expected to be present in the scene and could replace the original entity. To identify such entities, the model is provided with the image as input to choose objects that align with the scene's context. This method ensures that the counterfactual choice leverages the expected coherence between the scene and its potential entities while rigorously testing the model's attention to visual details.",
+        "bbox": [
+            89,
+            280,
+            483,
+            507
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            522,
+            222,
+            539
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Models We use the following models in our experiments. The detailed descriptions of each model are in Appendix C.",
+        "bbox": [
+            89,
+            547,
+            482,
+            578
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Phi3.5-Vision-Instruct (4B) [29]",
+            "- LLaVA Onevision (7B) [23]",
+            "Qwen2.5-VL-Instruct (7B, 72B) [43]",
+            "- InternVL-2.5-MPO (8B, 26B) [6]",
+            "- Idefics2 (8B) [22]",
+            "LLaVA NeXT Vicuna (13B) [24]",
+            "- Ovis2 (16B) [28]",
+            "GPT-4o [31]",
+            "Gemini 1.5 Pro [12]",
+            "- InternVL-3 (38B) [42]"
+        ],
+        "bbox": [
+            93,
+            579,
+            351,
+            729
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. MLLMs Benefit from Visual Cues",
+        "text_level": 1,
+        "bbox": [
+            89,
+            741,
+            387,
+            756
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Our first objective is to assess how effectively MLLMs leverage visual cues to resolve ambiguity in utterances. To this end, we systematically control the level of detail in the visual cues provided to the models and measure their accuracy in inferring the speaker's true intent. Performance is evaluated in both multiple-choice and free-form settings. For clarity, we primarily report multiple-choice accuracy, deferring free-form results to Appendix I.",
+        "bbox": [
+            89,
+            763,
+            483,
+            883
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We consider three levels of visual cues:",
+        "bbox": [
+            109,
+            885,
+            370,
+            898
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Language Models (LMs) receive no visual input, requiring models to rely on superficial textual priors such as common-sense knowledge of sarcasm or humor to determine intent.",
+            "- Socratic Models (SMs) [47] use text-only LMs but incorporate short image captions (up to two or three sentences) as additional input. This short generic caption may lack sufficient detail, which may be insufficient for accurately inferring intent. Each SM model generated its own image captions and used them in subsequent processing.",
+            "- Visual Language Models (VLMs) receive the raw image input, enabling a more direct interpretation of visual cues."
+        ],
+        "bbox": [
+            513,
+            90,
+            906,
+            272
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Results The results in Tab. 1 indicate that MLLMs can leverage visual cues, albeit to a limited extent, since SMs and VLMs consistently outperform LMs across all evaluated models. Additionally, more detailed visual input generally improves performance, with VLMs surpassing SMs in most cases, except in the case of proprietary models. This exception is further analyzed in Sec. 5.2. Both the VCR and Ego4D subsets exhibit similar performance trends, demonstrating the generalizability of our findings across both staged and real-world scenarios. Finally, the consistently low performance of LMs further reinforces the validity of our dataset as a multimodal benchmark.",
+        "bbox": [
+            511,
+            284,
+            908,
+            465
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2. Analysis on Failure Modes",
+        "text_level": 1,
+        "bbox": [
+            511,
+            476,
+            754,
+            491
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Here, we examine the ways in which models fail to infer the true intent and how these failure patterns vary with the level of visual cues provided. As shown in Fig. 3, our multiple-choice questions include three distinct types of incorrect answer candidates. We assess the model's raw visual understanding using the Fake Scene Understanding (FS) and Nonexistent Entity (NE) candidates, which test whether the model can correctly interpret the scene without being misled by fabricated or nonexistent elements. Conversely, the Superficial Understanding (SU) candidate evaluates the model's reasoning ability, testing whether it can go beyond surface-level perception to infer intent. We provide the full table of failure modes in the MCQ setting in Appendix I",
+        "bbox": [
+            511,
+            498,
+            908,
+            696
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Results Figure 4 illustrates how frequently each model selects different types of incorrect answers instead of the correct intent. Among the error types, Superficial Understanding (SU) is the most common. This indicates that while models generally succeed in recognizing basic visual details, they often fail to reason deeply about those visual cues to accurately infer the underlying intent of the speaker. However, proprietary models exhibit fewer SU-related errors, indicating stronger reasoning capabilities.",
+        "bbox": [
+            511,
+            703,
+            908,
+            839
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Moreover, stronger visual cues improve accuracy across all models and failure types. This improvement highlights the crucial role of visual conditioning in reducing both vision-based errors (FS and NE) and reasoning-related failures (SU).",
+        "bbox": [
+            511,
+            840,
+            908,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "1542",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/62bf245532a064985b3cff71bbdd8486b026798d8780805b5785c577faafa5ab.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"3\">VAGUE-VCR</td><td colspan=\"3\">VAGUE-Ego4D</td></tr><tr><td>LM (L)</td><td>SM (L+V)</td><td>VLM (L+V)</td><td>LM (L)</td><td>SM (L+V)</td><td>VLM (L+V)</td></tr><tr><td>Phi3.5-Vision-Instruct (4B)</td><td>26.6</td><td>35.3 (↑ 8.7)</td><td>46.0 (↑ 19.4)</td><td>22.5</td><td>31.1 (↑ 8.6)</td><td>42.4 (↑ 19.9)</td></tr><tr><td>LLaVA-Onevision (7B)</td><td>13.1</td><td>29.4 (↑ 16.3)</td><td>43.1 (↑ 30.0)</td><td>11.3</td><td>29.5 (↑ 18.2)</td><td>43.2 (↑ 31.9)</td></tr><tr><td>Qwen2.5-VL-Instruct (7B)</td><td>11.1</td><td>25.6 (↑ 14.5)</td><td>46.8 (↑ 35.7)</td><td>9.8</td><td>28.0 (↑ 18.2)</td><td>48.4 (↑ 38.6)</td></tr><tr><td>InternVL-2.5-MPO (8B)</td><td>23.0</td><td>48.4 (↑ 25.4)</td><td>63.9 (↑ 40.9)</td><td>24.2</td><td>54.0 (↑ 29.8)</td><td>66.8 (↑ 42.6)</td></tr><tr><td>Idefics2 (8B)</td><td>13.9</td><td>21.1 (↑ 7.2)</td><td>58.7 (↑ 44.8)</td><td>14.8</td><td>18.2 (↑ 3.4)</td><td>58.3 (↑ 43.5)</td></tr><tr><td>LLaVA-NeXT-vicuna (13B)</td><td>24.2</td><td>37.2 (↑ 13)</td><td>46.4 (↑ 22.2)</td><td>20.3</td><td>34.1 (↑ 13.8)</td><td>52.5 (↑ 32.2)</td></tr><tr><td>Ovis2 (16B)</td><td>21.9</td><td>23.8 (↑ 1.9)</td><td>24.5 (↑ 3.6)</td><td>20.5</td><td>25.3 (↑ 4.8)</td><td>25.7 (↑ 5.2)</td></tr><tr><td>InternVL-2.5-MPO (26B)</td><td>21.2</td><td>48.5 (↑ 27.3)</td><td>63.7 (↑ 42.5)</td><td>21.8</td><td>55.2 (↑ 33.4)</td><td>68.7 (↑ 46.9)</td></tr><tr><td>InternVL-3 (38B)</td><td>24.8</td><td>47.2 (↑ 22.4)</td><td>63.6 (↑ 38.8)</td><td>18.0</td><td>47.5 (↑ 29.5)</td><td>59.8 (↑ 41.8)</td></tr><tr><td>Qwen2.5-VL-Instruct (72B)</td><td>29.6</td><td>55.6 (↑ 26.0)</td><td>74.2 (↑ 44.6)</td><td>26.8</td><td>59.3 (↑ 32.5)</td><td>69.8 (↑ 43.0)</td></tr><tr><td>GPT-4o</td><td>46.4</td><td>69.5 (↑ 23.1)</td><td>65.1 (↑ 18.7)</td><td>48.2</td><td>67.5 (↑ 19.3)</td><td>63.6 (↑ 15.3)</td></tr><tr><td>Gemini-1.5-Pro</td><td>43.2</td><td>62.4 (↑ 19.2)</td><td>60.6 (↑ 17.4)</td><td>40.3</td><td>60.6 (↑ 20.3)</td><td>60.6 (↑ 20.3)</td></tr></table>",
+        "bbox": [
+            91,
+            88,
+            906,
+            337
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Experiments on the Multimodal Intention Disambiguation (MID) task with varying levels of visual cues. We report the accuracy  $(\\%)$  of the Multiple-Choice Question.  $\\uparrow$  indicates the performance gain from visual cues, i.e. increment compared to the LM setting.(L) denotes the use of language input only, while  $(\\mathrm{L} + \\mathrm{V})$  indicates the incorporation of visual cues. The noticeable increase in accuracy across LM, SM, and VLM demonstrates that the introduction of detailed visual cues is beneficial for the task.",
+        "bbox": [
+            89,
+            347,
+            906,
+            402
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f063bf44d3c5ca75ea7e8831036cb2b6bff129c5ecbd71bdc54b08faf9ac3d72.jpg",
+        "image_caption": [
+            "Figure 4. We present a bar plot to analyze the distribution of incorrect answer choices selected by each model. Each number represents how frequently a given choice was selected from the 1,677 items in the dataset. The counterfactual choice categories are FS (Fake Scene Understanding), SU (Superficial Understanding), and NE (Nonexistent Entity). We use distinct colors to represent LM (Language Models), SM (Socratic Models), and VLM (Visual-Language Models)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            417,
+            908,
+            580
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Notably, proprietary models perform better with captioned inputs (SM) than with raw images (VLM). A closer examination of Fig. 4 reveals that this discrepancy arises from vision-based failures (FS and NE) rather than reasoning-centric failure (SU). This indicates that their captioning ability potentially allows them to obtain more sophisticated information while reducing hallucination in the images. The supporting experiments and explanations are provided in Appendix F.",
+        "bbox": [
+            89,
+            672,
+            482,
+            794
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3. Comparison with human",
+        "text_level": 1,
+        "bbox": [
+            89,
+            803,
+            318,
+            819
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To validate our benchmark and establish an upper bound for performance, we assess human accuracy, highlighting the gap between existing models and human capability. This evaluation follows the multiple-choice setup within the VLM setting, using a subset of 400 samples. As shown",
+        "bbox": [
+            89,
+            824,
+            483,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "in Tab. 2, human performance reaches  $94\\%$ , demonstrating near-perfect accuracy. Although proprietary models and some large-sized models show a decent performance, a notable performance gap ( $\\sim 20\\%$ ) still exists compared to human evaluators. This result underscores the significant gap between AI models' multimodal reasoning capabilities and human-level understanding when inferring hidden intent, suggesting that visual perception alone, even when accurate, is insufficient without deeper cognitive integration. This performance gap is due to the models' tendency to rely on surface-level text rather than understanding deeper visual-textual implications. Thus, advancing multimodal reasoning likely requires models to integrate higher-order cognitive processes, like commonsense reasoning and pragmatic understanding, into visual interpretation tasks. Refer",
+        "bbox": [
+            511,
+            672,
+            908,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "1543",
+        "bbox": [
+            483,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/54538f880812ae8c5bba8b5ff31d0c85d4580b2bec406d2a9e499f683d40dfea.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Acc (%)</td><td>FS</td><td>SU</td><td>NE</td><td>Correct</td></tr><tr><td>Ovis2 (16B)</td><td>23.0</td><td>119</td><td>171</td><td>18</td><td>92</td></tr><tr><td>LLaVA-Onevision (7B)</td><td>41.0</td><td>43</td><td>183</td><td>10</td><td>164</td></tr><tr><td>Phi3.5-Vision-Instruct (4B)</td><td>44.3</td><td>60</td><td>132</td><td>31</td><td>177</td></tr><tr><td>Qwen2.5-VL-Instruct (7B)</td><td>47.0</td><td>47</td><td>152</td><td>13</td><td>188</td></tr><tr><td>LLaVA-NeXT-icuna (13B)</td><td>48.0</td><td>48</td><td>143</td><td>17</td><td>192</td></tr><tr><td>Idefics2 (8b)</td><td>57.0</td><td>28</td><td>120</td><td>24</td><td>228</td></tr><tr><td>Gemini-1.5-Pro</td><td>60.3</td><td>60</td><td>73</td><td>26</td><td>241</td></tr><tr><td>InternVL-3 (38B)</td><td>61.5</td><td>38</td><td>94</td><td>22</td><td>246</td></tr><tr><td>InternVL-2.5-MPO (8B)</td><td>61.8</td><td>42</td><td>95</td><td>16</td><td>247</td></tr><tr><td>GPT-4o</td><td>62.3</td><td>61</td><td>63</td><td>27</td><td>249</td></tr><tr><td>InternVL-2.5-MPO (26B)</td><td>63.0</td><td>36</td><td>101</td><td>11</td><td>252</td></tr><tr><td>Qwen2.5-VL-Instruct (72B)</td><td>72.3</td><td>39</td><td>60</td><td>12</td><td>289</td></tr><tr><td>Human</td><td>94.0</td><td>12</td><td>4</td><td>8</td><td>374</td></tr></table>",
+        "bbox": [
+            91,
+            88,
+            482,
+            277
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/1c4c73acf4369df3333c7bbaad8ef0c6b89d6f8a77aa4e034ea3936ac1d2dce3.jpg",
+        "table_caption": [
+            "Table 2. Performance across models and humans on a sampled set of 400 questions. The results show that humans outperform models by a margin of over  $20\\%$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td rowspan=\"2\">Type</td><td rowspan=\"2\">Acc (%)</td><td colspan=\"3\">Incorrect count</td></tr><tr><td>FS</td><td>SU</td><td>NE</td></tr><tr><td rowspan=\"4\">GPT-4o</td><td>SM</td><td>68.9</td><td>165</td><td>240</td><td>117</td></tr><tr><td>SM+CoT</td><td>69.5 (↑ 0.6)</td><td>165</td><td>241</td><td>105</td></tr><tr><td>VLM</td><td>64.6</td><td>226</td><td>226</td><td>141</td></tr><tr><td>VLM+CoT</td><td>66.4 (↑ 1.8)</td><td>162</td><td>156</td><td>85</td></tr><tr><td rowspan=\"4\">Gemini-1.5-Pro</td><td>SM</td><td>61.8</td><td>176</td><td>374</td><td>83</td></tr><tr><td>SM+CoT</td><td>61.0 (↓ 0.8)</td><td>190</td><td>367</td><td>94</td></tr><tr><td>VLM</td><td>60.6</td><td>249</td><td>264</td><td>145</td></tr><tr><td>VLM+CoT</td><td>64.4 (↑ 3.8)</td><td>213</td><td>267</td><td>117</td></tr></table>",
+        "bbox": [
+            91,
+            345,
+            478,
+            513
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Result of Chain-of-Thought (CoT) experiments on proprietary models, in both SM and VLM settings.  $\\uparrow$  and  $\\downarrow$  indicate an increase and decrease in accuracy when zero-shot CoT is applied.",
+        "bbox": [
+            89,
+            526,
+            482,
+            570
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "to Appendix E for details on the selected subset and human evaluation setup.",
+        "bbox": [
+            89,
+            595,
+            482,
+            628
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4. Chain-of-Thought Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            638,
+            370,
+            655
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Given the strong reasoning demands of multimodal intent deduction, we further explore the effectiveness of Chain-of-Thought (CoT) prompting [44] in enhancing the reasoning capabilities of MLLMs. The CoT prompt templates, provided in Fig. J21 and Fig. J22, are designed to explicitly ground the reasoning process, reducing hallucinations. While our main results focus on proprietary models, owing to their superior suitability for zero-shot CoT reasoning, we also present experiments and analyses for open-source models ranging from 4B to 72B in Appendix G.",
+        "bbox": [
+            89,
+            661,
+            482,
+            813
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Results As shown in Tab. 3, CoT prompting improves performance for raw image inputs (VLM), while showing no clear trend and remaining at a similar level for image caption inputs (SM). One possible explanation for this discrepancy is that CoT primarily enhances reasoning by improving",
+        "bbox": [
+            89,
+            825,
+            485,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "grounding and reducing hallucinations. Since image captions inherently contain fewer hallucinations, SMs see some benefit from CoT prompting albeit at the cost of reduced detail. Additionally, the performance improvements observed with CoT prompting are consistent across different types of false answer candidates, suggesting a generalizable effect in enhancing reasoning quality.",
+        "bbox": [
+            511,
+            90,
+            906,
+            198
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            511,
+            210,
+            633,
+            226
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We present VAGUE (Visual Contexts CIArify ambiGUous Expressions), a new benchmark aimed at assessing models' ability to interpret nuanced communication in complex multimodal scenarios. Our results show that models benefit from visual information when inferring the underlying intention of indirect expressions, as evidenced by their improved performance with increasing levels of visual cues. However, a significant disparity persists between machine capabilities and humans. To gain deeper insights into model inaccuracies, we design multiple-choice questions that explicitly address failure points, enabling a systematic and quantitative evaluation of the reasons behind performance. The primary challenge identified is the tendency of multimodal models to inadequately integrate visual cues, relying instead on the literal interpretation of textual information. This shortcoming highlights the need for further research, and we anticipate that VAGUE will open up promising avenues for developing systems capable of deeper multimodal reasoning to enhance AI's ability to engage in human-like interactions.",
+        "bbox": [
+            511,
+            234,
+            908,
+            523
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7. Limitations",
+        "text_level": 1,
+        "bbox": [
+            511,
+            536,
+            635,
+            551
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "First, we acknowledge that there are cultural and linguistic limitations. We incorporate sarcastic, humorous, and idiomatic implicatures in generating indirect expressions. However, since the initial data drafts are created by a model (GPT-4o [31]), they may reflect cultural biases present in its training data. To prevent any potential ethical issues, all human annotators are instructed to remove any content considered problematic or discriminatory during the rating and filtering process. Also, all textual expressions in VAGUE are limited to English. Therefore, we encourage future exploration of indirect expressions across diverse languages and cultures. The second limitation pertains to the dependency of certain meta-information on the quality of the parent dataset and the performance of the models utilized during the dataset generation. We observe that bounding box annotations from YOLOv11 [18] are occasionally duplicated, leading to a reported number of people higher than actually present. Likewise, the tagging model RAM [49] sometimes misidentified objects. Therefore, we remove any corrupted instances that could undermine the integrity of the task during the human rating and filtering process.",
+        "bbox": [
+            511,
+            561,
+            908,
+            878
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "1544",
+        "bbox": [
+            483,
+            944,
+            514,
+            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] Ashwin Baluja. Text is not all you need: Multimodal prompting helps llms understand humor. arXiv preprint arXiv:2412.05315, 2024. 3",
+            "[2] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 1",
+            "[3] Matteo Bortoletto, Constantin Ruhdorfer, Lei Shi, and Andreas Bulling. Explicit modelling of theory of mind for belief prediction in nonverbal social interactions. In Proc. 27th European Conference on Artificial Intelligence (ECAI), pages 1-8, 2024. 2",
+            "[4] Tuhin Chakrabarty, Arkadiy Saakyan, Debanjan Ghosh, and Smaranda Muresan. FLUTE: Figurative language understanding through textual explanations. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7139-7159, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 2",
+            "[5] Zhawnen Chen, Tianchun Wang, Yizhou Wang, Michal Kosinski, Xiang Zhang, Yun Fu, and Sheng Li. Through the theory of mind's eye: Reading minds with multimodal video large language models. arXiv preprint arXiv:2406.13763, 2024. 2",
+            "[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 6, 3",
+            "[7] Christine Cuskley, Rebecca Woods, and Molly Flaherty. The limitations of large language models for understanding human language and cognition. Open Mind, 8:1058-1083, 2024. 2",
+            "[8] Hang Du, Guoshun Nan, Sicheng Zhang, Binzhu Xie, Junrui Xu, Hehe Fan, Qimei Cui, Xiaofeng Tao, and Xudong Jiang. Docmsu: A comprehensive benchmark for document-level multimodal sarcasm understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17933-17941, 2024. 2",
+            "[9] Jiafei Duan, Samson Yu, Nicholas Tan, Li Yi, and Cheston Tan. Boss: A benchmark for human belief prediction in object-context scenarios. arXiv preprint arXiv:2206.10665, 2022. 2",
+            "[10] Alessandro Favero, Luca Zancato, Matthew Trager, Siddharth Choudhary, Pramuditha Perera, Alessandro Achille, Ashwin Swaminathan, and Stefano Soatto. Multi-modal hallucination control by visual information grounding. CVPR 2024, 2024. 4",
+            "[11] Kanishk Gandhi, Jan-Philipp Fraenkel, Tobias Gerstenberg, and Noah D. Goodman. Understanding social reasoning in language models with language models. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023), Datasets and Benchmarks Track, 2023. Dataset and Benchmarks Track. 2",
+            "[12] Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 6, 3"
+        ],
+        "bbox": [
+            93,
+            114,
+            483,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[13] Md Kamrul Hasan, Wasifur Rahman, AmirAli Bagher Zadeh, Jianyuan Zhong, Md Iftekhar Tanveer, Louis-Philippe Morency, and Mohammed (Ehsan) Hoque. UR-FUNNY: A multimodal language dataset for understanding humor. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2046–2056, Hong Kong, China, 2019. Association for Computational Linguistics. 2",
+            "[14] Jack Hessel, Ana Marasovic, Jena D. Hwang, Lillian Lee, Jeff Da, Rowan Zellers, Robert Mankoff, and Yejin Choi. Do androids laugh at electric sheep? humor \"understanding\" benchmarks from the new yorker caption contest. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 688-714, Toronto, Canada, 2023. Association for Computational Linguistics. 2",
+            "[15] EunJeong Hwang and Vered Shwartz. Memecap: A dataset for captioning and interpreting memes. arXiv preprint arXiv:2305.13703, 2023. 3",
+            "[16] Prince Jha, Krishanu Maity, Raghav Jain, Apoorv Verma, Sriparna Saha, and Pushpak Bhattacharyya. Meme-ingful analysis: Enhanced understanding of cyberbullying in memes through multimodal explanations. arXiv preprint arXiv:2401.09899, 2024.3",
+            "[17] Chuanyang Jin, Yutong Wu, Jing Cao, Jiannan Xiang, Yen Ling Kuo, Zhiting Hu, Tomer Ullman, Antonio Torralba, Joshua Tenenbaum, and Tianmin Shu. MMToM-QA: Multimodal theory of mind question answering. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16077-16102, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2",
+            "[18] Rahima Khanam and Muhammad Hussain. Yolov11: An overview of the key architectural enhancements, 2024. 4, 8",
+            "[19] Shivani Kumar, Atharva Kulkarni, Md Shad Akhtar, and Tan-moy Chakraborty. When did you become so smart, oh wise one?! sarcasm explanation in multi-modal multi-party dialogues. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5956-5968, Dublin, Ireland, 2022. Association for Computational Linguistics. 2",
+            "[20] Shivani Kumar, Ishani Mondai, Md Shad Akhtar, and Tanmoy Chakraborty. Explaining (sarcastic) utterances to enhance affect understanding in multimodal dialogues. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. AAAI Press, 2023. 2",
+            "[21] Yash Kumar Lal and Mohaddeseh Bastan. SBU figures it out: Models explain figurative language. In Proceedings of the 3rd Workshop on Figurative Language Processing (FLP), pages 143-149, Abu Dhabi, United Arab Emirates (Hybrid), 2022. Association for Computational Linguistics. 2",
+            "[22] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language mod"
+        ],
+        "bbox": [
+            516,
+            92,
+            906,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "1545",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "els? Advances in Neural Information Processing Systems, 37: 87874-87907, 2025. 6, 3",
+            "[23] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6, 3",
+            "[24] 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, 2024. 6, 3",
+            "[25] Kevin Y. Li, Sachin Goyal, Joao D. Semedo, and J. Zico Kolter. Inference optimal vlms need fewer visual tokens and more parameters, 2025. 4",
+            "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision – ECCV 2014, pages 740–755, Cham, 2014. Springer International Publishing. 1, 3",
+            "[27] Annie Louis, Dan Roth, and Filip Radlinski. \"I'd rather just go to bed\": Understanding indirect answers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7411-7425, Online, 2020. Association for Computational Linguistics. 2",
+            "[28] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 6, 3",
+            "[29] Microsoft. Phi-3 technical report: A highly capable language model locally on your phone, 2024. 6, 3",
+            "[30] Maryam Sadat Mirzaei, Kourosh Meshgi, and Satoshi Sekine What is the real intention behind this question? dataset collection and intention classification. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13606-13622, Toronto, Canada, 2023. Association for Computational Linguistics. 2",
+            "[31] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5, 6, 8, 3",
+            "[32] Kun Ouyang, Liqiang Jing, Xuemeng Song, Meng Liu, Yu-peng Hu, and Liqiang Nie. Sentiment-enhanced graph-based sarcasm explanation in dialogue. arXiv:2402.11414, 2024. 2",
+            "[33] 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, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 3, 4",
+            "[34] David Premack and Guy Woodruff. Does the chimpanzee have a theory of mind? Behavioral and brain sciences, 1(4): 515-526, 1978. 2",
+            "[35] Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket Tandon, Chris Pal, Hugo Larochelle, Aaron Courville, and Bernt Schiele. Movie description. International Journal of Computer Vision, 2017. 4",
+            "[36] Melanie Sclar, Sachin Kumar, Peter West, Alane Suhr, Yejin Choi, and Yulia Tsvetkov. Minding language models' (lack of) theory of mind: A plug-and-play multi-character belief"
+        ],
+        "bbox": [
+            91,
+            90,
+            483,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "tracker. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), pages 13960-13980, Toronto, Canada, 2023. Association for Computational Linguistics. 2",
+            "[37] Haojun Shi, Suyu Ye, Xinyu Fang, Chuanyang Jin, Leyla Isik, Yen-Ling Kuo, and Tianmin Shu. Muma-tom: Multi-modal multi-agent theory of mind, 2024. 2",
+            "[38] Settaluri Sravanthi, Meet Doshi, Pavan Tankala, Rudra Murthy, Raj Dabre, and Pushpak Bhattacharyya. PUB: A pragmatics understanding benchmark for assessing LLMs' pragmatics capabilities. In Findings of the Association for Computational Linguistics: ACL 2024, pages 12075-12097, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2",
+            "[39] Chuandong Su, Fumiyo Fukumoto, Xiaoxi Huang, Jiyi Li, Rongbo Wang, and Zhiqun Chen. DeepMet: A reading comprehension paradigm for token-level metaphor detection. In Proceedings of the Second Workshop on Figurative Language Processing, pages 30-39, Online, 2020. Association for Computational Linguistics. 2",
+            "[40] Junya Takayama, Tomoyuki Kajiwara, and Yuki Arase. DIRECT: Direct and indirect responses in conversational text corpus. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1980-1989, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2",
+            "[41] Ego4d Team. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18995-19012, 2022. 2, 4",
+            "[42] OpenGVLab Team. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models, 2025. 6, 3",
+            "[43] Qwen Team. Qwen2.5-v1 technical report, 2025. 6, 3",
+            "[44] 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. Advances in neural information processing systems, 35:24824-24837, 2022. 8",
+            "[45] Alex Wilf, Leena Mathur, Sheryl Mathew, Claire Ko, Youssof Kebe, Paul Pu Liang, and Louis-Philippe Morency. Social-iq 2.0 challenge: Benchmarking multimodal social understanding. https://github.com/abwilf/Social-IQ-2.0-Challenge, 2023.2",
+            "[46] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6713-6724, 2018. 2, 4, 1",
+            "[47] Andy Zeng, Maria Attarian, Krzysztof Marcin Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael S Ryoo, Vikas Sindhwani, Johnny Lee, et al. Socratic models: Composing zero-shot multimodal reasoning with language. In The Eleventh International Conference on Learning Representations, 2023. 2, 6",
+            "[48] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations, 2020. 4"
+        ],
+        "bbox": [
+            516,
+            92,
+            906,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "1546",
+        "bbox": [
+            483,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "ref_text",
+        "text": "[49] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize anything: A strong image tagging model. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1724-1732, 2024. 4, 8, 1",
+        "bbox": [
+            91,
+            90,
+            486,
+            176
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "1547",
+        "bbox": [
+            483,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/321d7adc-1768-447e-aad1-8d77c13702f1_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:048b39b5ad94deb989147ab1faf39a05a487a5307433e61be817438ea8058139
+size 1392500

2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/full.md

@@ -0,0 +1,312 @@
+# VAGUE: Visual Contexts Clarify Ambiguous Expressions
+
+Heejeong  $\mathrm{Nam}^{*1}$ , Jinwoo Ahn $^{*2}$ , Keummin  $\mathrm{Ka}^3$ , Jiwan Chung $^3$ , and Youngjae  $\mathrm{Yu}^{\dagger 3}$
+
+Brown University
+
+$^{2}$ UC Berkeley
+
+3Yonsei University
+
+# Abstract
+
+Human communication often relies on visual cues to resolve ambiguity. While humans can intuitively integrate these cues, AI systems often find it challenging to engage in sophisticated multimodal reasoning. We introduce VAGUE, a benchmark evaluating multimodal AI systems' ability to integrate visual context for intent disambiguation. VAGUE consists of 1.6K ambiguous textual expressions, each paired with an image and multiple-choice interpretations, where the correct answer is only apparent with visual context. The dataset spans both staged, complex (Visual Commonsense Reasoning) and natural, personal (Ego4D) scenes, ensuring diversity. Our experiments reveal that existing multimodal AI models struggle to infer the speaker's true intent. While performance consistently improves from the introduction of more visual cues, the overall accuracy remains far below human performance, highlighting a critical gap in multimodal reasoning. Analysis of failure cases demonstrates that current models fail to distinguish true intent from superficial correlations in the visual scene, indicating that they perceive images but do not effectively reason with them. We release our code and data at https://hazel-heejeong-nam.github.io/vague/.
+
+# 1. Introduction
+
+Human communication is inherently contextual; for example, exclaiming "Hey, this is a disaster!" upon seeing a cluttered room conveys frustration or exaggeration rather than referring to an actual catastrophe. Without surrounding cues, textual dialogues can be ambiguous, making it difficult for models to accurately capture intent and nuance.
+
+We consider the case of visual contextual cues. Consider Fig. 1, which depicts a speaker making a remark in a certain situation. Without specifying contexts introduced
+
+![](images/e2a8dbca01ca7f9cf2de59e57ff233135cba5e760ade85d2b736c88db6c48b0f.jpg)
+
+Hey person1, looks like you're planning a party for the neighborhood squirrels.
+
+![](images/bb8ba85a6b859c036abd1d76f868a67d75e2753b51dd3ebe0919d8588c4362f9.jpg)
+
+![](images/9b146a4622b5effe44329b07299bd76f8a88078e788e652b22affa5b4f4e0d2a.jpg)  
+Visual Context 1
+
+![](images/509ea48f33bcd7e74ab1e7eabd010e041181ef09744a26e395c2c99caa9b002f.jpg)
+
+![](images/1bdd4bc388a4d7802e8453cf79b21f4526a541f221b7420bec508b25359cb457.jpg)  
+Visual Context 2
+
+![](images/87898e9bddf2979562ffd934554fc2f25bb756e79aba17f08f153bf2b21ff6c4.jpg)  
+The speaker wants person1 to...
+
+![](images/f39f695adddd5796b67d5a45b0ac7a1636c83143ef23955c8f89b99045f1c29a.jpg)  
+close the cereal box  
+hand out the envelop to the squirrels  
+Figure 1. A motivating example demonstrating the importance of visual context in understanding intention. Without a predetermined context, a single expression can convey multiple different intentions. The textual expression and Context 1 are from our dataset, while Context 2 is generated using DALL-E 3 [2] to help understanding.
+
+from visual cues, the speaker's intention can vary, thus remaining ambiguous. This implies that the visual contexts play important roles in communication, raising the question: can AI systems integrate visual cues with ambiguous dialogue to infer the speaker's intent?
+
+We introduce Visual Contexts ClArify ambiGUous Expressions (VAGUE), a benchmark consisting of 1.6K ambiguous textual expressions, each paired with a single image. VAGUE aims to model diverse and natural human-to-human interactions by setting each image as the speaker's viewpoint, where the speaker implicitly requests a certain action from a person within their field of view. We define the problem addressed through this setup as Multimodal Intention Disambiguation (MID), which involves reasoning about the most plausible request conditioned on visual context. Each sample in VAGUE is annotated with four multiple-choice candi
+
+dates, ensuring clarity and preventing multiple valid answers caused by paraphrasing or hierarchical inclusion of meaning. The dataset is meticulously curated to ensure visual dependency; the ground-truth candidate is only preferable when considering the visual context. VAGUE includes visual scenes from both artificial sources (Visual Commonsense Reasoning [46]) and real-world scenarios (Ego4D [41]), capturing a broad spectrum of scene complexity and naturalness. The textual expressions in VAGUE are initially generated by GPT-4o [31] following instructions, then reviewed through extensive human rating and filtering to ensure naturalness and alignment with the corresponding images.
+
+Experiments on VAGUE demonstrate that existing multimodal AI models struggle to infer a speaker's true intent in a multimodal setting. First, although models can leverage visual context—as seen by a performance progression from text-only Language Models (LMs) to pipelined Socratic Models (SMs) [47] and ultimately to end-to-end Visual Language Models (VLMs)—their overall accuracy remains significantly lower than that of humans, indicating a failure to capture the true intent. A closer analysis of failure cases reveals that the primary source of error is the models' inability to distinguish the true intent from a superficial understanding of the visual context. In other words, even though these multimodal systems can perceive the image content, they cannot effectively use this information to reason about the speaker's true intent.
+
+In conclusion, we introduce a benchmark that exposes the limitations of current models in integrating visual cues with intent comprehension and identifies their primary failure mode. We anticipate that VAGUE will serve as a testing ground for the development of future multimodal conversational or embodied agents—systems that combine robust visual perception with nuanced conversational reasoning to effectively respond to user requests in complex scenes.
+
+Our contributions are threefold:
+
+- VAGUE: a novel benchmark for evaluating multimodal intention disambiguation. Validated through extensive human filtering, VAGUE is designed for robust quantitative assessment by ensuring both the ambiguity of queries and the visual (in)dependency to answer candidates.  
+- Carefully curated 1,677 scene images sourced from VCR [46] and Ego4D [41], capturing a wide range of scene complexity, diversity, and naturalness to ensure VAGUE's generalizability across various contexts.  
+- Experimental results highlighting a critical challenge in multimodal intention disambiguation: while existing models can perceive visual cues, they fail to effectively integrate this information into reasoning to deduce the speaker's true intent.
+
+# 2. Related Work
+
+# 2.1. Multimodal Theory of Mind
+
+Theory of Mind (ToM) refers to the ability to infer and reason about the intentions of others based on available information [34], where recent language models still struggle with relevant tasks [7] highlighting the need for dedicated research in this area. Initially, various methods and benchmarks have been proposed in unimodal settings, relying on text-based approaches [11, 36]. However, these methods often fail to capture the richness of real-world interactions, which often require integrating both linguistic and visual cues.
+
+Moving beyond text-only contexts, recent work has incorporated visual information. MMToM [17] introduces a benchmark where models must process both visual and textual cues to solve question-answering tasks related to ToM. The BOSS dataset [9] is a multimodal dataset collected in situations where nonverbal communication is required. It is used to evaluate whether human beliefs can be inferred based on nonverbal cues during social interactions. Similarly, Chen et al. (2024) [5] propose a Video ToM model that leverages key video frames and transcripts, demonstrating improved reasoning on the Social-IQ 2.0 dataset [45]. MuMA-ToM [37] further extends this direction by assessing ToM reasoning in multi-agent interactions, evaluating a model's ability to infer human beliefs and goals based on video and text inputs. MToMnet [3] introduces a ToM-based neural network that integrates contextual cues, such as scene videos and object locations, with person-specific cues, to predict human beliefs in specific scenarios.
+
+However, progress in multimodal ToM remains constrained not only by the scarcity of high-quality datasets [5] but also by the lack of explicit consideration for the ambiguity and indirectness inherent in human communication.
+
+# 2.2. Multimodal Implicature Understanding
+
+Implicature and the ambiguity that arises from it naturally emerge in everyday human conversation, requiring pragmatic understanding [38]. Early research on implicature understanding has primarily been conducted in text-only settings [27, 30, 40], with some studies specifically focusing on figurative language and metaphor [4, 21, 39]. However, since the ambiguity of standalone text is inherently limited, recent studies have extended to multiple modalities. One example is multimodal sarcasm understanding (MSU). WITS [19] and MOSES [20] are benchmarks for sarcasm explanation, both providing the speaker's emotion and voice tone as cues. DocMSU [8] is a document-level benchmark for sarcasm localization and detection. To improve MSU, EDGE [32], a graph-based approach, achieved strong performance. UR-FUNNY [13] is a benchmark for multimodal humor comprehension, incorporating facial expressions and voice tones as in MSU [19, 20]. Hessel et al. (2023) [14] intro
+
+Q. Select the option that best explains the underlying intention of the utterance based on the given image.
+
+![](images/5e3b0a1b1a27751d011a6dcbe11cb25d073beac26e2f82c50585b0bb63611535.jpg)  
+Figure 2. Description of the Multimodal Intention Disambiguation (MID) task in a Multiple-Choice Question format: Given an input image  $(I)$  and an indirect expression  $(p_i)$ , the goal is to infer the speaker's hidden intent  $(T)$  and select the most likely answer.
+
+Hey person1, spot the difference, this parking's a bit too special isn't it?
+
+![](images/d314ec8014afa1b1655e660cbf27f3b14d3fc17bd409690d07d9a53e5c01bf22.jpg)
+
+a) The speaker wants person1 to admire the unusually decorated motorcycle in the parking lot.  
+b) The speaker wants person1 to enjoy playing a puzzle game and spot the difference.  
+c) The speaker wants person1 to move the sedan because it's in a handicapped parking spot.  
+d) The speaker wants person1 to move the sedan because it's parked in front of a fire hydrant.
+
+duced a benchmark derived from a Cartoon Caption Contest, exploring humor identification and explanation. Baluja et al. (2024) [1] demonstrated that models benefit from multimodal cues in humor understanding. Memes also involve implicature, with multimodal datasets such as MemeCap [15] and MultiBully-Ex [16].
+
+However, the cues used in multimodal implicature understanding remain simple, primarily appearing in images with a single main object or person, overlooking the importance of interactions between multiple objects and people in real-world scenarios. These limitations underscore the need for more complex cues, as addressed in VAGUE.
+
+# 3. Multimodal Intention Disambiguation
+
+In this section, we outline the structure and rationale behind the format of our primary task, which we term Multimodal Intention Disambiguation (MID). Then, we further specify the necessary components that form the basis of the task.
+
+# 3.1. Problem Setting
+
+Each MID problem comprises an input image  $I$ , a direct text expression  $p_d$ , and an indirect text expression  $p_i$ . Here,
+
+the direct expression  $p_d$  clearly shows the underlying intention of the corresponding  $p_i$  and serves as an essential intermediate step of generating  $p_i$ .
+
+To clarify our problem, we assume that all reasoning is confined to the depicted scene and that each expression is spoken by a human who intends for the listener to take a particular action based on the situation. The ultimate objective of the task is to interpret the hidden intention  $T$  effectively by leveraging the contextual cues within the image.
+
+To evaluate how well models capture such intentions, we adopt a multiple-choice (MCQ) format as the primary setup, as shown in Fig. 2. This decision reflects the fact that certain prompts can lead to multiple plausible outcomes, driven by hierarchical relations (e.g., pick up the chips - snack - food) or by the inherent uncertainty of what action best satisfies the speaker's goal (e.g., an indirect prompt complaining about darkness could be addressed by either turning on a light or opening curtains). Exploring all possible valid interpretations is labor-intensive and often infeasible. Consequently, each MID instance is presented as four distinct options, one correct and three intentionally designed to be incorrect for different reasons (see Sec. 4.2.3), challenging models in both linguistic and visual reasoning.
+
+Formally, let  $C$  be the set of all multiple-choice options  $c_{n}$ . Given an image  $I$  and an indirect prompt  $p_i$ , the task is to select the most valid interpretation of  $p_i$  from the predefined options, conditioned on the visual context in  $I$ . We define this task as follows:
+
+$$
+T (I, p _ {i}) := \operatorname {a r g m a x} _ {c _ {n} \in C} \Pr \left(c _ {n} \mid I, p _ {i}\right). \tag {1}
+$$
+
+# 3.2. Direct and Indirect Expressions
+
+By the design of our task, curating effective input prompts  $p_d$  and  $p_i$  is crucial for ensuring accurate interpretation. What makes a good prompt, though? In this section, we define and explain the criteria that both direct and indirect expressions must satisfy. For more details on good and bad examples for each criterion, please refer to our Appendix A.
+
+# 3.2.1. Directness: Relevance and Solvability
+
+Relevance The direct prompt is an utterance from the speaker that explicitly conveys its intended meaning without ambiguity. However, it is equally important that this intention aligns with the visual context of the scene. For example, a direct prompt  $p_d$  such as "Hey person1, I want you to stop the fireworks" clearly expresses its intended action. However, if the corresponding image, as shown in Fig. 2, contains no elements related to fireworks, the prompt is misaligned with the scene. Thus, a direct prompt must not only reveal its intention but also maintain relevance to the image. In the context of our task, relevance is determined by whether a human can reasonably establish a connection between the prompt and the depicted scene.
+
+Solvability Relevance alone does not guarantee that a prompt is useful. As outlined in Sec. 3.1, the prompt  $p_d$  must explicitly request an action that the listener can reasonably perform. This introduces solvability, which requires that the prompt present a clear and actionable problem. A solvable prompt defines a specific issue that can be addressed independently, ensuring that the listener is not left with multiple competing actions to choose from.
+
+# 3.2.2. Indirectness: Consistency and Ambiguity
+
+Consistency Indirect prompts are designed to obscure their true intention, but they must still convey the same underlying intention as their direct counterpart—essentially requesting the same solution. Since indirect prompts are derived from direct prompts, consistency serves as a key criterion. We define a direct prompt  $p_d$  and an indirect prompt  $p_i$  as consistent if their interpretations could potentially align in intention. The term "potentially" is used because indirect prompts, by nature, may have multiple valid interpretations. For instance, in the earlier "this is a disaster!" example, it conveys distress but allows for multiple reasonable responses, such as cleaning the room or providing reassurance.
+
+Ambiguity If an indirect prompt is entirely consistent with its direct counterpart without introducing any additional complexity, it becomes indistinguishable from a direct prompt. Therefore, an indirect prompt should conceal its underlying intention, which we define as ambiguity. The key principle behind this criterion is that neither the specific action required nor the key entity involved should be explicitly or implicitly mentioned within the prompt. Once these two elements are concealed, further refinements—such as adjusting the tone to be more indirect, sarcastic, or humorous—can enhance the overall nuance and difficulty of interpretation.
+
+# 4. VAGUE Benchmark Construction
+
+VAGUE is a novel benchmark that extends single-modal or simple multimodal ambiguity to more realistic domains and evaluates whether concurrent vision-language models can perform human-like reasoning with complex visual contexts. It comprises 1,677 images, with 1,144 sourced from the VCR [46] dataset and 533 from Ego4D [41], covering diverse contextual scenarios as well as real-world human interactions. On average, VAGUE contains seven objects and four people per image. Each image is paired with a direct expression  $p_d$ , an indirect expression  $p_i$ , and four multiple-choice answers, along with relevant meta-information. All textual components are generated using GPT-4o, then refined through extensive human rating, selection, and filtering, ensuring a carefully curated benchmark dataset for testing and advancing multimodal reasoning. We provide detailed benchmark statistics and a diversity analysis in Appendix B.2.
+
+# 4.1. Visual Data Curation
+
+# 4.1.1. Sampling
+
+VCR [46] The VCR dataset consists of 110K movie scenes sourced from the Large Scale Movie Description Challenge [35] and YouTube clips. These images are curated based on an "interestingness" criterion [46], ensuring the presence of at least two people, which promotes interactive scenarios. To prevent redundancy in our dataset, we sample 10K images while carefully avoiding neighboring frames, as adjacent frames exhibit minimal variation. This selection process preserves the contextual diversity of the dataset while maintaining its focus on complex, multi-entity interactions.
+
+Ego4D [41] While VCR provides a wide range of contextual diversity, it often includes artificially composed settings that may not fully capture real-world interactions. To address this, we integrate frames from the Ego4D dataset, which offers a more naturalistic depiction of human interactions. We specifically leverage the AV (Audio-Visual), which indicate conversational exchanges between individuals, to ensure the presence of people in the selected frames. Similar to VCR, we avoid neighboring frames to maintain diversity and filter out heavily blurry images to enhance data quality. This process results in 888 candidate images from 94 videos, which serve as the basis for further text processing.
+
+# 4.1.2. Object Extraction
+
+To ensure the complexity of the visual information, we extract a list of physical objects present in each image using a tagging model, RAM [49]. This step allows us to easily identify scenes with sufficient visual detail. In the case of VCR, many scenes are relatively simple, often containing only a few objects. Therefore, we sort VCR images by the number of detected objects and retained the top 4,000 as candidates for text processing, ensuring that our benchmark primarily consists of rich visual cues in contextually diverse scenes. Please refer to Appendix B.3 for more details.
+
+# 4.1.3. Person Indicator
+
+In our task, we assume that the speaker is outside the scene, viewing the image and talking to a person in it. However, identifying the addressee is not always straightforward, as images mostly contain multiple individuals. While grounding the specific person referenced in the utterance could introduce additional complexity, it is not the primary focus of our evaluation. To clarify the listener, we assign an indicator tag to each person in the image as shown in Fig. 3. For VCR, we use their existing annotations [46], while for Ego4D, where bounding boxes are not fully available, we employ YOLOv11 [18] to detect and annotate humans. While this provides a straightforward method for grounding the target person, it requires models to perform basic Optical Character Recognition (OCR). Therefore, we conduct exper
+
+![](images/9e1c2cc12c6b85ca62f868102b42281b9ec99a2fd7df1e4755d09e2cd23bc692.jpg)  
+Figure 3. Overview of the data generation process. Based on human-defined criteria and instructions, GPT-4o [31] generates initial data, which are then rated and filtered by humans to ensure quality. Since generating high-quality indirect expressions from raw images is challenging, the process follows these steps: generating a direct expression and intention from the image, creating an indirect expression using information from the previous step, and producing answer candidates based on all the information gathered so far. In the answer candidates, FS stands for Fake Scene Understanding, SU stands for Superficial Understanding, and NE stands for Nonexistent Entity. FS evaluates global hallucination, where the model misinterprets visual context from an entirely different image (e.g., an outdoor camping scene with a candle), while NE addresses local hallucination, where only a specific object is replaced with a fabricated one (e.g., curtains).
+
+iments to evaluate the OCR capabilities of the models used in our study. See Appendix B.4 for details.
+
+# 4.2. Multimodal Expression Synthesis
+
+As shown in Fig. 3, VAGUE's textual expressions are generated first by the model and undergo an extensive process of human rating and filtering. We use GPT-4o [31] for all text processing. The instructions used for generating direct  $(p_d)$  and indirect  $(p_i)$  expressions are provided in Appendix B.5, while the instructions for generating answer candidates for multiple-choice questions are detailed in Appendix B.7.
+
+# 4.2.1. Direct Expressions
+
+The direct expression  $p_d$  serves as a crucial foundation for crafting the indirect one  $p_i$ , as both share the same underlying intention. To ensure that  $p_d$  adhere to the principles of relevance and solvability discussed in Sec. 3.2.1, we generate  $p_d$  conditioned on the input image  $I$  and a task prompt that explicitly defines these criteria. During the generation process, we also instruct the model to output a solution triplet in the format: (subject, action, object). Since direct expressions explicitly state their intentions, extracting each component of the solution triplet is straightforward. To maintain consistency with the visual context, the "object" in the triplet is restricted to physical objects we extracted in Sec. 4.1.2.
+
+After generating  $p_d$  for all candidate images, human raters evaluate each prompt based on relevance and solvability, assigning scores from 1 to 5. Only those prompts that receive a rating of 4 or 5 are retained for further use. The detailed rating criteria for human verification and an example of the rating process are provided in Fig. J13.
+
+# 4.2.2. Indirect Expressions
+
+To ensure that the indirect expressions  $p_i$  maintain both ambiguity and fluency while aligning with the true intent  $T$ , we adopt a two-stage process: proposal and selection.
+
+In the initial step, the model is prompted to generate three distinct candidate options. Each candidate follows the criteria outlined in Section 3.2.2, but with explicit instructions to incorporate different linguistic strategies: sarcasm, humor, and meme/idiomatic expressions, respectively. This approach ensures diversity in the generated responses. In the second step, human annotators evaluate the three candidates and select the one that best aligns with the intended indirectness. The selected prompt is then rated on a scale from 1 to 5 based on more specific criteria. Only those prompts that receive a score of 3 or higher are retained for use in the dataset. The detailed rating criteria for human verification and an example of rating process are provided in Fig. J14
+
+# 4.2.3. Counterfactual Choices
+
+Generating high-quality counterfactual choices is crucial. To enable detailed analysis of model weaknesses in multimodal intent disambiguation, we design interpretable counterfactual choices that provide more plausible alternatives.
+
+Fake Scene Understanding The first counterfactual choice is an interpretation that could arise when the model largely misinterprets the image. This process is conducted in two steps. In the first step, a fake caption is generated by assuming an imaginary scene that can be aligned with the indirect expression but is inconsistent with the true intent. The caption of fake scene is then combined with the speaker's
+
+indirect statement to derive the most likely interpretation.
+
+Superficial Understanding The subsequent choice corresponds to an interpretation generated when the model fails to deeply reason about the implicit intent of the sentence and instead relies on surface-level meaning. We enforce the model to focus only on the literal wording, without considering any implied or deeper meaning of the indirect sentence. These answer choices are generated alongside the indirect expression  $p_i$ . During the indirect selection phase, the corresponding superficially understood choice is selected together, maintaining coherence between them.
+
+Nonexistent Entity The last choice arises when the model interprets the text correctly but fails to adequately consider the details of the image, resulting in a plausible yet incorrect choice. This resembles the correct answer in structure, but replaces the key object in the solution with one that does not exist in the image. To prevent the task from becoming too easy by generating highly irrelevant objects, we constrain the substituted object to one that, while absent from the image, is highly expected to be present in the scene and could replace the original entity. To identify such entities, the model is provided with the image as input to choose objects that align with the scene's context. This method ensures that the counterfactual choice leverages the expected coherence between the scene and its potential entities while rigorously testing the model's attention to visual details.
+
+# 5. Experiments
+
+Models We use the following models in our experiments. The detailed descriptions of each model are in Appendix C.
+
+- Phi3.5-Vision-Instruct (4B) [29]  
+- LLaVA Onevision (7B) [23]  
+Qwen2.5-VL-Instruct (7B, 72B) [43]  
+- InternVL-2.5-MPO (8B, 26B) [6]  
+- Idefics2 (8B) [22]  
+LLaVA NeXT Vicuna (13B) [24]  
+- Ovis2 (16B) [28]  
+GPT-4o [31]  
+Gemini 1.5 Pro [12]  
+- InternVL-3 (38B) [42]
+
+# 5.1. MLLMs Benefit from Visual Cues
+
+Our first objective is to assess how effectively MLLMs leverage visual cues to resolve ambiguity in utterances. To this end, we systematically control the level of detail in the visual cues provided to the models and measure their accuracy in inferring the speaker's true intent. Performance is evaluated in both multiple-choice and free-form settings. For clarity, we primarily report multiple-choice accuracy, deferring free-form results to Appendix I.
+
+We consider three levels of visual cues:
+
+- Language Models (LMs) receive no visual input, requiring models to rely on superficial textual priors such as common-sense knowledge of sarcasm or humor to determine intent.  
+- Socratic Models (SMs) [47] use text-only LMs but incorporate short image captions (up to two or three sentences) as additional input. This short generic caption may lack sufficient detail, which may be insufficient for accurately inferring intent. Each SM model generated its own image captions and used them in subsequent processing.  
+- Visual Language Models (VLMs) receive the raw image input, enabling a more direct interpretation of visual cues.
+
+Results The results in Tab. 1 indicate that MLLMs can leverage visual cues, albeit to a limited extent, since SMs and VLMs consistently outperform LMs across all evaluated models. Additionally, more detailed visual input generally improves performance, with VLMs surpassing SMs in most cases, except in the case of proprietary models. This exception is further analyzed in Sec. 5.2. Both the VCR and Ego4D subsets exhibit similar performance trends, demonstrating the generalizability of our findings across both staged and real-world scenarios. Finally, the consistently low performance of LMs further reinforces the validity of our dataset as a multimodal benchmark.
+
+# 5.2. Analysis on Failure Modes
+
+Here, we examine the ways in which models fail to infer the true intent and how these failure patterns vary with the level of visual cues provided. As shown in Fig. 3, our multiple-choice questions include three distinct types of incorrect answer candidates. We assess the model's raw visual understanding using the Fake Scene Understanding (FS) and Nonexistent Entity (NE) candidates, which test whether the model can correctly interpret the scene without being misled by fabricated or nonexistent elements. Conversely, the Superficial Understanding (SU) candidate evaluates the model's reasoning ability, testing whether it can go beyond surface-level perception to infer intent. We provide the full table of failure modes in the MCQ setting in Appendix I
+
+Results Figure 4 illustrates how frequently each model selects different types of incorrect answers instead of the correct intent. Among the error types, Superficial Understanding (SU) is the most common. This indicates that while models generally succeed in recognizing basic visual details, they often fail to reason deeply about those visual cues to accurately infer the underlying intent of the speaker. However, proprietary models exhibit fewer SU-related errors, indicating stronger reasoning capabilities.
+
+Moreover, stronger visual cues improve accuracy across all models and failure types. This improvement highlights the crucial role of visual conditioning in reducing both vision-based errors (FS and NE) and reasoning-related failures (SU).
+
+<table><tr><td rowspan="2">Model</td><td colspan="3">VAGUE-VCR</td><td colspan="3">VAGUE-Ego4D</td></tr><tr><td>LM (L)</td><td>SM (L+V)</td><td>VLM (L+V)</td><td>LM (L)</td><td>SM (L+V)</td><td>VLM (L+V)</td></tr><tr><td>Phi3.5-Vision-Instruct (4B)</td><td>26.6</td><td>35.3 (↑ 8.7)</td><td>46.0 (↑ 19.4)</td><td>22.5</td><td>31.1 (↑ 8.6)</td><td>42.4 (↑ 19.9)</td></tr><tr><td>LLaVA-Onevision (7B)</td><td>13.1</td><td>29.4 (↑ 16.3)</td><td>43.1 (↑ 30.0)</td><td>11.3</td><td>29.5 (↑ 18.2)</td><td>43.2 (↑ 31.9)</td></tr><tr><td>Qwen2.5-VL-Instruct (7B)</td><td>11.1</td><td>25.6 (↑ 14.5)</td><td>46.8 (↑ 35.7)</td><td>9.8</td><td>28.0 (↑ 18.2)</td><td>48.4 (↑ 38.6)</td></tr><tr><td>InternVL-2.5-MPO (8B)</td><td>23.0</td><td>48.4 (↑ 25.4)</td><td>63.9 (↑ 40.9)</td><td>24.2</td><td>54.0 (↑ 29.8)</td><td>66.8 (↑ 42.6)</td></tr><tr><td>Idefics2 (8B)</td><td>13.9</td><td>21.1 (↑ 7.2)</td><td>58.7 (↑ 44.8)</td><td>14.8</td><td>18.2 (↑ 3.4)</td><td>58.3 (↑ 43.5)</td></tr><tr><td>LLaVA-NeXT-vicuna (13B)</td><td>24.2</td><td>37.2 (↑ 13)</td><td>46.4 (↑ 22.2)</td><td>20.3</td><td>34.1 (↑ 13.8)</td><td>52.5 (↑ 32.2)</td></tr><tr><td>Ovis2 (16B)</td><td>21.9</td><td>23.8 (↑ 1.9)</td><td>24.5 (↑ 3.6)</td><td>20.5</td><td>25.3 (↑ 4.8)</td><td>25.7 (↑ 5.2)</td></tr><tr><td>InternVL-2.5-MPO (26B)</td><td>21.2</td><td>48.5 (↑ 27.3)</td><td>63.7 (↑ 42.5)</td><td>21.8</td><td>55.2 (↑ 33.4)</td><td>68.7 (↑ 46.9)</td></tr><tr><td>InternVL-3 (38B)</td><td>24.8</td><td>47.2 (↑ 22.4)</td><td>63.6 (↑ 38.8)</td><td>18.0</td><td>47.5 (↑ 29.5)</td><td>59.8 (↑ 41.8)</td></tr><tr><td>Qwen2.5-VL-Instruct (72B)</td><td>29.6</td><td>55.6 (↑ 26.0)</td><td>74.2 (↑ 44.6)</td><td>26.8</td><td>59.3 (↑ 32.5)</td><td>69.8 (↑ 43.0)</td></tr><tr><td>GPT-4o</td><td>46.4</td><td>69.5 (↑ 23.1)</td><td>65.1 (↑ 18.7)</td><td>48.2</td><td>67.5 (↑ 19.3)</td><td>63.6 (↑ 15.3)</td></tr><tr><td>Gemini-1.5-Pro</td><td>43.2</td><td>62.4 (↑ 19.2)</td><td>60.6 (↑ 17.4)</td><td>40.3</td><td>60.6 (↑ 20.3)</td><td>60.6 (↑ 20.3)</td></tr></table>
+
+Table 1. Experiments on the Multimodal Intention Disambiguation (MID) task with varying levels of visual cues. We report the accuracy  $(\%)$  of the Multiple-Choice Question.  $\uparrow$  indicates the performance gain from visual cues, i.e. increment compared to the LM setting.(L) denotes the use of language input only, while  $(\mathrm{L} + \mathrm{V})$  indicates the incorporation of visual cues. The noticeable increase in accuracy across LM, SM, and VLM demonstrates that the introduction of detailed visual cues is beneficial for the task.
+
+![](images/f063bf44d3c5ca75ea7e8831036cb2b6bff129c5ecbd71bdc54b08faf9ac3d72.jpg)  
+Figure 4. We present a bar plot to analyze the distribution of incorrect answer choices selected by each model. Each number represents how frequently a given choice was selected from the 1,677 items in the dataset. The counterfactual choice categories are FS (Fake Scene Understanding), SU (Superficial Understanding), and NE (Nonexistent Entity). We use distinct colors to represent LM (Language Models), SM (Socratic Models), and VLM (Visual-Language Models).
+
+Notably, proprietary models perform better with captioned inputs (SM) than with raw images (VLM). A closer examination of Fig. 4 reveals that this discrepancy arises from vision-based failures (FS and NE) rather than reasoning-centric failure (SU). This indicates that their captioning ability potentially allows them to obtain more sophisticated information while reducing hallucination in the images. The supporting experiments and explanations are provided in Appendix F.
+
+# 5.3. Comparison with human
+
+To validate our benchmark and establish an upper bound for performance, we assess human accuracy, highlighting the gap between existing models and human capability. This evaluation follows the multiple-choice setup within the VLM setting, using a subset of 400 samples. As shown
+
+in Tab. 2, human performance reaches  $94\%$ , demonstrating near-perfect accuracy. Although proprietary models and some large-sized models show a decent performance, a notable performance gap ( $\sim 20\%$ ) still exists compared to human evaluators. This result underscores the significant gap between AI models' multimodal reasoning capabilities and human-level understanding when inferring hidden intent, suggesting that visual perception alone, even when accurate, is insufficient without deeper cognitive integration. This performance gap is due to the models' tendency to rely on surface-level text rather than understanding deeper visual-textual implications. Thus, advancing multimodal reasoning likely requires models to integrate higher-order cognitive processes, like commonsense reasoning and pragmatic understanding, into visual interpretation tasks. Refer
+
+<table><tr><td>Model</td><td>Acc (%)</td><td>FS</td><td>SU</td><td>NE</td><td>Correct</td></tr><tr><td>Ovis2 (16B)</td><td>23.0</td><td>119</td><td>171</td><td>18</td><td>92</td></tr><tr><td>LLaVA-Onevision (7B)</td><td>41.0</td><td>43</td><td>183</td><td>10</td><td>164</td></tr><tr><td>Phi3.5-Vision-Instruct (4B)</td><td>44.3</td><td>60</td><td>132</td><td>31</td><td>177</td></tr><tr><td>Qwen2.5-VL-Instruct (7B)</td><td>47.0</td><td>47</td><td>152</td><td>13</td><td>188</td></tr><tr><td>LLaVA-NeXT-icuna (13B)</td><td>48.0</td><td>48</td><td>143</td><td>17</td><td>192</td></tr><tr><td>Idefics2 (8b)</td><td>57.0</td><td>28</td><td>120</td><td>24</td><td>228</td></tr><tr><td>Gemini-1.5-Pro</td><td>60.3</td><td>60</td><td>73</td><td>26</td><td>241</td></tr><tr><td>InternVL-3 (38B)</td><td>61.5</td><td>38</td><td>94</td><td>22</td><td>246</td></tr><tr><td>InternVL-2.5-MPO (8B)</td><td>61.8</td><td>42</td><td>95</td><td>16</td><td>247</td></tr><tr><td>GPT-4o</td><td>62.3</td><td>61</td><td>63</td><td>27</td><td>249</td></tr><tr><td>InternVL-2.5-MPO (26B)</td><td>63.0</td><td>36</td><td>101</td><td>11</td><td>252</td></tr><tr><td>Qwen2.5-VL-Instruct (72B)</td><td>72.3</td><td>39</td><td>60</td><td>12</td><td>289</td></tr><tr><td>Human</td><td>94.0</td><td>12</td><td>4</td><td>8</td><td>374</td></tr></table>
+
+Table 2. Performance across models and humans on a sampled set of 400 questions. The results show that humans outperform models by a margin of over  $20\%$ .  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Type</td><td rowspan="2">Acc (%)</td><td colspan="3">Incorrect count</td></tr><tr><td>FS</td><td>SU</td><td>NE</td></tr><tr><td rowspan="4">GPT-4o</td><td>SM</td><td>68.9</td><td>165</td><td>240</td><td>117</td></tr><tr><td>SM+CoT</td><td>69.5 (↑ 0.6)</td><td>165</td><td>241</td><td>105</td></tr><tr><td>VLM</td><td>64.6</td><td>226</td><td>226</td><td>141</td></tr><tr><td>VLM+CoT</td><td>66.4 (↑ 1.8)</td><td>162</td><td>156</td><td>85</td></tr><tr><td rowspan="4">Gemini-1.5-Pro</td><td>SM</td><td>61.8</td><td>176</td><td>374</td><td>83</td></tr><tr><td>SM+CoT</td><td>61.0 (↓ 0.8)</td><td>190</td><td>367</td><td>94</td></tr><tr><td>VLM</td><td>60.6</td><td>249</td><td>264</td><td>145</td></tr><tr><td>VLM+CoT</td><td>64.4 (↑ 3.8)</td><td>213</td><td>267</td><td>117</td></tr></table>
+
+Table 3. Result of Chain-of-Thought (CoT) experiments on proprietary models, in both SM and VLM settings.  $\uparrow$  and  $\downarrow$  indicate an increase and decrease in accuracy when zero-shot CoT is applied.
+
+to Appendix E for details on the selected subset and human evaluation setup.
+
+# 5.4. Chain-of-Thought Experiments
+
+Given the strong reasoning demands of multimodal intent deduction, we further explore the effectiveness of Chain-of-Thought (CoT) prompting [44] in enhancing the reasoning capabilities of MLLMs. The CoT prompt templates, provided in Fig. J21 and Fig. J22, are designed to explicitly ground the reasoning process, reducing hallucinations. While our main results focus on proprietary models, owing to their superior suitability for zero-shot CoT reasoning, we also present experiments and analyses for open-source models ranging from 4B to 72B in Appendix G.
+
+Results As shown in Tab. 3, CoT prompting improves performance for raw image inputs (VLM), while showing no clear trend and remaining at a similar level for image caption inputs (SM). One possible explanation for this discrepancy is that CoT primarily enhances reasoning by improving
+
+grounding and reducing hallucinations. Since image captions inherently contain fewer hallucinations, SMs see some benefit from CoT prompting albeit at the cost of reduced detail. Additionally, the performance improvements observed with CoT prompting are consistent across different types of false answer candidates, suggesting a generalizable effect in enhancing reasoning quality.
+
+# 6. Conclusion
+
+We present VAGUE (Visual Contexts CIArify ambiGUous Expressions), a new benchmark aimed at assessing models' ability to interpret nuanced communication in complex multimodal scenarios. Our results show that models benefit from visual information when inferring the underlying intention of indirect expressions, as evidenced by their improved performance with increasing levels of visual cues. However, a significant disparity persists between machine capabilities and humans. To gain deeper insights into model inaccuracies, we design multiple-choice questions that explicitly address failure points, enabling a systematic and quantitative evaluation of the reasons behind performance. The primary challenge identified is the tendency of multimodal models to inadequately integrate visual cues, relying instead on the literal interpretation of textual information. This shortcoming highlights the need for further research, and we anticipate that VAGUE will open up promising avenues for developing systems capable of deeper multimodal reasoning to enhance AI's ability to engage in human-like interactions.
+
+# 7. Limitations
+
+First, we acknowledge that there are cultural and linguistic limitations. We incorporate sarcastic, humorous, and idiomatic implicatures in generating indirect expressions. However, since the initial data drafts are created by a model (GPT-4o [31]), they may reflect cultural biases present in its training data. To prevent any potential ethical issues, all human annotators are instructed to remove any content considered problematic or discriminatory during the rating and filtering process. Also, all textual expressions in VAGUE are limited to English. Therefore, we encourage future exploration of indirect expressions across diverse languages and cultures. The second limitation pertains to the dependency of certain meta-information on the quality of the parent dataset and the performance of the models utilized during the dataset generation. We observe that bounding box annotations from YOLOv11 [18] are occasionally duplicated, leading to a reported number of people higher than actually present. Likewise, the tagging model RAM [49] sometimes misidentified objects. Therefore, we remove any corrupted instances that could undermine the integrity of the task during the human rating and filtering process.
+
+# References
+
+[1] Ashwin Baluja. Text is not all you need: Multimodal prompting helps llms understand humor. arXiv preprint arXiv:2412.05315, 2024. 3  
+[2] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 1  
+[3] Matteo Bortoletto, Constantin Ruhdorfer, Lei Shi, and Andreas Bulling. Explicit modelling of theory of mind for belief prediction in nonverbal social interactions. In Proc. 27th European Conference on Artificial Intelligence (ECAI), pages 1-8, 2024. 2  
+[4] Tuhin Chakrabarty, Arkadiy Saakyan, Debanjan Ghosh, and Smaranda Muresan. FLUTE: Figurative language understanding through textual explanations. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7139-7159, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 2  
+[5] Zhawnen Chen, Tianchun Wang, Yizhou Wang, Michal Kosinski, Xiang Zhang, Yun Fu, and Sheng Li. Through the theory of mind's eye: Reading minds with multimodal video large language models. arXiv preprint arXiv:2406.13763, 2024. 2  
+[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 6, 3  
+[7] Christine Cuskley, Rebecca Woods, and Molly Flaherty. The limitations of large language models for understanding human language and cognition. Open Mind, 8:1058-1083, 2024. 2  
+[8] Hang Du, Guoshun Nan, Sicheng Zhang, Binzhu Xie, Junrui Xu, Hehe Fan, Qimei Cui, Xiaofeng Tao, and Xudong Jiang. Docmsu: A comprehensive benchmark for document-level multimodal sarcasm understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17933-17941, 2024. 2  
+[9] Jiafei Duan, Samson Yu, Nicholas Tan, Li Yi, and Cheston Tan. Boss: A benchmark for human belief prediction in object-context scenarios. arXiv preprint arXiv:2206.10665, 2022. 2  
+[10] Alessandro Favero, Luca Zancato, Matthew Trager, Siddharth Choudhary, Pramuditha Perera, Alessandro Achille, Ashwin Swaminathan, and Stefano Soatto. Multi-modal hallucination control by visual information grounding. CVPR 2024, 2024. 4  
+[11] Kanishk Gandhi, Jan-Philipp Fraenkel, Tobias Gerstenberg, and Noah D. Goodman. Understanding social reasoning in language models with language models. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023), Datasets and Benchmarks Track, 2023. Dataset and Benchmarks Track. 2  
+[12] Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 6, 3
+
+[13] Md Kamrul Hasan, Wasifur Rahman, AmirAli Bagher Zadeh, Jianyuan Zhong, Md Iftekhar Tanveer, Louis-Philippe Morency, and Mohammed (Ehsan) Hoque. UR-FUNNY: A multimodal language dataset for understanding humor. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2046–2056, Hong Kong, China, 2019. Association for Computational Linguistics. 2  
+[14] Jack Hessel, Ana Marasovic, Jena D. Hwang, Lillian Lee, Jeff Da, Rowan Zellers, Robert Mankoff, and Yejin Choi. Do androids laugh at electric sheep? humor "understanding" benchmarks from the new yorker caption contest. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 688-714, Toronto, Canada, 2023. Association for Computational Linguistics. 2  
+[15] EunJeong Hwang and Vered Shwartz. Memecap: A dataset for captioning and interpreting memes. arXiv preprint arXiv:2305.13703, 2023. 3  
+[16] Prince Jha, Krishanu Maity, Raghav Jain, Apoorv Verma, Sriparna Saha, and Pushpak Bhattacharyya. Meme-ingful analysis: Enhanced understanding of cyberbullying in memes through multimodal explanations. arXiv preprint arXiv:2401.09899, 2024.3  
+[17] Chuanyang Jin, Yutong Wu, Jing Cao, Jiannan Xiang, Yen Ling Kuo, Zhiting Hu, Tomer Ullman, Antonio Torralba, Joshua Tenenbaum, and Tianmin Shu. MMToM-QA: Multimodal theory of mind question answering. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16077-16102, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2  
+[18] Rahima Khanam and Muhammad Hussain. Yolov11: An overview of the key architectural enhancements, 2024. 4, 8  
+[19] Shivani Kumar, Atharva Kulkarni, Md Shad Akhtar, and Tan-moy Chakraborty. When did you become so smart, oh wise one?! sarcasm explanation in multi-modal multi-party dialogues. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5956-5968, Dublin, Ireland, 2022. Association for Computational Linguistics. 2  
+[20] Shivani Kumar, Ishani Mondai, Md Shad Akhtar, and Tanmoy Chakraborty. Explaining (sarcastic) utterances to enhance affect understanding in multimodal dialogues. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. AAAI Press, 2023. 2  
+[21] Yash Kumar Lal and Mohaddeseh Bastan. SBU figures it out: Models explain figurative language. In Proceedings of the 3rd Workshop on Figurative Language Processing (FLP), pages 143-149, Abu Dhabi, United Arab Emirates (Hybrid), 2022. Association for Computational Linguistics. 2  
+[22] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language mod
+
+els? Advances in Neural Information Processing Systems, 37: 87874-87907, 2025. 6, 3  
+[23] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6, 3  
+[24] 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, 2024. 6, 3  
+[25] Kevin Y. Li, Sachin Goyal, Joao D. Semedo, and J. Zico Kolter. Inference optimal vlms need fewer visual tokens and more parameters, 2025. 4  
+[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision – ECCV 2014, pages 740–755, Cham, 2014. Springer International Publishing. 1, 3  
+[27] Annie Louis, Dan Roth, and Filip Radlinski. "I'd rather just go to bed": Understanding indirect answers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7411-7425, Online, 2020. Association for Computational Linguistics. 2  
+[28] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 6, 3  
+[29] Microsoft. Phi-3 technical report: A highly capable language model locally on your phone, 2024. 6, 3  
+[30] Maryam Sadat Mirzaei, Kourosh Meshgi, and Satoshi Sekine What is the real intention behind this question? dataset collection and intention classification. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13606-13622, Toronto, Canada, 2023. Association for Computational Linguistics. 2  
+[31] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5, 6, 8, 3  
+[32] Kun Ouyang, Liqiang Jing, Xuemeng Song, Meng Liu, Yu-peng Hu, and Liqiang Nie. Sentiment-enhanced graph-based sarcasm explanation in dialogue. arXiv:2402.11414, 2024. 2  
+[33] 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, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 3, 4  
+[34] David Premack and Guy Woodruff. Does the chimpanzee have a theory of mind? Behavioral and brain sciences, 1(4): 515-526, 1978. 2  
+[35] Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket Tandon, Chris Pal, Hugo Larochelle, Aaron Courville, and Bernt Schiele. Movie description. International Journal of Computer Vision, 2017. 4  
+[36] Melanie Sclar, Sachin Kumar, Peter West, Alane Suhr, Yejin Choi, and Yulia Tsvetkov. Minding language models' (lack of) theory of mind: A plug-and-play multi-character belief
+
+tracker. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), pages 13960-13980, Toronto, Canada, 2023. Association for Computational Linguistics. 2  
+[37] Haojun Shi, Suyu Ye, Xinyu Fang, Chuanyang Jin, Leyla Isik, Yen-Ling Kuo, and Tianmin Shu. Muma-tom: Multi-modal multi-agent theory of mind, 2024. 2  
+[38] Settaluri Sravanthi, Meet Doshi, Pavan Tankala, Rudra Murthy, Raj Dabre, and Pushpak Bhattacharyya. PUB: A pragmatics understanding benchmark for assessing LLMs' pragmatics capabilities. In Findings of the Association for Computational Linguistics: ACL 2024, pages 12075-12097, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2  
+[39] Chuandong Su, Fumiyo Fukumoto, Xiaoxi Huang, Jiyi Li, Rongbo Wang, and Zhiqun Chen. DeepMet: A reading comprehension paradigm for token-level metaphor detection. In Proceedings of the Second Workshop on Figurative Language Processing, pages 30-39, Online, 2020. Association for Computational Linguistics. 2  
+[40] Junya Takayama, Tomoyuki Kajiwara, and Yuki Arase. DIRECT: Direct and indirect responses in conversational text corpus. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1980-1989, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2  
+[41] Ego4d Team. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18995-19012, 2022. 2, 4  
+[42] OpenGVLab Team. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models, 2025. 6, 3  
+[43] Qwen Team. Qwen2.5-v1 technical report, 2025. 6, 3  
+[44] 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. Advances in neural information processing systems, 35:24824-24837, 2022. 8  
+[45] Alex Wilf, Leena Mathur, Sheryl Mathew, Claire Ko, Youssof Kebe, Paul Pu Liang, and Louis-Philippe Morency. Social-iq 2.0 challenge: Benchmarking multimodal social understanding. https://github.com/abwilf/Social-IQ-2.0-Challenge, 2023.2  
+[46] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6713-6724, 2018. 2, 4, 1  
+[47] Andy Zeng, Maria Attarian, Krzysztof Marcin Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael S Ryoo, Vikas Sindhwani, Johnny Lee, et al. Socratic models: Composing zero-shot multimodal reasoning with language. In The Eleventh International Conference on Learning Representations, 2023. 2, 6  
+[48] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations, 2020. 4
+
+[49] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize anything: A strong image tagging model. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1724-1732, 2024. 4, 8, 1

2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba7a1da700708ff8ed1331c24bb4074d291f99e3a6ce7b95ac436f4c672d79ee
+size 451798

2025/VALLR_ Visual ASR Language Model for Lip Reading/88644d63-e953-49ce-b2d3-f0418686524b_content_list.json

@@ -0,0 +1,1825 @@
+[
+    {
+        "type": "text",
+        "text": "VALLR: Visual ASR Language Model for Lip Reading",
+        "text_level": 1,
+        "bbox": [
+            222,
+            130,
+            776,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Marshall Thomas Edward Fish Richard Bowden University of Surrey",
+        "bbox": [
+            287,
+            181,
+            709,
+            217
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "mt00893@surrey.ac.uk edward.fish@surrey.ac.uk r.bowden@surrey.ac.uk",
+        "bbox": [
+            181,
+            219,
+            805,
+            233
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            248,
+            256,
+            326,
+            272
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Lip Reading, or Visual Automatic Speech Recognition (V-ASR), is a complex task requiring the interpretation of spoken language exclusively from visual cues, primarily lip movements and facial expressions. This task is especially challenging due to the absence of auditory information and the inherent ambiguity when visually distinguishing phonemes that have overlapping visemes, where different phonemes appear identical on the lips. Current methods typically attempt to predict words or characters directly from these visual cues, but this approach frequently encounters high error rates due to coarticulation effects and viseme ambiguity. We propose a novel two-stage, phoneme-centric framework for Visual Automatic Speech Recognition (V-ASR) that addresses these longstanding challenges. First, our model predicts a compact sequence of phonemes from visual inputs using a Video Transformer with a CTC head, thereby reducing the task complexity and achieving robust speaker invariance. This phoneme output then serves as the input to a fine-tuned Large Language Model (LLM), which reconstructs coherent words and sentences by leveraging broader linguistic context. Unlike existing methods that either predict words directly or rely on large-scale multimodal pre-training, our approach explicitly encodes intermediate linguistic structure while remaining highly data efficient. We demonstrate state-of-the-art performance on two challenging datasets, LRS2 and LRS3, where our method achieves significant reductions in Word Error Rate (WER) achieving a SOTA WER of 18.7 on LRS3 despite using  $99.4\\%$  less labelled video data than the next best approach. Code is available here: https://gitlab.surrey.ac.uk/",
+        "bbox": [
+            86,
+            297,
+            485,
+            765
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            91,
+            800,
+            220,
+            814
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Lip Reading, or Visual Automatic Speech Recognition (V-ASR), involves interpreting spoken language from visual cues such as lip movements and facial expressions. As a natural skill, humans use it to supplement auditory information, and as a technology, it has profound potential for en",
+        "bbox": [
+            89,
+            824,
+            483,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/9ae34c4388a4f19abe6b22865a145bffff65591cbe287bfbeb007bdb92f19060.jpg",
+        "image_caption": [
+            "Figure 1. Comparison between different models' performances in WER for Visual Automatic Speech Recognition on the LRS3 dataset [3] when compared with the amount of labelled training data. Circle size (green) denotes scale of pre-training data, while fine-tuned models (grey) are not pre-trained. Our model (orange) outperforms all existing approaches with just 30 hours of video training data, no self-supervised pre-training, and without the requirement for additional labeled visual data during fine-tuning."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            258,
+            913,
+            460
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "hancing accessibility, particularly for the Deaf and hard-of-hearing communities, as well as for applications in noisy or privacy-sensitive environments [1]. Despite substantial advancements in ASR, the visual-only interpretation of speech remains an unresolved challenge due to inherent uncertainties in lip movements and their complex temporal dynamics. A central difficulty in V-ASR stems from the inherent ambiguity of visemes, the visual equivalents of phonemes, which often appear nearly identical for different sounds (e.g., 'p' and 'b') [1, 7]. Further complicating this task are coarticulation effects [33], where adjacent sounds blur one another's articulation, making phoneme boundaries visually unclear. Although some recent methods have focussed on robust intermediary representations to deal with these cases, such as leveraging subword-level predictions [41], or quantized latent representations [13], these approaches can still fail to capture broader contextual cues, limiting their ability to resolve visually similar phonemes that might otherwise be",
+        "bbox": [
+            511,
+            628,
+            908,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV 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": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "2846",
+        "bbox": [
+            482,
+            944,
+            516,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/85b1f3f5dae7dfb135fece6cf0a7069536516d7e82e8e5024ee5d2691a090d08.jpg",
+        "image_caption": [
+            "Figure 2. An overview of our approach. First we extract facial regions from 16 frames of input video. We apply random pixel masking at  $50\\%$  probability, add positional embedding, and encode visual features via a vision transformer encoder (ViT-base [14]). We implement temporal downsampling via 1D convolution and then use a CTC linear head to predict sequences of phonemes. During training, we also fine-tune an LLM to reconstruct sentences from phonemes with a text-only dataset [52]. During inference, the phonemes from the CTC head are processed via the LLM to reconstruct the predicted text. This can be performed end-to-end or in two stages, depending on available resources."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            89,
+            85,
+            906,
+            333
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "distinguishable through sentence-level semantics.",
+        "bbox": [
+            89,
+            463,
+            416,
+            478
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "One approach to this challenge is to leverage large language models (LLM's) and attempt to map raw lip movements directly to sentences, thus bypassing the issue of viseme ambiguity [58, 59]. However, bridging the gap between high-dimensional, unstructured visual inputs and detailed textual representations is resource intensive and typically requires huge datasets, or specialized architectures. Moreover, these purely end-to-end pipelines often lack an interpretable intermediary step, making it unclear how the model is handling viseme ambiguity at a finer linguistic granularity. This can lead to hallucination effects, which become more critical in lip reading technologies designed for accessibility, or security applications.",
+        "bbox": [
+            88,
+            493,
+            482,
+            690
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our two-stage, phoneme-centric method addresses both issues. In the first stage, we map short windows of video frames to a discrete, interpretable phoneme sequence. This intermediate representation is significantly easier to learn and more robust against speaker or environmental variations, as phonemes abstract away speaker-specific attributes. Subsequently, in the second stage, we fine-tune an LLM for the task of reconstructing sentences from phonemes. By segmenting the problem, we combine the interpretability and efficiency of phoneme-based prediction with the sophisticated contextual reasoning of modern LLMs. This phoneme  $\\rightarrow$  sentence pipeline also aligns with psycholinguistic models of speech perception, where",
+        "bbox": [
+            88,
+            703,
+            482,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "phoneme-level processing naturally precedes lexical access [31, 32].",
+        "bbox": [
+            511,
+            463,
+            906,
+            493
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Several advantages emerge from this design:",
+        "bbox": [
+            511,
+            501,
+            808,
+            516
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(i) Compact Target Space: Predicting only 38 phoneme classes avoids learning entire word-level vocabularies, making the model more data-efficient and simpler to train.",
+            "(ii) LLM-Enhanced Accuracy: A large language model, pre-trained for phoneme  $\\rightarrow$  sentence reconstruction, removes the requirement of direct word-level predictions from RGB data, reducing the need for complex multimodal training.",
+            "(iii) Data Availability: Exploiting easily available phoneme  $\\rightarrow$  sentence text pairs eliminates reliance on extensive lip-reading video data for pre-training.",
+            "(iv) Interpretability & Error Analysis: The two-stage design yields an explicit phoneme-level output, constraining recognition errors to local phoneme segments, which the LLM subsequently repairs at the word level reducing errors at a sentence level."
+        ],
+        "bbox": [
+            496,
+            523,
+            906,
+            779
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            511,
+            792,
+            653,
+            806
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Here, we review key categories of approaches and highlight their advantages and limitations in the context of our proposed phoneme-centric, language-aware approach.",
+        "bbox": [
+            511,
+            816,
+            906,
+            863
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Lip Reading: Early methods predominantly relied on statistical models such as Hidden Markov Models (HMMs)",
+        "bbox": [
+            511,
+            869,
+            906,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "2847",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "to aggregate temporal sequences of lip movements alongside hand-crafted feature extractors [1, 9, 35, 36, 40, 62]. With the advent of deep learning, coupled with the availability of large-scale datasets such as Grid [12], LRW [11], LRS2 [47], and LRS3 [3], significant progress has been made in recent years. Initially, Convolutional Neural Networks (CNNs) were applied to the task of word recognition [11], with additional temporal processing via RNN's [30, 54, 55]. With greater compute came the possibility to interpret full sentences with approaches applying Automatic Speech Recognition (ASR) methodologies to the Visual Automatic Speech Recognition (V-ASR) task. These included sequence-to-sequence models [2], CTC approaches [5, 46], and hybrid models [37, 38]. These models improved recognition accuracy by learning hierarchical representations of visual features and modelling temporal dependencies. The application of transformer models [14, 21, 44, 50, 53] in V-ASR brought further performance gains in combination with extensive self-supervised pretraining. Methods such as AV-HuBERT [45] and LiteVSR [24] leveraged self-supervised learning to learn robust representations from large amounts of unlabelled video data.",
+        "bbox": [
+            89,
+            90,
+            483,
+            424
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "AV-ASR vs V-ASR: Lip reading approaches can be separated into Visual Automatic Speech Recognition (V-ASR) and Audio-Visual Automatic Speech Recognition (AV-ASR). In the AV-ASR task, labelled audio data is available at training time which can be used to guide the visual encoder in various ways. These can include generating synthetic additional pre-training data [25, 27], or simply fusing audio-visual input features [44-46, 57]. Other methods have used knowledge distillation to inductively transfer knowledge from pre-trained ASR models to visual encoders [4, 28]. However, these methods depend on audio data or pre-trained ASR models, which may not be available or reliable in scenarios where lip reading is most needed, such as noisy environments. V-ASR, on the other hand, is a more challenging task that relies only on visual data in the training stage [1, 56, 60, 62]. Recent methods have also shown the effectiveness of adapting visual inputs to pre-trained ASR models [13, 24, 42] without the need for audio data during training. This method is particularly powerful since the ASR model already has the ability to reconstruct latent representations to words from the pre-training task, however it relies on large amounts of additional data to learn robust visual representations and map them to the audio latent space [42].",
+        "bbox": [
+            89,
+            435,
+            483,
+            799
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Lip Reading with LLMs: Recent approaches have explored visual attention mechanisms for sub-word units [15, 39, 41, 48] to capture fine-grained linguistic features in lip reading. However, reconstructing these units into coherent sentences remains challenging for networks lacking robust contextual and semantic reasoning capabilities. Early",
+        "bbox": [
+            89,
+            809,
+            483,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "phoneme-based methods [16, 49] similarly struggled with effective decoding to word-level representations and these methods were outpaced in performance by end-to-end approaches [5] which could leverage extensive data. As such they have not been fully explored in the context of modern architectures and techniques.",
+        "bbox": [
+            511,
+            90,
+            905,
+            181
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For example, recent approaches which have integrated Large Language Models (LLMs) [43, 51] into lip-reading pipelines [58, 59] have started from mapping visual features directly to LLM text embeddings. However, by omitting explicit intermediate representations, these approaches inherit phonetic ambiguities that propagate through the network. This manifests as increased word error rates and reduced robustness, thus requiring substantially more training data to achieve generalization.",
+        "bbox": [
+            511,
+            189,
+            906,
+            325
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our work re-examines phonemes as an interpretable discrete representation bridging visual and linguistic domains. Rather than forcing LLMs to perform visual  $\\rightarrow$  sentence translation requiring expensive visual-text alignment – we constrain the LLM's role to phoneme  $\\rightarrow$  sentence reconstruction. This decomposition offers three key advantages: (1) Phoneme-to-text translation is a well-constrained task requiring only lightweight LLM fine-tuning using abundant textual corpora; (2) Speaker-independent phoneme representations eliminate the need for visual adaptation in the language model; (3) The modular architecture prevents error propagation between visual and linguistic domains while maintaining interpretability.",
+        "bbox": [
+            511,
+            334,
+            908,
+            531
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            513,
+            544,
+            604,
+            559
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we present our two-stage phoneme-centric approach to visual-only lip reading. Figure 2 provides an overview of the framework. Our goal is to learn a function",
+        "bbox": [
+            511,
+            569,
+            905,
+            614
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} f: \\quad \\mathbf {X} = \\left\\{x _ {1}, x _ {2}, \\dots , x _ {T} \\right\\} \\\\ \\longmapsto \\mathbf {P h} = \\left\\{p h _ {1}, p h _ {2}, \\dots , p h _ {m} \\right\\} \\tag {1} \\\\ \\longmapsto \\mathbf {S} = \\left\\{s _ {1}, s _ {2}, \\dots , s _ {M} \\right\\}. \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            566,
+            627,
+            905,
+            681
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{X} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$  is a video sequence of  $T$  frames (with height  $H$  and width  $W$ ),  $\\mathbf{Ph}$  denotes the phoneme sequence, and  $\\mathbf{S}$  is the reconstructed sentence. Specifically, we decompose the task into:",
+        "bbox": [
+            511,
+            693,
+            905,
+            755
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Video  $\\rightarrow$  Phoneme: Map the sequence of video frames to phonemes using a Vision Transformer and CTC head.",
+            "2. Phoneme  $\\rightarrow$  Sentence: Convert phonemes to a coherent word sequence with a fine-tuned Large Language Model (LLM)."
+        ],
+        "bbox": [
+            517,
+            763,
+            905,
+            838
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The model comprises four primary components:",
+        "bbox": [
+            511,
+            847,
+            834,
+            863
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "- Visual Feature Extractor: Captures relevant features from each frame via a Vision Transformer.",
+        "bbox": [
+            511,
+            869,
+            905,
+            898
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "2848",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Adapter Network with Temporal Downsampling: Reduces sequence length to accommodate CTC alignment.",
+            "- CTC Head: Predicts phoneme sequences without requiring explicit temporal labels.",
+            "- LLM for Phoneme-to-Sentence Reconstruction: Translates predicted phonemes into final word sequences."
+        ],
+        "bbox": [
+            89,
+            90,
+            480,
+            181
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1. Pre-Processing Pipeline",
+        "text_level": 1,
+        "bbox": [
+            89,
+            191,
+            307,
+            208
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Each frame  $x_{i}$  is first passed through a face-detection model [26] to identify and crop the speaker's face region, centered on the speaker's lips. The detected region is then cropped, resized to  $224 \\times 224$ , and normalized for batch processing.",
+        "bbox": [
+            89,
+            214,
+            480,
+            289
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Video Transformer",
+        "text_level": 1,
+        "bbox": [
+            89,
+            299,
+            272,
+            314
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We employ a Vision Transformer (ViT) [44] to encode the spatio-temporal information in the face region. Let ViT(\\cdot) denote this encoding function. The input  $\\mathbf{X} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$  is chunked into patches, embedded, and equipped with positional encodings before passing through multiple transformer blocks. We obtain:",
+        "bbox": [
+            89,
+            321,
+            480,
+            411
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {Z} = \\operatorname {V i T} (\\mathbf {X}) \\in \\mathbb {R} ^ {T \\times D}, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            194,
+            422,
+            480,
+            441
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $D$  is the transformer output dimension per frame. During training, we randomly mask portions of the input patches to promote robust feature learning and improve generalization.",
+        "bbox": [
+            89,
+            454,
+            482,
+            513
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Adapter Network with Temporal Downsampling",
+        "text_level": 1,
+        "bbox": [
+            89,
+            523,
+            482,
+            556
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The sequence length  $T$  of the ViT output can be prohibitively large for the subsequent CTC module. To reduce the temporal dimension, we apply a sequence of 1D convolutions and pooling layers, as shown in Table ?? Formally,",
+        "bbox": [
+            89,
+            561,
+            482,
+            623
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {G} _ {\\text {d o w n}} = \\operatorname {A d a p t} (\\mathbf {Z}) \\in \\mathbb {R} ^ {T ^ {\\prime} \\times C _ {\\text {a d p a t e r}}}, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            158,
+            633,
+            480,
+            652
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $T' < T$  and  $C_{\\mathrm{adapter}}$  is an intermediate feature dimension aligned to the CTC head input.",
+        "bbox": [
+            89,
+            665,
+            482,
+            696
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4. CTC Head",
+        "text_level": 1,
+        "bbox": [
+            89,
+            705,
+            210,
+            720
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "An MLP with one hidden layer transforms  $\\mathbf{G}_{\\mathrm{down}} \\in \\mathbb{R}^{T' \\times C_{\\mathrm{adapter}}} \\text{ into phoneme logits:}$",
+        "bbox": [
+            89,
+            728,
+            482,
+            758
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {H} _ {\\mathrm {c t c}} = \\operatorname {M L P} \\left(\\mathbf {G} _ {\\text {d o w n}}\\right) \\in \\mathbb {R} ^ {T ^ {\\prime} \\times N _ {\\text {p h o n e m e s}}}, \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            143,
+            770,
+            480,
+            789
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $N_{\\text{phonemes}}$  is the size of our phoneme set (39 English phonemes plus a blank symbol). We then apply a logarithmic softmax along the phoneme dimension to obtain log probabilities  $\\log p_{\\text{ctc}}(ph_t \\mid \\mathbf{X})$  at each time step.",
+        "bbox": [
+            89,
+            801,
+            482,
+            863
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Since lip movements do not align cleanly with discrete phoneme boundaries, we adopt the Connectionist Temporal",
+        "bbox": [
+            89,
+            869,
+            482,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/1b981d1a719a18feb087fe2df03f3b6b59d1263151fbcdd9f66245eeb8421ddb.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Stage</td><td>Module</td><td>In → Out Channels</td><td>Kernel</td><td>Stride</td><td>Padding</td><td>Downsample Factor</td><td>Output Length</td></tr><tr><td rowspan=\"3\">1</td><td>Conv1d</td><td>feature_size → adapter_dim</td><td>5</td><td>2</td><td>2</td><td>×2</td><td>L/2</td></tr><tr><td>BatchNorm1d</td><td>adapter_dim → adapter_dim</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td>ReLU</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td rowspan=\"3\">2</td><td>Conv1d</td><td>adapter_dim → adapter_dim</td><td>3</td><td>2</td><td>1</td><td>×2</td><td>L/4</td></tr><tr><td>BatchNorm1d</td><td>adapter_dim → adapter_dim</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td>ReLU</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td rowspan=\"3\">3</td><td>Conv1d</td><td>adapter_dim → adapter_dim</td><td>3</td><td>2</td><td>1</td><td>×2</td><td>L/8</td></tr><tr><td>BatchNorm1d</td><td>adapter_dim → adapter_dim</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td>ReLU</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td rowspan=\"3\">4</td><td>Conv1d</td><td>adapter_dim → adapter_dim</td><td>3</td><td>3</td><td>1</td><td>×3</td><td>L/24</td></tr><tr><td>BatchNorm1d</td><td>adapter_dim → adapter_dim</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td>ReLU</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td>5</td><td>AvgPool1d</td><td>adapter_dim → adapter_dim</td><td>5</td><td>8</td><td>0</td><td>×8</td><td>L/192</td></tr><tr><td rowspan=\"2\">6</td><td>Linear</td><td>adapter_dim → CTC_dim</td><td>-</td><td>-</td><td>-</td><td>×1</td><td>L/192</td></tr><tr><td>ReLU</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr></table>",
+        "bbox": [
+            516,
+            88,
+            923,
+            229
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Architecture of the adapter for temporal downsampling, where the input length is  $T = 1568$  and final output length is  $T = 8$ .",
+        "bbox": [
+            511,
+            239,
+            903,
+            280
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Classification (CTC) loss [18]:",
+        "bbox": [
+            513,
+            305,
+            718,
+            320
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C T C}} = - \\ln \\left(\\sum_ {\\alpha \\in \\mathcal {A} (\\mathbf {P h})} P _ {\\mathrm {c t c}} (\\alpha \\mid \\mathbf {X})\\right), \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            325,
+            903,
+            362
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{Ph} = (ph_1, \\ldots, ph_m)$  is the ground-truth phoneme sequence,  $\\mathcal{A}(\\mathbf{Ph})$  is the set of all valid alignments, and  $P_{\\mathrm{ctc}}(\\alpha \\mid \\mathbf{X})$  is the probability of alignment  $\\alpha$ . Minimizing  $\\mathcal{L}_{\\mathrm{CTC}}$  enables the model to learn frame-to-phoneme mappings without explicit per-frame labels. At inference, we perform beam search over the CTC log probabilities to decode the final phoneme sequence.",
+        "bbox": [
+            511,
+            369,
+            906,
+            476
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.5. Phoneme-to-Sentence Reconstruction Using LLM",
+        "text_level": 1,
+        "bbox": [
+            511,
+            483,
+            903,
+            513
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The predicted phoneme sequence  $\\widehat{\\mathbf{Ph}}$  is derived by decoding the CTC logits from Eq. (4) via beam search. We then feed  $\\widehat{\\mathbf{Ph}}$  into a Large Language Model (LLM) that is finetuned to map phonemes to sentences:",
+        "bbox": [
+            511,
+            518,
+            905,
+            582
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {S} = \\operatorname {L L M} \\left(\\widehat {\\mathbf {P h}}\\right) = \\left(s _ {1}, s _ {2}, \\dots , s _ {M}\\right). \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            589,
+            903,
+            609
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "LoRA Fine-Tuning. To efficiently adapt the LLM, we employ Low-Rank Adaptation (LoRA) [22], injecting two low-rank matrices  $\\mathbf{A}$  and  $\\mathbf{B}$  into each linear layer. This greatly reduces the number of trainable parameters compared to full-model fine-tuning. Specifically, for a linear layer  $\\mathbf{W}_{\\mathrm{orig}} \\in \\mathbb{R}^{d \\times d}$ :",
+        "bbox": [
+            511,
+            623,
+            905,
+            715
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {W} ^ {\\prime} = \\mathbf {W} _ {\\text {o r i g}} + \\mathbf {A B}, \\quad \\mathbf {A} \\in \\mathbb {R} ^ {d \\times r}, \\mathbf {B} \\in \\mathbb {R} ^ {r \\times d}, \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            535,
+            722,
+            903,
+            739
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $r \\ll d$ . Only  $\\mathbf{A}$  and  $\\mathbf{B}$  are updated during training.",
+        "bbox": [
+            511,
+            747,
+            893,
+            762
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Pretraining LLM Objective. We fine-tune the LLM on a large phoneme  $\\rightarrow$  sentence corpus generated from WikiText [34]. Let  $\\mathbf{S} = (s_1, \\ldots, s_M)$  be the ground-truth sentence for a phoneme sequence  $\\mathbf{Ph}$ . We employ the cross-entropy loss:",
+        "bbox": [
+            511,
+            779,
+            905,
+            854
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} = - \\sum_ {t = 1} ^ {M} \\ln p \\left(s _ {t} \\mid \\mathbf {P h}, s _ {1: t - 1}\\right), \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            584,
+            862,
+            903,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "2849",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $p(s_{t} \\mid \\mathbf{Ph}, s_{1:t-1})$  is the likelihood of predicting the correct word  $s_{t}$ , given the phoneme sequence and previously generated words. During inference, we prompt the LLM with  $\\widehat{\\mathbf{Ph}}$  to generate  $\\widehat{\\mathbf{S}}$ .",
+        "bbox": [
+            89,
+            90,
+            480,
+            154
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.6. Training Procedure",
+        "text_level": 1,
+        "bbox": [
+            89,
+            162,
+            277,
+            179
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Stage 1: Video  $\\rightarrow$  Phoneme. We train the visual extractor, adapter, and CTC head to minimize  $\\mathcal{L}_{\\mathrm{CTC}}$  (Eq. 5) on video-phoneme pairs. By learning without strict alignment constraints, the model remains flexible to a wide range of speaking speeds and styles.",
+        "bbox": [
+            89,
+            184,
+            482,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Stage 2: Phoneme  $\\rightarrow$  Sentence. We independently fine-tune the LLM via cross-entropy loss (Eq. 8) on text-based phoneme  $\\rightarrow$  sentence datasets. This leverages easily obtained text data, allowing the LLM to learn phonetic-linguistic mappings separately from visual modeling.",
+        "bbox": [
+            89,
+            273,
+            482,
+            349
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Inference. We compose the trained modules to form:",
+        "bbox": [
+            89,
+            361,
+            442,
+            376
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {X} \\xrightarrow {f _ {\\mathrm {v i s}}} \\widehat {\\mathbf {P h}} \\xrightarrow {f _ {\\mathrm {L L M}}} \\widehat {\\mathbf {S}},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            202,
+            388,
+            370,
+            407
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\widehat{\\mathbf{Ph}}$  is decoded from CTC logits, and  $\\widehat{\\mathbf{S}}$  is the final sentence. This approach is highly modular and data-efficient, as each stage (visual and linguistic) is learned with its own objective and dataset, yet easily integrated at inference time.",
+        "bbox": [
+            89,
+            419,
+            482,
+            497
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            511,
+            223,
+            527
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This section outlines the datasets and pre-processing methods used for training the model architecture.",
+        "bbox": [
+            89,
+            537,
+            482,
+            566
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Data",
+        "text_level": 1,
+        "bbox": [
+            89,
+            575,
+            165,
+            590
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The visual feature extractor, the adapter with downsampling, and CTC head are trained on the LRS2 [47] and LRS3 [3] datasets. The LLM is fine-tuned on the WikiText [34] dataset.",
+        "bbox": [
+            89,
+            598,
+            482,
+            657
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "LRS2 [47]: The LRS2 dataset consists of 144,482 video clips of spoken sentences from BBC television, consisting of approximately 224.5 hours of footage and each sentences is up to 100 characters in length. The videos are divided into a pre-training set with 96,318 utterances (195 hours), a training set with 45,839 utterances (28 hours), a validation set with 1,082 utterances (0.6 hours) and a test set with 1,243 utterances (0.5 hours).",
+        "bbox": [
+            89,
+            666,
+            482,
+            787
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "LRS3 [3]: The LRS3 dataset describes the largest public audio-visual English dataset collected consists of clips from over 5,000 TED and TEDx talks totaling 438.9 hours. It contains 438.9 hours with 151,819 utterances. Specifically, there are 118,516 utterances in the 'pre-train' set (408 hours), 31,982 utterances in the 'train-val' set (30 hours) and 1,321 utterances in the 'test' set (0.9 hours).",
+        "bbox": [
+            89,
+            794,
+            482,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In our setting, unlike other approaches, we do not utilise the pre-training set and train our model on only the 28-hour and 30-hour partitions.",
+        "bbox": [
+            511,
+            90,
+            903,
+            137
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Phoneme Sentence Pairs. Finally, the dataset used for training the LLM was the WikiText [34] dataset with sentences converted into lists of phonemes using the CMU dictionary [52]. These phonemes were then masked and paired with the original sentences.",
+        "bbox": [
+            511,
+            143,
+            903,
+            219
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            511,
+            228,
+            697,
+            243
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Following previous works [11, 47] we adopt Word Error Rate [23] (WER) as our evaluation metric for Lip-Reading. WER calculates the percentage of errors in the predicted text compared to the ground truth, accounting for substitutions, insertions, and deletions. Similar to previous studies [3, 11, 41, 42, 47], we report results on all recent AV-ASR and V-ASR approaches.",
+        "bbox": [
+            511,
+            250,
+            903,
+            356
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5. Results",
+        "text_level": 1,
+        "bbox": [
+            513,
+            369,
+            599,
+            383
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we present the results of our method and compare them against those of existing SOTA approaches.",
+        "bbox": [
+            511,
+            395,
+            903,
+            426
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Public vs Private Data on LRS3 [3]: In Table 1, we compare different V-ASR and AV-ASR models based on their WER on the LRS3 dataset, we split the models into two categories; Fully Supervised models with publicly available data and models trained on large-scale non-publicly available datasets.",
+        "bbox": [
+            511,
+            431,
+            905,
+            522
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The Fully Supervised Models rely solely on publicly available labelled datasets with the best WER being  $40.6\\%$ , which was achieved by VTP [41]. The non-public dataset models leverage extensive, proprietary datasets, achieving much lower WERs with the best being  $12.5\\%$  by LP [10].",
+        "bbox": [
+            511,
+            530,
+            903,
+            606
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "LRS3 [3]: In Table 1 we compare our results against other methods based on results from the LRS3 dataset. The results in Table 1 demonstrate a consistent trend: models incorporating extensive self-supervised pre-training followed by fine-tuning achieve lower WERs than those using only supervised approaches. However, our model achieves the lowest WER on the LRS3 dataset at  $18.7\\%$  without self-supervised pre-training or additional fine-tuning data.",
+        "bbox": [
+            511,
+            613,
+            905,
+            734
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "LRS2 [47]: In Table 2, we compare our method against several other state-of-the-art approaches for visual-only lip reading, focusing on Word Error Rate (WER) on the LRS2 dataset. From Table 2 we can see that increasing the amount of data available helps improve a model's WER. However, we can also see that even without increasing the size of the dataset, our model still outperforms other models that only use publicly available data, achieving the lowest WER at  $20.8\\%$ .",
+        "bbox": [
+            511,
+            741,
+            905,
+            876
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Therefore as shown in Tables 2 and 3, we can see that our",
+        "bbox": [
+            513,
+            885,
+            903,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "2850",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/4bb0f2d946f065d9dbc0973f67487c4e9b6a98182344b3591187de9a2b301982.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Unlabeled (hrs)</td><td>Labeled (hrs)</td><td>WER (%)</td></tr><tr><td colspan=\"4\">Fully supervised models with publicly available data</td></tr><tr><td>ASR distillation [4]</td><td>-</td><td>590</td><td>68.8</td></tr><tr><td>Conv-Seq2Seq [60]</td><td>-</td><td>855</td><td>60.1</td></tr><tr><td>Discriminative AVSR [57]</td><td>-</td><td>590</td><td>57.8</td></tr><tr><td>Hyb. + Conformer [29]</td><td>-</td><td>590</td><td>43.3</td></tr><tr><td>VTP [41]</td><td>-</td><td>698</td><td>40.6</td></tr><tr><td colspan=\"4\">Trained on large-scale non-publicly available datasets</td></tr><tr><td>Deep-AV-SR [2]</td><td>-</td><td>1,519</td><td>58.9</td></tr><tr><td>Large-scale AV-SR [46]</td><td>-</td><td>3,886</td><td>55.1</td></tr><tr><td>RNN-T [30]</td><td>-</td><td>31,000</td><td>33.6</td></tr><tr><td>VTP [41]</td><td>-</td><td>2,676</td><td>30.7</td></tr><tr><td>ViT-3D [44]</td><td>-</td><td>90,000</td><td>17.0</td></tr><tr><td>LP [10]</td><td>-</td><td>1,000,000</td><td>12.5</td></tr><tr><td colspan=\"4\">Self-supervised pre-training + Supervised fine-tuning on LRS3</td></tr><tr><td>LiRA [28]</td><td>433</td><td>30</td><td>71.9</td></tr><tr><td>ASR distillation [4]</td><td>334</td><td>590</td><td>59.8</td></tr><tr><td>LiteVSR [24]</td><td>639</td><td>59</td><td>45.7</td></tr><tr><td>ES3 [61]</td><td>433</td><td>30</td><td>43.5</td></tr><tr><td>AV-HuBERT Large [45]</td><td>1,759</td><td>30</td><td>32.5</td></tr><tr><td>Lip2Vec [13]</td><td>1,759</td><td>30</td><td>31.2</td></tr><tr><td>Sub-Word[41]</td><td>2,676</td><td>2,676</td><td>30.7</td></tr><tr><td>SynthVSR [25]</td><td>3,652</td><td>438</td><td>27.9</td></tr><tr><td>Whisper [42]</td><td>1,759</td><td>30</td><td>25.5</td></tr><tr><td>Auto-AVSR [27]</td><td>1,759</td><td>433</td><td>25.0</td></tr><tr><td>LLaMA-AVSR [8]</td><td>-</td><td>1756</td><td>24.0</td></tr><tr><td>RAVeN [19]</td><td>1759</td><td>433</td><td>23.1</td></tr><tr><td>USR [20]</td><td>1,326</td><td>433</td><td>21.5</td></tr><tr><td colspan=\"4\">Supervised finetuning only on LRS3</td></tr><tr><td>Ours</td><td>-</td><td>30</td><td>18.7</td></tr></table>",
+        "bbox": [
+            94,
+            88,
+            496,
+            460
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/2309680e9fc8df15b81d4649f6ee4748410ca448eec1e3eceae9d15922ee1af3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            635,
+            169,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/7754bf5651d54cce94bc54485751dcdadbf2b73bbffaf38ff8144f09528b3941.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            171,
+            636,
+            194,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5241836870d9d791c95f8c9159eb00903de7d664b498f268b9fcd5a3aec71514.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            196,
+            635,
+            225,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/276819e67a971584a49e4d208e6f4455af7997117101357a4747b9be5c95485c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            225,
+            635,
+            253,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/513f852c48a8ad848e828353e103328bee41306c6fde1fbc3a53f93a75960893.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            254,
+            635,
+            281,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/015424b644c8663954efc1f779652ecdfffb7b59c66d495a4885f69ecca87fda.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            282,
+            635,
+            308,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/0aa3d8102ac7868c0e1bcdc5bebd27111bf0d66867bd2c741b52605324fd8002.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            308,
+            635,
+            334,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/4cf7ce9a562052ae9dd55dc21fbc0e6a62e0fbb85558cbb0d60521588ee3f667.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            336,
+            635,
+            362,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/0a1a3fecf95803ab3592d83a37edcb4301f5f641561753b2b269d90f613aa441.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            364,
+            635,
+            390,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/a4c6362e1c7e6d602efec1dab005d55e5828f46a9a4f415b27088b2f964df35e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            392,
+            635,
+            419,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5af5aa88bf55dd63f55f2c27f5fb8caddf21de23c2f3aca92a2e4b6dfa0146fa.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            419,
+            635,
+            446,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/861cf9126b1057f0a3876f194369d5f406de253a33de9d56ccb82e45cfed29e5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            447,
+            635,
+            473,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/9fa901c1b1ede2859a34c2e0b0be5fd01e9a440a82cb5245f16a3afbdcb64885.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            473,
+            635,
+            500,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/4409cabae598a4e370a820bb58e44f6d466d251279aa3f089505cb0d8a8076d4.jpg",
+        "table_caption": [
+            "Table 2. Comparison of Word Error Rate (WER) across different models for visual-only speech recognition on the LRS3 dataset. The table is divided into three sections: fully supervised models trained on publicly available data, models trained on large-scale non-public datasets, and models that leverage self-supervised pre-training with supervised fine-tuning on LRS3 [3]. Each entry details the amount of labelled and unlabelled data used and the resulting WER. Our approach, tested with various language models, achieves the lowest WER of  $22.1\\%$  with the Llama 3.2-3B model, demonstrating the effectiveness of our phoneme-based, LLM-assisted approach."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Predicted \nPhoneme</td><td>W, AA, Z</td><td>DH, AH</td><td>K, AY, N, D</td><td>AH, V</td></tr><tr><td>Predicted \nWord</td><td>Was</td><td>the</td><td>kind</td><td>of</td></tr><tr><td>GT</td><td>Was</td><td>the</td><td>kind</td><td>of</td></tr><tr><td>Predicted \nPhoneme</td><td>AH, B, S, EH, SH, \nAH, N</td><td>DH, AE, T</td><td>HH, AE, D</td><td>HH, ER</td></tr><tr><td>Predicted \nWord</td><td>obsession</td><td>that</td><td>had</td><td>her</td></tr><tr><td>GT</td><td>obsession</td><td>that</td><td>had</td><td>her</td></tr></table>",
+        "bbox": [
+            94,
+            676,
+            500,
+            773
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/c720ad6324dd4c6cc3507a3073064c3edc26b7b5fb6b9cf8752d83efaa366110.jpg",
+        "table_caption": [
+            "Figure 3. Example of the model's phonetic and sentence outputs from a sample in the LRS3 dataset [3]. The table illustrates the model's ability to predict a sequence of phonemes from visual input, which are then reconstructed into a coherent sentence by the LLM. In this example all of the phonemes are predicted correctly and the words are recreated correctly."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Unlabeled (hrs)</td><td>Labeled (hrs)</td><td>WER (%)</td></tr><tr><td colspan=\"4\">Fully supervised models with publicly available data</td></tr><tr><td>Auto-AV-SR [27]</td><td>-</td><td>818</td><td>27.9</td></tr><tr><td>Auto-AV-SR [27]</td><td>-</td><td>3448</td><td>14.6</td></tr><tr><td colspan=\"4\">Self-supervised pre-training + Supervised finetuning on LRS2 for AVSR</td></tr><tr><td>LiRA [28]</td><td>1,759</td><td>433</td><td>38.8</td></tr><tr><td>ES3 [61]</td><td>1,759</td><td>223</td><td>26.7</td></tr><tr><td>Sub-Word [41]</td><td>2,676</td><td>2,676</td><td>22.6</td></tr><tr><td>RAVeN [19]</td><td>1,759</td><td>223</td><td>17.9</td></tr><tr><td>USR [20]</td><td>1,759</td><td>223</td><td>15.4</td></tr><tr><td colspan=\"4\">Supervised finetuning only on LRS2</td></tr><tr><td>Ours</td><td>-</td><td>28</td><td>20.8</td></tr></table>",
+        "bbox": [
+            517,
+            88,
+            919,
+            260
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Comparison of Word Error Rate (WER) for various self-supervised pre-training and supervised fine-tuning methods evaluated on the LRS2 [47] dataset. Our method achieves competitive performance with other methods, without the requirement of extensive pre-training, additional labelled data, or additional audio modality as in the best performing approaches. We are the only method to train on only the LRS2 train dataset.",
+        "bbox": [
+            511,
+            270,
+            906,
+            368
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "method provides a lower WER with the same amount, or less data, without the requirement of self-supervised pretraining as current approaches [4, 13, 24, 25, 27, 28, 42, 45]. This large increase in performance and generalisation can be explained by the powerful combination of the multi-task objectives and the ability of the LLM to reconstruct sentences accurately even when phonemes are predicted incorrectly. In Fig. 4, we find that the phonetic output of the visual encoder is sometimes incorrect or misses phonemes, but the LLM is able to still reconstruct the correct word (as in the example  $z \\rightarrow s$ ). We also show how the model still struggles with tricky homophones such as your and you're. In Fig. 3, we show an additional example of the model's outputs at both phonetic and word levels, demonstrating how phonemes are correctly predicted and combined by the LLM for sentence reconstruction. For further examples please see the appendix.",
+        "bbox": [
+            511,
+            393,
+            906,
+            651
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. Ablations",
+        "text_level": 1,
+        "bbox": [
+            511,
+            659,
+            624,
+            672
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we aim to understand the performance contribution of each component in the network and their performance on visual  $\\rightarrow$  phoneme prediction and phoneme  $\\rightarrow$  word reconstruction, alongside ablations related to architectural decisions, and the limitations of the approach.",
+        "bbox": [
+            511,
+            681,
+            905,
+            758
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Visual  $\\rightarrow$  Phoneme: The confusion matrix in Fig. 5 shows phoneme prediction performance of the model on the LRS3 [3] dataset. The matrix indicates that some phonemes are misclassified more frequently than others, as seen by the off-diagonal elements, with the worst missed classifications being highlighted with a red bounding box. This bias could be due to the similarity in the visual representation of these sounds, making it challenging for the model to distinguish between them. For instance, /S/ and /Z/ may often be con-",
+        "bbox": [
+            511,
+            763,
+            906,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "2851",
+        "bbox": [
+            482,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/4e8fc1ec0b0814031a2d09ff8bf89e914bb5512b94eccdc80d0511cedb025fbc.jpg",
+        "image_caption": [
+            "Figure 4. Comparison of the model's phonetic and sentence outputs with ground truth from a sample in the LRS3 [3] dataset. Red shows an incorrect prediction, M shows a missing prediction and green shows a correction from phonemes to words. In this example, even though the model incorrectly predicts certain phonemes, the LLM can correctly recreate the word but struggles to recreate homophones."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            87,
+            500,
+            119
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/8e711fe5e567d49e2d32bb3875ffaa3e363b360ab9840fb1300f0144e21978e0.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Predicted \nPhoneme</td><td>Y, UH, R</td><td>JH, AH, Z, T</td><td>W, AH, N</td><td>AH, V</td></tr><tr><td>Predicted \nWord</td><td>Your</td><td>just</td><td>one</td><td>of</td></tr><tr><td>GT</td><td>You&#x27;re</td><td>just</td><td>one</td><td>of</td></tr><tr><td>Predicted \nPhoneme</td><td>DH, AH</td><td>P, IY, P, &lt;M&gt;, L</td><td>HH, UW</td><td>HH, AE, S, N, T</td></tr><tr><td>Predicted \nWord</td><td>the</td><td>people</td><td>who</td><td>hasn&#x27;t</td></tr><tr><td>GT</td><td>the</td><td>people</td><td>who</td><td>hasn&#x27;t</td></tr></table>",
+        "bbox": [
+            93,
+            128,
+            500,
+            224
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "fused due to similar lip movements, leading to miss classifications between these phonemes.",
+        "bbox": [
+            89,
+            359,
+            482,
+            390
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f7d26b30e521eb6a97cd06f3b94f654e12b2cf2e2cd055439660d384e40abed9.jpg",
+        "image_caption": [
+            "Figure 5. Confusion matrix showing the performance on isolated phonemes of the LRS3 [3] dataset. We observe a very high match rate between the predicted phonemes and the ground truth. In red, we show the most difficult phonemes for our model to identify."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            89,
+            400,
+            501,
+            691
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Phoneme  $\\rightarrow$  Word: In Table 4 we take the top 5 words in alphabetical order that are misclassified when predicting words directly from phonemes after the LLM pretraining stage. We can observe that these are words that may occur infrequently in the training set and are missclassified due to some phonetic similarities between the words.",
+        "bbox": [
+            89,
+            771,
+            482,
+            863
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Different LLMs: In this ablation study, we evaluate three different Language models on our model architecture to",
+        "bbox": [
+            89,
+            869,
+            483,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/9d3d4061eb0bfc8f590cb45171de428e906f503bc03b483afafcb52a7cd45f4b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Phonemes</td><td>True Word</td><td>Predicted Word</td></tr><tr><td>/EH/ /R/ /AH/ /N/ /S/ /AH/ /N/</td><td>Aaronson</td><td>Erasmus</td></tr><tr><td>/AA/ /B/ /ER/ /G/</td><td>Aaberg</td><td>Braggartly</td></tr><tr><td>/AH/ /S/ /EH/ /R /AH/</td><td>Acerra</td><td>Esraa</td></tr><tr><td>/AE/ /S/ /AH/ /T/ /EY /T/</td><td>Acetate</td><td>Asquette</td></tr><tr><td>/AH/ /K/ /AH/ /S/ /T/ /AH/ /M/ /D/</td><td>Accustomed</td><td>Accosted</td></tr></table>",
+        "bbox": [
+            517,
+            87,
+            919,
+            167
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "understand the impact on word error rate (WER) performance. The first model, GPT 2 [43], serves as our baseline, consisting of the default GPT 2 [43] without fine-tuning on lip-reading datasets. This model achieves a WER of  $23.9\\%$ , indicating the core components' foundational capabilities. The second model, Llama 3.2-1B [51] introduces fine-tuning on phoneme word pairs, enhancing the model's contextual understanding of phoneme sequences. With this addition, Llama 3.2-1B [51] has a substantial improvement, reducing the WER to  $22.8\\%$ . Finally, Llama 3.2-3B [51] also incorporates fine-tuning for phoneme-to-word reconstruction. However, Llama 3.2-3B [51] is a larger model with more parameters than the Llama 3.2-1B [51] model, further lowering the WER to  $18.7\\%$ . This progression highlights the significant contributions of fine-tuning for purpose-specific LLMs.",
+        "bbox": [
+            511,
+            290,
+            906,
+            532
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/cb11ee012b8e5c7758268e42e9f87c27a56e49644ba4c76ffcde0a43b1da4199.jpg",
+        "table_caption": [
+            "Table 4. Examples of misclassification's between predicted words and ground truth based on phoneme sequences. Each row presents a sequence of phonemes, the intended (true) word, and the model's predicted word. Differences in predictions illustrate common misclassifications, often due to phonetic similarities between visually indistinct sounds."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Parameters (B)</td><td>WER (%)</td></tr><tr><td>GPT-2 Small [43]</td><td>0.12</td><td>33.9</td></tr><tr><td>Llama 3.2-1B [51]</td><td>1.0</td><td>22.8</td></tr><tr><td>Llama 3.2-3B [51]</td><td>3.0</td><td>18.7</td></tr></table>",
+        "bbox": [
+            527,
+            546,
+            893,
+            625
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 5. Comparison of Word Error Rate (WER) for different large language models (LLMs) tested with our model. The table includes each model's parameter count (in billions) and resulting WER on the LRS3 [3] dataset. GPT-2 [43] Small with 0.12B parameters has the highest WER at  $23.9\\%$ . In contrast, Llama 3.2-3B [51] achieves the best WER of  $18.7\\%$ , highlighting that increased model capacity improves phoneme-to-word reconstruction accuracy.",
+        "bbox": [
+            511,
+            633,
+            906,
+            746
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Freezing the CTC Head: Inspired by works that directly map visual features to pretrained ASR networks [42] we investigate the effect of initialising the CTC head in our model with weights from a pre-trained Wav2Vec2 ASR model [6]. The CTC head plays a crucial role in mapping visual features to the phoneme sequences, and we hypothesized that initialising and freezing its parameters could reduce computational overhead while preserving feature alignment from audio pre-training. To test this hypothesis a baseline model",
+        "bbox": [
+            511,
+            763,
+            908,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "2852",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "with a non-frozen CTC head and a comparison model with a frozen CTC head were trained and a comparison of the results was made as shown in Table 6. Freezing the CTC head led to a slight increase in WER of  $0.4\\%$ , indicating a decrease in performance. This result suggests that while the frozen CTC head still provided a foundation for phoneme alignment, we did not require extensive audio pre-training to obtain superior performance.",
+        "bbox": [
+            89,
+            90,
+            480,
+            210
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Varying Sample Sizes: To understand the impact of additional data on performance, we vary the quantity of data used to train both the Video  $\\rightarrow$  Phoneme network and the Phoneme  $\\rightarrow$  Sentence LLM and investigate the effect this has on the WER when evaluated on the LRS3 [3] dataset. By varying the percentages of data used in training, we can see the effect on the model in Figure 6. Decreasing the amount of data used to train both models has the expected effect on the WER by increasing it. However, increasing the amount of training data for the Feature Extractor has a negligible effect after 30 hours of data. However, by increasing the amount of training data for the LLM we can also increase the effectiveness of the model and decrease the total WER down further to  $17.5\\%$ . The additional data is randomly sampled from both AvSpeech [17] for the visual encoder, and a BookCorpus [63] replica to match the quantity of data in WikiText.",
+        "bbox": [
+            91,
+            220,
+            483,
+            477
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/9039c2c35aa9c82d656a47cc63694fed24bdefa3cff3157abb122551594fff59.jpg",
+        "image_caption": [
+            "Figure 6. Graph showing the effect of varying the sample size of the training datasets for both the Video  $\\rightarrow$  Phoneme network (in blue) and for the Phoneme  $\\rightarrow$  Sentence LLM (in red). The graph shows that decreasing the amount of data negatively impacts the model for both networks, increasing the amount of training data for the Video  $\\rightarrow$  Phoneme network has negligible effect and increasing the amount of training data for the Phoneme  $\\rightarrow$  Sentence LLM has a noticeable positive effect."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            507,
+            464,
+            720
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "End-To-End: For our final study, we investigate the effect of removing the video  $\\rightarrow$  phoneme stage and train the",
+        "bbox": [
+            89,
+            869,
+            482,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "model to directly predict sentences from the video features, reproducing the method in [41] but with the same ViT encoder [44], LLama LLM [51], and limited data set. As shown in Table 6 we achieve aWER of  $25.6\\%$  on LRS3, demonstrating that the more parameterised transformer encoder improves performance over the CNN encoder used in [41]. However, our two stage approach still performs significantly better.",
+        "bbox": [
+            511,
+            90,
+            903,
+            212
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/f233a11b44b49368eb2ad65b5c63bc99b6dc35c265d76a66a54f5192fb4b4bff.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>WER (%)</td></tr><tr><td>Ours (Two Stage)</td><td>18.7</td></tr><tr><td>ViT + LLM (End to End)</td><td>25.6</td></tr><tr><td>Sub-Word (End to End) [41]</td><td>30.7</td></tr></table>",
+        "bbox": [
+            560,
+            227,
+            859,
+            305
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6. Comparison of Word Error Rate (WER) between using the intermediary Phonemes and skipping the CTC head. Removing the intermediary representation increases the WER from 18.7 to 25.6, showing the importance of the intermediary step and the effectiveness of the phoneme centric fine-tuning.",
+        "bbox": [
+            511,
+            314,
+            906,
+            385
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion and Limitations",
+        "text_level": 1,
+        "bbox": [
+            511,
+            410,
+            769,
+            426
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We have introduced a two-stage, phoneme-centric framework for visual-only lip reading that first predicts phonemes from lip movements and then reconstructs coherent sentences via a fine-tuned LLM. This design sharply reduces word error rates on benchmark datasets (LRS2 [47] and LRS3 [3]), providing improvements of  $3.5\\%$  and  $6.8\\%$  WER, respectively, over existing state-of-the-art approaches that focus on end-to-end connectionist approaches.",
+        "bbox": [
+            511,
+            435,
+            906,
+            571
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Despite these gains, the approach remains limited by the visual ambiguity of certain phonemes. Lip movements alone cannot always differentiate minimal pairs such as /s/ vs. /z/, and words that share near-identical phoneme sequences (e.g., \"Hello\" vs. \"Hallow\") can still result in misclassification. We plan to address these issues by refining the phoneme-to-word generation process, potentially through deeper linguistic context modeling and more specialized phoneme embeddings. Future work may also explore speaker-adaptive fine-tuning or multi-frame alignment strategies to better handle subtle visual distinctions.",
+        "bbox": [
+            511,
+            582,
+            908,
+            750
+        ],
+        "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 by the SNSF project 'SMILE II' (CRSII5 193686), the Innosuisse IICT Flagship (PFFS-21-47), EPSRC grant APP24554 (SignGPT-EP/Z535370/1) and through funding from Google.org via the AI for Global Goals scheme. This work reflects only the author's views and the funders are not responsible for any use that may be made of the information it contains.",
+        "bbox": [
+            511,
+            787,
+            906,
+            887
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "2853",
+        "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] Emre Afacan and Gültekin Demirci. A survey on automatic lip-reading: Challenges and future directions. Multimedia Tools and Applications, 2019. 1, 3",
+            "[2] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 2018. 3, 6",
+            "[3] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 1, 3, 5, 6, 7, 8",
+            "[4] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2143-2147. IEEE, 2020. 3, 6",
+            "[5] Yannis M Assael, Brendan Shillingford, Shimon Whiteson, and Nando De Freitas. Lipnet: End-to-end sentence-level lipreading. arXiv preprint arXiv:1611.01599, 2016. 3",
+            "[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 2020. 7",
+            "[7] Helen L Bear, Richard W Harvey, Barry-John Theobald, and Yuxuan Lan. Which phoneme-to-viseme maps best improve visual-only computer lip-reading? In International symposium on visual computing. Springer, 2014. 1",
+            "[8] Umberto Cappellazzo, Minsu Kim, Honglie Chen, Pingchuan Ma, Stavros Petridis, Daniele Falavigna, Alessio Brutti, and Maja Pantic. Large language models are strong audio-visual speech recognition learners, 2025. 6",
+            "[9] Luca Cappelletta and Naomi Harte. Viseme definitions comparison for visual-only speech recognition. In 2011 19th European Signal Processing Conference. IEEE, 2011. 3",
+            "[10] Oscar Chang, Hank Liao, Dmitriy Serdyuk, Ankit Shahy, and Olivier Siohan. Conformer is all you need for visual speech recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024. 5, 6",
+            "[11] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian Conference on Computer Vision, 2016. 3, 5",
+            "[12] Martin Cooke, Jon Barker, Stuart Cunningham, and Xu Shao. Grid av speech corpus sample. The Journal of the Accoustical Society of America, 2013. 3",
+            "[13] Yasser Abdelaziz Dahou Djilali, Sanath Narayan, Haithem Boussaid, Ebtessam Almazrouei, and Merouane Debbah. Lip2vec: Efficient and robust visual speech recognition via latent-to-latent visual to audio representation mapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3, 6",
+            "[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3",
+            "[15] Randa El-Bialy, Daqing Chen, Souheil Fenghour, Walid Hussein, Perry Xiao, Omar H. Karam, and Bo Li. Developing phoneme-based lip-reading sentences system for silent"
+        ],
+        "bbox": [
+            93,
+            114,
+            482,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "speech recognition. CAAI Transactions on Intelligence Technology, 8(1):129-138, 2023. 3",
+            "[16] Randa El-Bialy, Daqing Chen, Souheil Fenghour, Walid Hussein, Perry Xiao, Omar H Karam, and Bo Li. Developing phoneme-based lip-reading sentences system for silent speech recognition. CAAI Transactions on Intelligence Technology, 2023. 3",
+            "[17] A. Ephrat, I. Mosseri, O. Lang, T. Dekel, K Wilson, A. Hassidim, W. T. Freeman, and M. Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 8",
+            "[18] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, 2006. 4",
+            "[19] Alexandros Haliassos, Pingchuan Ma, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Jointly learning visual and auditory speech representations from raw data. arXiv preprint arXiv:2212.06246, 2022. 6",
+            "[20] Alexandros Haliassos, Rodrigo Mira, Honglie Chen, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Unified speech recognition: A single model for auditory, visual, and audio-visual inputs. arXiv preprint arXiv:2411.02256, 2024. 6",
+            "[21] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 2021. 3",
+            "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 2022. 4",
+            "[23] F. Jelinek, L. Bahl, and R. Mercer. Design of a linguistic statistical decoder for the recognition of continuous speech. IEEE Transactions on Information Theory, 21(3), 1975. 5",
+            "[24] Hendrik Laux, Emil Mededovic, Ahmed Hallawa, Lukas Martin, Arne Peine, and Anke Schmeink. Litevsr: Efficient visual speech recognition by learning from speech representations of unlabeled data. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024. 3, 6",
+            "[25] Xubo Liu, Egor Lakomkin, Konstantinos Vougioukas, Pingchuan Ma, Honglie Chen, Ruiming Xie, Morrie Doulaty, Niko Moritz, Jachym Kolar, Stavros Petridis, et al. Synthvsr: Scaling up visual speech recognition with synthetic supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 6",
+            "[26] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 4",
+            "[27] Pingchuan Ma and Alexandros et al. Haliassos. Auto-avsr: Audio-visual speech recognition with automatic labels. In"
+        ],
+        "bbox": [
+            516,
+            93,
+            903,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "2854",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023. 3, 6",
+            "[28] Pingchuan Ma and Rodrigo et al. Mira. Lira: Learning visual speech representations from audio through self-supervision. arXiv, 2021. 3, 6",
+            "[29] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with conformers. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021. 6",
+            "[30] Takaki Makino, Hank Liao, Yannis Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In IEEE automatic speech recognition and understanding workshop (ASRU). IEEE, 2019. 3, 6",
+            "[31] William D Marslen-Wilson. Functional parallelism in spoken word-recognition. Cognition, 25(1-2), 1987. 2",
+            "[32] William D Marslen-Wilson and Alan Welsh. Processing interactions and lexical access during word recognition in continuous speech. Cognitive psychology, 10(1), 1978. 2",
+            "[33] Harry McGurk and John MacDonald. Hearing lips and seeing voices. Nature, 1976. 1",
+            "[34] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. 4, 5",
+            "[35] Ara V Nefian, Luhong Liang, Xiaobo Pi, Liu Xiaoxiang, Crusoe Mao, and Kevin Murphy. A coupled hmm for audiovisual speech recognition. In 2002 IEEE International Conference on Acoustics, Speech, and Signal Processing. IEEE, 2002. 3",
+            "[36] Eng-Jon Ong and Richard Bowden. Learning temporal signatures for lip reading. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops). IEEE, 2011. 3",
+            "[37] Stavros Petridis and Maja Pantic. Deep complementary bottleneck features for visual speech recognition. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3",
+            "[38] Stavros Petridis, Themos Stafylakis, Pingchuan Ma, Georgios Tzimiropoulos, and Maja Pantic. Audio-visual speech recognition with a hybrid ctc/attention architecture. In 2018 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2018. 3",
+            "[39] Javad Peymanfard, Vahid Saeedi, Mohammad Reza Mohammadi, Hossein Zeinali, and Nasser Mozayani. Leveraging visemes for better visual speech representation and lip reading, 2023. 3",
+            "[40] Gerasimos Potamianos, Chalapathy Neti, Guillaume Gravier, Ashutosh Garg, and Andrew W Senior. Recent advances in the automatic recognition of audiovisual speech. Proceedings of the IEEE, 91(9), 2003. 3",
+            "[41] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 5, 6, 8",
+            "[42] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Speech recognition models are strong lip-readers. In Proc. Interspeech 2024, 2024. 3, 5, 6, 7"
+        ],
+        "bbox": [
+            91,
+            90,
+            483,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[43] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. arXiv, 2018. 3, 7",
+            "[44] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition for single and multi-person video. arXiv preprint arXiv:2201.10439, 2022. 3, 4, 6, 8",
+            "[45] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. *ICLR*, 2022. 3, 6",
+            "[46] Brendan Shillingford, Yannis Assael, Matthew W Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, et al. Large-scale visual speech recognition. InterSpeech, 2018. 3, 6",
+            "[47] Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Lip reading sentences in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 3, 5, 6, 8",
+            "[48] Themos Stafylakis and Georgios Tzimiropoulos. Zero-shot keyword spotting for visual speech recognition in-the-wild. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-529, 2018. 3",
+            "[49] Kwanchiva Thangthai, Helen L Bear, and Richard Harvey. Comparing phonemes and visemes with dnn-based lipreading. arXiv preprint arXiv:1805.02924, 2018. 3",
+            "[50] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35, 2022. 3",
+            "[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3, 7, 8",
+            "[52] Carnegie Mellon University. The cmu pronouncing dictionary, 1993. http://www.speech.cs.cmu.edu/cgi-bin/cmudict.2,5",
+            "[53] 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, 2017. 3",
+            "[54] Michael Wand, Jan Koutnik, and Jürgen Schmidhuber. Lipreading with long short-term memory. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3",
+            "[55] Michael Wand, Jan Koutnik, and Jürgen Schmidhuber. Lipreading with long short-term memory. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3",
+            "[56] Xinshuo Weng and Kris Kitani. Learning spatio-temporal features with two-stream deep 3d cnn for lipreading. arXiv preprint arXiv:1905.02540, 2019. 3",
+            "[57] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, 2020. 3, 6",
+            "[58] Jeong Hun Yeo, Seunghee Han, Minsu Kim, and Yong Man Ro. Where visual speech meets language: Vsp-llm frame"
+        ],
+        "bbox": [
+            516,
+            92,
+            906,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "2855",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "work for efficient and context-aware visual speech processing. arXiv preprint arXiv:2402.15151, 2024. 2, 3",
+            "[59] Jeong Hun Yeo, Chae Won Kim, Hyunjun Kim, Hyeongseop Rha, Seunghee Han, Wen-Huang Cheng, and Yong Man Ro. Personalized lip reading: Adapting to your unique lip movements with vision and language. arXiv preprint arXiv:2409.00986, 2024. 2, 3",
+            "[60] Xingxuan Zhang, Feng Cheng, and Shilin Wang. Spatiotemporal fusion based convolutional sequence learning for lip reading. In Proceedings of the IEEE/CVF International conference on Computer Vision, 2019. 3, 6",
+            "[61] Yuanhang Zhang, Shuang Yang, Shiguang Shan, and Xilin Chen. Es3: Evolving self-supervised learning of robust audio-visual speech representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 6",
+            "[62] Ziheng Zhou, Guoying Zhao, Xiaopeng Hong, and Matti Pietikainen. A review of recent advances in visual speech decoding. Image and vision computing, 32(9), 2014. 3",
+            "[63] Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, 2015. 8"
+        ],
+        "bbox": [
+            91,
+            90,
+            482,
+            443
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "2856",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2025/VALLR_ Visual ASR Language Model for Lip Reading/88644d63-e953-49ce-b2d3-f0418686524b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c314d458f2fa8c83c775c6acd30b3e01aeec072736d457fd9c94a2db792ba2b8
+size 893483

2025/VALLR_ Visual ASR Language Model for Lip Reading/full.md

@@ -0,0 +1,374 @@
+# VALLR: Visual ASR Language Model for Lip Reading
+
+Marshall Thomas Edward Fish Richard Bowden University of Surrey
+
+mt00893@surrey.ac.uk edward.fish@surrey.ac.uk r.bowden@surrey.ac.uk
+
+# Abstract
+
+Lip Reading, or Visual Automatic Speech Recognition (V-ASR), is a complex task requiring the interpretation of spoken language exclusively from visual cues, primarily lip movements and facial expressions. This task is especially challenging due to the absence of auditory information and the inherent ambiguity when visually distinguishing phonemes that have overlapping visemes, where different phonemes appear identical on the lips. Current methods typically attempt to predict words or characters directly from these visual cues, but this approach frequently encounters high error rates due to coarticulation effects and viseme ambiguity. We propose a novel two-stage, phoneme-centric framework for Visual Automatic Speech Recognition (V-ASR) that addresses these longstanding challenges. First, our model predicts a compact sequence of phonemes from visual inputs using a Video Transformer with a CTC head, thereby reducing the task complexity and achieving robust speaker invariance. This phoneme output then serves as the input to a fine-tuned Large Language Model (LLM), which reconstructs coherent words and sentences by leveraging broader linguistic context. Unlike existing methods that either predict words directly or rely on large-scale multimodal pre-training, our approach explicitly encodes intermediate linguistic structure while remaining highly data efficient. We demonstrate state-of-the-art performance on two challenging datasets, LRS2 and LRS3, where our method achieves significant reductions in Word Error Rate (WER) achieving a SOTA WER of 18.7 on LRS3 despite using  $99.4\%$  less labelled video data than the next best approach. Code is available here: https://gitlab.surrey.ac.uk/
+
+# 1. Introduction
+
+Lip Reading, or Visual Automatic Speech Recognition (V-ASR), involves interpreting spoken language from visual cues such as lip movements and facial expressions. As a natural skill, humans use it to supplement auditory information, and as a technology, it has profound potential for en
+
+![](images/9ae34c4388a4f19abe6b22865a145bffff65591cbe287bfbeb007bdb92f19060.jpg)  
+Figure 1. Comparison between different models' performances in WER for Visual Automatic Speech Recognition on the LRS3 dataset [3] when compared with the amount of labelled training data. Circle size (green) denotes scale of pre-training data, while fine-tuned models (grey) are not pre-trained. Our model (orange) outperforms all existing approaches with just 30 hours of video training data, no self-supervised pre-training, and without the requirement for additional labeled visual data during fine-tuning.
+
+hancing accessibility, particularly for the Deaf and hard-of-hearing communities, as well as for applications in noisy or privacy-sensitive environments [1]. Despite substantial advancements in ASR, the visual-only interpretation of speech remains an unresolved challenge due to inherent uncertainties in lip movements and their complex temporal dynamics. A central difficulty in V-ASR stems from the inherent ambiguity of visemes, the visual equivalents of phonemes, which often appear nearly identical for different sounds (e.g., 'p' and 'b') [1, 7]. Further complicating this task are coarticulation effects [33], where adjacent sounds blur one another's articulation, making phoneme boundaries visually unclear. Although some recent methods have focussed on robust intermediary representations to deal with these cases, such as leveraging subword-level predictions [41], or quantized latent representations [13], these approaches can still fail to capture broader contextual cues, limiting their ability to resolve visually similar phonemes that might otherwise be
+
+![](images/85b1f3f5dae7dfb135fece6cf0a7069536516d7e82e8e5024ee5d2691a090d08.jpg)  
+Figure 2. An overview of our approach. First we extract facial regions from 16 frames of input video. We apply random pixel masking at  $50\%$  probability, add positional embedding, and encode visual features via a vision transformer encoder (ViT-base [14]). We implement temporal downsampling via 1D convolution and then use a CTC linear head to predict sequences of phonemes. During training, we also fine-tune an LLM to reconstruct sentences from phonemes with a text-only dataset [52]. During inference, the phonemes from the CTC head are processed via the LLM to reconstruct the predicted text. This can be performed end-to-end or in two stages, depending on available resources.
+
+distinguishable through sentence-level semantics.
+
+One approach to this challenge is to leverage large language models (LLM's) and attempt to map raw lip movements directly to sentences, thus bypassing the issue of viseme ambiguity [58, 59]. However, bridging the gap between high-dimensional, unstructured visual inputs and detailed textual representations is resource intensive and typically requires huge datasets, or specialized architectures. Moreover, these purely end-to-end pipelines often lack an interpretable intermediary step, making it unclear how the model is handling viseme ambiguity at a finer linguistic granularity. This can lead to hallucination effects, which become more critical in lip reading technologies designed for accessibility, or security applications.
+
+Our two-stage, phoneme-centric method addresses both issues. In the first stage, we map short windows of video frames to a discrete, interpretable phoneme sequence. This intermediate representation is significantly easier to learn and more robust against speaker or environmental variations, as phonemes abstract away speaker-specific attributes. Subsequently, in the second stage, we fine-tune an LLM for the task of reconstructing sentences from phonemes. By segmenting the problem, we combine the interpretability and efficiency of phoneme-based prediction with the sophisticated contextual reasoning of modern LLMs. This phoneme  $\rightarrow$  sentence pipeline also aligns with psycholinguistic models of speech perception, where
+
+phoneme-level processing naturally precedes lexical access [31, 32].
+
+Several advantages emerge from this design:
+
+(i) Compact Target Space: Predicting only 38 phoneme classes avoids learning entire word-level vocabularies, making the model more data-efficient and simpler to train.  
+(ii) LLM-Enhanced Accuracy: A large language model, pre-trained for phoneme  $\rightarrow$  sentence reconstruction, removes the requirement of direct word-level predictions from RGB data, reducing the need for complex multimodal training.  
+(iii) Data Availability: Exploiting easily available phoneme  $\rightarrow$  sentence text pairs eliminates reliance on extensive lip-reading video data for pre-training.  
+(iv) Interpretability & Error Analysis: The two-stage design yields an explicit phoneme-level output, constraining recognition errors to local phoneme segments, which the LLM subsequently repairs at the word level reducing errors at a sentence level.
+
+# 2. Related Work
+
+Here, we review key categories of approaches and highlight their advantages and limitations in the context of our proposed phoneme-centric, language-aware approach.
+
+Lip Reading: Early methods predominantly relied on statistical models such as Hidden Markov Models (HMMs)
+
+to aggregate temporal sequences of lip movements alongside hand-crafted feature extractors [1, 9, 35, 36, 40, 62]. With the advent of deep learning, coupled with the availability of large-scale datasets such as Grid [12], LRW [11], LRS2 [47], and LRS3 [3], significant progress has been made in recent years. Initially, Convolutional Neural Networks (CNNs) were applied to the task of word recognition [11], with additional temporal processing via RNN's [30, 54, 55]. With greater compute came the possibility to interpret full sentences with approaches applying Automatic Speech Recognition (ASR) methodologies to the Visual Automatic Speech Recognition (V-ASR) task. These included sequence-to-sequence models [2], CTC approaches [5, 46], and hybrid models [37, 38]. These models improved recognition accuracy by learning hierarchical representations of visual features and modelling temporal dependencies. The application of transformer models [14, 21, 44, 50, 53] in V-ASR brought further performance gains in combination with extensive self-supervised pretraining. Methods such as AV-HuBERT [45] and LiteVSR [24] leveraged self-supervised learning to learn robust representations from large amounts of unlabelled video data.
+
+AV-ASR vs V-ASR: Lip reading approaches can be separated into Visual Automatic Speech Recognition (V-ASR) and Audio-Visual Automatic Speech Recognition (AV-ASR). In the AV-ASR task, labelled audio data is available at training time which can be used to guide the visual encoder in various ways. These can include generating synthetic additional pre-training data [25, 27], or simply fusing audio-visual input features [44-46, 57]. Other methods have used knowledge distillation to inductively transfer knowledge from pre-trained ASR models to visual encoders [4, 28]. However, these methods depend on audio data or pre-trained ASR models, which may not be available or reliable in scenarios where lip reading is most needed, such as noisy environments. V-ASR, on the other hand, is a more challenging task that relies only on visual data in the training stage [1, 56, 60, 62]. Recent methods have also shown the effectiveness of adapting visual inputs to pre-trained ASR models [13, 24, 42] without the need for audio data during training. This method is particularly powerful since the ASR model already has the ability to reconstruct latent representations to words from the pre-training task, however it relies on large amounts of additional data to learn robust visual representations and map them to the audio latent space [42].
+
+Lip Reading with LLMs: Recent approaches have explored visual attention mechanisms for sub-word units [15, 39, 41, 48] to capture fine-grained linguistic features in lip reading. However, reconstructing these units into coherent sentences remains challenging for networks lacking robust contextual and semantic reasoning capabilities. Early
+
+phoneme-based methods [16, 49] similarly struggled with effective decoding to word-level representations and these methods were outpaced in performance by end-to-end approaches [5] which could leverage extensive data. As such they have not been fully explored in the context of modern architectures and techniques.
+
+For example, recent approaches which have integrated Large Language Models (LLMs) [43, 51] into lip-reading pipelines [58, 59] have started from mapping visual features directly to LLM text embeddings. However, by omitting explicit intermediate representations, these approaches inherit phonetic ambiguities that propagate through the network. This manifests as increased word error rates and reduced robustness, thus requiring substantially more training data to achieve generalization.
+
+Our work re-examines phonemes as an interpretable discrete representation bridging visual and linguistic domains. Rather than forcing LLMs to perform visual  $\rightarrow$  sentence translation requiring expensive visual-text alignment – we constrain the LLM's role to phoneme  $\rightarrow$  sentence reconstruction. This decomposition offers three key advantages: (1) Phoneme-to-text translation is a well-constrained task requiring only lightweight LLM fine-tuning using abundant textual corpora; (2) Speaker-independent phoneme representations eliminate the need for visual adaptation in the language model; (3) The modular architecture prevents error propagation between visual and linguistic domains while maintaining interpretability.
+
+# 3. Method
+
+In this section, we present our two-stage phoneme-centric approach to visual-only lip reading. Figure 2 provides an overview of the framework. Our goal is to learn a function
+
+$$
+\begin{array}{l} f: \quad \mathbf {X} = \left\{x _ {1}, x _ {2}, \dots , x _ {T} \right\} \\ \longmapsto \mathbf {P h} = \left\{p h _ {1}, p h _ {2}, \dots , p h _ {m} \right\} \tag {1} \\ \longmapsto \mathbf {S} = \left\{s _ {1}, s _ {2}, \dots , s _ {M} \right\}. \\ \end{array}
+$$
+
+where  $\mathbf{X} \in \mathbb{R}^{T \times H \times W \times 3}$  is a video sequence of  $T$  frames (with height  $H$  and width  $W$ ),  $\mathbf{Ph}$  denotes the phoneme sequence, and  $\mathbf{S}$  is the reconstructed sentence. Specifically, we decompose the task into:
+
+1. Video  $\rightarrow$  Phoneme: Map the sequence of video frames to phonemes using a Vision Transformer and CTC head.  
+2. Phoneme  $\rightarrow$  Sentence: Convert phonemes to a coherent word sequence with a fine-tuned Large Language Model (LLM).
+
+The model comprises four primary components:
+
+- Visual Feature Extractor: Captures relevant features from each frame via a Vision Transformer.
+
+- Adapter Network with Temporal Downsampling: Reduces sequence length to accommodate CTC alignment.  
+- CTC Head: Predicts phoneme sequences without requiring explicit temporal labels.  
+- LLM for Phoneme-to-Sentence Reconstruction: Translates predicted phonemes into final word sequences.
+
+# 3.1. Pre-Processing Pipeline
+
+Each frame  $x_{i}$  is first passed through a face-detection model [26] to identify and crop the speaker's face region, centered on the speaker's lips. The detected region is then cropped, resized to  $224 \times 224$ , and normalized for batch processing.
+
+# 3.2. Video Transformer
+
+We employ a Vision Transformer (ViT) [44] to encode the spatio-temporal information in the face region. Let ViT(\cdot) denote this encoding function. The input  $\mathbf{X} \in \mathbb{R}^{T \times H \times W \times 3}$  is chunked into patches, embedded, and equipped with positional encodings before passing through multiple transformer blocks. We obtain:
+
+$$
+\mathbf {Z} = \operatorname {V i T} (\mathbf {X}) \in \mathbb {R} ^ {T \times D}, \tag {2}
+$$
+
+where  $D$  is the transformer output dimension per frame. During training, we randomly mask portions of the input patches to promote robust feature learning and improve generalization.
+
+# 3.3. Adapter Network with Temporal Downsampling
+
+The sequence length  $T$  of the ViT output can be prohibitively large for the subsequent CTC module. To reduce the temporal dimension, we apply a sequence of 1D convolutions and pooling layers, as shown in Table ?? Formally,
+
+$$
+\mathbf {G} _ {\text {d o w n}} = \operatorname {A d a p t} (\mathbf {Z}) \in \mathbb {R} ^ {T ^ {\prime} \times C _ {\text {a d p a t e r}}}, \tag {3}
+$$
+
+where  $T' < T$  and  $C_{\mathrm{adapter}}$  is an intermediate feature dimension aligned to the CTC head input.
+
+# 3.4. CTC Head
+
+An MLP with one hidden layer transforms  $\mathbf{G}_{\mathrm{down}} \in \mathbb{R}^{T' \times C_{\mathrm{adapter}}} \text{ into phoneme logits:}$
+
+$$
+\mathbf {H} _ {\mathrm {c t c}} = \operatorname {M L P} \left(\mathbf {G} _ {\text {d o w n}}\right) \in \mathbb {R} ^ {T ^ {\prime} \times N _ {\text {p h o n e m e s}}}, \tag {4}
+$$
+
+where  $N_{\text{phonemes}}$  is the size of our phoneme set (39 English phonemes plus a blank symbol). We then apply a logarithmic softmax along the phoneme dimension to obtain log probabilities  $\log p_{\text{ctc}}(ph_t \mid \mathbf{X})$  at each time step.
+
+Since lip movements do not align cleanly with discrete phoneme boundaries, we adopt the Connectionist Temporal
+
+<table><tr><td>Stage</td><td>Module</td><td>In → Out Channels</td><td>Kernel</td><td>Stride</td><td>Padding</td><td>Downsample Factor</td><td>Output Length</td></tr><tr><td rowspan="3">1</td><td>Conv1d</td><td>feature_size → adapter_dim</td><td>5</td><td>2</td><td>2</td><td>×2</td><td>L/2</td></tr><tr><td>BatchNorm1d</td><td>adapter_dim → adapter_dim</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td>ReLU</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td rowspan="3">2</td><td>Conv1d</td><td>adapter_dim → adapter_dim</td><td>3</td><td>2</td><td>1</td><td>×2</td><td>L/4</td></tr><tr><td>BatchNorm1d</td><td>adapter_dim → adapter_dim</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td>ReLU</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td rowspan="3">3</td><td>Conv1d</td><td>adapter_dim → adapter_dim</td><td>3</td><td>2</td><td>1</td><td>×2</td><td>L/8</td></tr><tr><td>BatchNorm1d</td><td>adapter_dim → adapter_dim</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td>ReLU</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td rowspan="3">4</td><td>Conv1d</td><td>adapter_dim → adapter_dim</td><td>3</td><td>3</td><td>1</td><td>×3</td><td>L/24</td></tr><tr><td>BatchNorm1d</td><td>adapter_dim → adapter_dim</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td>ReLU</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td>5</td><td>AvgPool1d</td><td>adapter_dim → adapter_dim</td><td>5</td><td>8</td><td>0</td><td>×8</td><td>L/192</td></tr><tr><td rowspan="2">6</td><td>Linear</td><td>adapter_dim → CTC_dim</td><td>-</td><td>-</td><td>-</td><td>×1</td><td>L/192</td></tr><tr><td>ReLU</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr></table>
+
+Table 1. Architecture of the adapter for temporal downsampling, where the input length is  $T = 1568$  and final output length is  $T = 8$ .
+
+Classification (CTC) loss [18]:
+
+$$
+\mathcal {L} _ {\mathrm {C T C}} = - \ln \left(\sum_ {\alpha \in \mathcal {A} (\mathbf {P h})} P _ {\mathrm {c t c}} (\alpha \mid \mathbf {X})\right), \tag {5}
+$$
+
+where  $\mathbf{Ph} = (ph_1, \ldots, ph_m)$  is the ground-truth phoneme sequence,  $\mathcal{A}(\mathbf{Ph})$  is the set of all valid alignments, and  $P_{\mathrm{ctc}}(\alpha \mid \mathbf{X})$  is the probability of alignment  $\alpha$ . Minimizing  $\mathcal{L}_{\mathrm{CTC}}$  enables the model to learn frame-to-phoneme mappings without explicit per-frame labels. At inference, we perform beam search over the CTC log probabilities to decode the final phoneme sequence.
+
+# 3.5. Phoneme-to-Sentence Reconstruction Using LLM
+
+The predicted phoneme sequence  $\widehat{\mathbf{Ph}}$  is derived by decoding the CTC logits from Eq. (4) via beam search. We then feed  $\widehat{\mathbf{Ph}}$  into a Large Language Model (LLM) that is finetuned to map phonemes to sentences:
+
+$$
+\mathbf {S} = \operatorname {L L M} \left(\widehat {\mathbf {P h}}\right) = \left(s _ {1}, s _ {2}, \dots , s _ {M}\right). \tag {6}
+$$
+
+LoRA Fine-Tuning. To efficiently adapt the LLM, we employ Low-Rank Adaptation (LoRA) [22], injecting two low-rank matrices  $\mathbf{A}$  and  $\mathbf{B}$  into each linear layer. This greatly reduces the number of trainable parameters compared to full-model fine-tuning. Specifically, for a linear layer  $\mathbf{W}_{\mathrm{orig}} \in \mathbb{R}^{d \times d}$ :
+
+$$
+\mathbf {W} ^ {\prime} = \mathbf {W} _ {\text {o r i g}} + \mathbf {A B}, \quad \mathbf {A} \in \mathbb {R} ^ {d \times r}, \mathbf {B} \in \mathbb {R} ^ {r \times d}, \tag {7}
+$$
+
+where  $r \ll d$ . Only  $\mathbf{A}$  and  $\mathbf{B}$  are updated during training.
+
+Pretraining LLM Objective. We fine-tune the LLM on a large phoneme  $\rightarrow$  sentence corpus generated from WikiText [34]. Let  $\mathbf{S} = (s_1, \ldots, s_M)$  be the ground-truth sentence for a phoneme sequence  $\mathbf{Ph}$ . We employ the cross-entropy loss:
+
+$$
+\mathcal {L} _ {\mathrm {C E}} = - \sum_ {t = 1} ^ {M} \ln p \left(s _ {t} \mid \mathbf {P h}, s _ {1: t - 1}\right), \tag {8}
+$$
+
+where  $p(s_{t} \mid \mathbf{Ph}, s_{1:t-1})$  is the likelihood of predicting the correct word  $s_{t}$ , given the phoneme sequence and previously generated words. During inference, we prompt the LLM with  $\widehat{\mathbf{Ph}}$  to generate  $\widehat{\mathbf{S}}$ .
+
+# 3.6. Training Procedure
+
+Stage 1: Video  $\rightarrow$  Phoneme. We train the visual extractor, adapter, and CTC head to minimize  $\mathcal{L}_{\mathrm{CTC}}$  (Eq. 5) on video-phoneme pairs. By learning without strict alignment constraints, the model remains flexible to a wide range of speaking speeds and styles.
+
+Stage 2: Phoneme  $\rightarrow$  Sentence. We independently fine-tune the LLM via cross-entropy loss (Eq. 8) on text-based phoneme  $\rightarrow$  sentence datasets. This leverages easily obtained text data, allowing the LLM to learn phonetic-linguistic mappings separately from visual modeling.
+
+Inference. We compose the trained modules to form:
+
+$$
+\mathbf {X} \xrightarrow {f _ {\mathrm {v i s}}} \widehat {\mathbf {P h}} \xrightarrow {f _ {\mathrm {L L M}}} \widehat {\mathbf {S}},
+$$
+
+where  $\widehat{\mathbf{Ph}}$  is decoded from CTC logits, and  $\widehat{\mathbf{S}}$  is the final sentence. This approach is highly modular and data-efficient, as each stage (visual and linguistic) is learned with its own objective and dataset, yet easily integrated at inference time.
+
+# 4. Experiments
+
+This section outlines the datasets and pre-processing methods used for training the model architecture.
+
+# 4.1. Data
+
+The visual feature extractor, the adapter with downsampling, and CTC head are trained on the LRS2 [47] and LRS3 [3] datasets. The LLM is fine-tuned on the WikiText [34] dataset.
+
+LRS2 [47]: The LRS2 dataset consists of 144,482 video clips of spoken sentences from BBC television, consisting of approximately 224.5 hours of footage and each sentences is up to 100 characters in length. The videos are divided into a pre-training set with 96,318 utterances (195 hours), a training set with 45,839 utterances (28 hours), a validation set with 1,082 utterances (0.6 hours) and a test set with 1,243 utterances (0.5 hours).
+
+LRS3 [3]: The LRS3 dataset describes the largest public audio-visual English dataset collected consists of clips from over 5,000 TED and TEDx talks totaling 438.9 hours. It contains 438.9 hours with 151,819 utterances. Specifically, there are 118,516 utterances in the 'pre-train' set (408 hours), 31,982 utterances in the 'train-val' set (30 hours) and 1,321 utterances in the 'test' set (0.9 hours).
+
+In our setting, unlike other approaches, we do not utilise the pre-training set and train our model on only the 28-hour and 30-hour partitions.
+
+Phoneme Sentence Pairs. Finally, the dataset used for training the LLM was the WikiText [34] dataset with sentences converted into lists of phonemes using the CMU dictionary [52]. These phonemes were then masked and paired with the original sentences.
+
+# 4.2. Evaluation Metrics
+
+Following previous works [11, 47] we adopt Word Error Rate [23] (WER) as our evaluation metric for Lip-Reading. WER calculates the percentage of errors in the predicted text compared to the ground truth, accounting for substitutions, insertions, and deletions. Similar to previous studies [3, 11, 41, 42, 47], we report results on all recent AV-ASR and V-ASR approaches.
+
+# 5. Results
+
+In this section, we present the results of our method and compare them against those of existing SOTA approaches.
+
+Public vs Private Data on LRS3 [3]: In Table 1, we compare different V-ASR and AV-ASR models based on their WER on the LRS3 dataset, we split the models into two categories; Fully Supervised models with publicly available data and models trained on large-scale non-publicly available datasets.
+
+The Fully Supervised Models rely solely on publicly available labelled datasets with the best WER being  $40.6\%$ , which was achieved by VTP [41]. The non-public dataset models leverage extensive, proprietary datasets, achieving much lower WERs with the best being  $12.5\%$  by LP [10].
+
+LRS3 [3]: In Table 1 we compare our results against other methods based on results from the LRS3 dataset. The results in Table 1 demonstrate a consistent trend: models incorporating extensive self-supervised pre-training followed by fine-tuning achieve lower WERs than those using only supervised approaches. However, our model achieves the lowest WER on the LRS3 dataset at  $18.7\%$  without self-supervised pre-training or additional fine-tuning data.
+
+LRS2 [47]: In Table 2, we compare our method against several other state-of-the-art approaches for visual-only lip reading, focusing on Word Error Rate (WER) on the LRS2 dataset. From Table 2 we can see that increasing the amount of data available helps improve a model's WER. However, we can also see that even without increasing the size of the dataset, our model still outperforms other models that only use publicly available data, achieving the lowest WER at  $20.8\%$ .
+
+Therefore as shown in Tables 2 and 3, we can see that our
+
+<table><tr><td>Method</td><td>Unlabeled (hrs)</td><td>Labeled (hrs)</td><td>WER (%)</td></tr><tr><td colspan="4">Fully supervised models with publicly available data</td></tr><tr><td>ASR distillation [4]</td><td>-</td><td>590</td><td>68.8</td></tr><tr><td>Conv-Seq2Seq [60]</td><td>-</td><td>855</td><td>60.1</td></tr><tr><td>Discriminative AVSR [57]</td><td>-</td><td>590</td><td>57.8</td></tr><tr><td>Hyb. + Conformer [29]</td><td>-</td><td>590</td><td>43.3</td></tr><tr><td>VTP [41]</td><td>-</td><td>698</td><td>40.6</td></tr><tr><td colspan="4">Trained on large-scale non-publicly available datasets</td></tr><tr><td>Deep-AV-SR [2]</td><td>-</td><td>1,519</td><td>58.9</td></tr><tr><td>Large-scale AV-SR [46]</td><td>-</td><td>3,886</td><td>55.1</td></tr><tr><td>RNN-T [30]</td><td>-</td><td>31,000</td><td>33.6</td></tr><tr><td>VTP [41]</td><td>-</td><td>2,676</td><td>30.7</td></tr><tr><td>ViT-3D [44]</td><td>-</td><td>90,000</td><td>17.0</td></tr><tr><td>LP [10]</td><td>-</td><td>1,000,000</td><td>12.5</td></tr><tr><td colspan="4">Self-supervised pre-training + Supervised fine-tuning on LRS3</td></tr><tr><td>LiRA [28]</td><td>433</td><td>30</td><td>71.9</td></tr><tr><td>ASR distillation [4]</td><td>334</td><td>590</td><td>59.8</td></tr><tr><td>LiteVSR [24]</td><td>639</td><td>59</td><td>45.7</td></tr><tr><td>ES3 [61]</td><td>433</td><td>30</td><td>43.5</td></tr><tr><td>AV-HuBERT Large [45]</td><td>1,759</td><td>30</td><td>32.5</td></tr><tr><td>Lip2Vec [13]</td><td>1,759</td><td>30</td><td>31.2</td></tr><tr><td>Sub-Word[41]</td><td>2,676</td><td>2,676</td><td>30.7</td></tr><tr><td>SynthVSR [25]</td><td>3,652</td><td>438</td><td>27.9</td></tr><tr><td>Whisper [42]</td><td>1,759</td><td>30</td><td>25.5</td></tr><tr><td>Auto-AVSR [27]</td><td>1,759</td><td>433</td><td>25.0</td></tr><tr><td>LLaMA-AVSR [8]</td><td>-</td><td>1756</td><td>24.0</td></tr><tr><td>RAVeN [19]</td><td>1759</td><td>433</td><td>23.1</td></tr><tr><td>USR [20]</td><td>1,326</td><td>433</td><td>21.5</td></tr><tr><td colspan="4">Supervised finetuning only on LRS3</td></tr><tr><td>Ours</td><td>-</td><td>30</td><td>18.7</td></tr></table>
+
+![](images/2309680e9fc8df15b81d4649f6ee4748410ca448eec1e3eceae9d15922ee1af3.jpg)
+
+![](images/7754bf5651d54cce94bc54485751dcdadbf2b73bbffaf38ff8144f09528b3941.jpg)
+
+![](images/5241836870d9d791c95f8c9159eb00903de7d664b498f268b9fcd5a3aec71514.jpg)
+
+![](images/276819e67a971584a49e4d208e6f4455af7997117101357a4747b9be5c95485c.jpg)
+
+![](images/513f852c48a8ad848e828353e103328bee41306c6fde1fbc3a53f93a75960893.jpg)
+
+![](images/015424b644c8663954efc1f779652ecdfffb7b59c66d495a4885f69ecca87fda.jpg)
+
+![](images/0aa3d8102ac7868c0e1bcdc5bebd27111bf0d66867bd2c741b52605324fd8002.jpg)
+
+![](images/4cf7ce9a562052ae9dd55dc21fbc0e6a62e0fbb85558cbb0d60521588ee3f667.jpg)
+
+![](images/0a1a3fecf95803ab3592d83a37edcb4301f5f641561753b2b269d90f613aa441.jpg)
+
+![](images/a4c6362e1c7e6d602efec1dab005d55e5828f46a9a4f415b27088b2f964df35e.jpg)
+
+![](images/5af5aa88bf55dd63f55f2c27f5fb8caddf21de23c2f3aca92a2e4b6dfa0146fa.jpg)
+
+![](images/861cf9126b1057f0a3876f194369d5f406de253a33de9d56ccb82e45cfed29e5.jpg)
+
+![](images/9fa901c1b1ede2859a34c2e0b0be5fd01e9a440a82cb5245f16a3afbdcb64885.jpg)
+
+Table 2. Comparison of Word Error Rate (WER) across different models for visual-only speech recognition on the LRS3 dataset. The table is divided into three sections: fully supervised models trained on publicly available data, models trained on large-scale non-public datasets, and models that leverage self-supervised pre-training with supervised fine-tuning on LRS3 [3]. Each entry details the amount of labelled and unlabelled data used and the resulting WER. Our approach, tested with various language models, achieves the lowest WER of  $22.1\%$  with the Llama 3.2-3B model, demonstrating the effectiveness of our phoneme-based, LLM-assisted approach.  
+
+<table><tr><td>Predicted 
+Phoneme</td><td>W, AA, Z</td><td>DH, AH</td><td>K, AY, N, D</td><td>AH, V</td></tr><tr><td>Predicted 
+Word</td><td>Was</td><td>the</td><td>kind</td><td>of</td></tr><tr><td>GT</td><td>Was</td><td>the</td><td>kind</td><td>of</td></tr><tr><td>Predicted 
+Phoneme</td><td>AH, B, S, EH, SH, 
+AH, N</td><td>DH, AE, T</td><td>HH, AE, D</td><td>HH, ER</td></tr><tr><td>Predicted 
+Word</td><td>obsession</td><td>that</td><td>had</td><td>her</td></tr><tr><td>GT</td><td>obsession</td><td>that</td><td>had</td><td>her</td></tr></table>
+
+Figure 3. Example of the model's phonetic and sentence outputs from a sample in the LRS3 dataset [3]. The table illustrates the model's ability to predict a sequence of phonemes from visual input, which are then reconstructed into a coherent sentence by the LLM. In this example all of the phonemes are predicted correctly and the words are recreated correctly.  
+
+<table><tr><td>Method</td><td>Unlabeled (hrs)</td><td>Labeled (hrs)</td><td>WER (%)</td></tr><tr><td colspan="4">Fully supervised models with publicly available data</td></tr><tr><td>Auto-AV-SR [27]</td><td>-</td><td>818</td><td>27.9</td></tr><tr><td>Auto-AV-SR [27]</td><td>-</td><td>3448</td><td>14.6</td></tr><tr><td colspan="4">Self-supervised pre-training + Supervised finetuning on LRS2 for AVSR</td></tr><tr><td>LiRA [28]</td><td>1,759</td><td>433</td><td>38.8</td></tr><tr><td>ES3 [61]</td><td>1,759</td><td>223</td><td>26.7</td></tr><tr><td>Sub-Word [41]</td><td>2,676</td><td>2,676</td><td>22.6</td></tr><tr><td>RAVeN [19]</td><td>1,759</td><td>223</td><td>17.9</td></tr><tr><td>USR [20]</td><td>1,759</td><td>223</td><td>15.4</td></tr><tr><td colspan="4">Supervised finetuning only on LRS2</td></tr><tr><td>Ours</td><td>-</td><td>28</td><td>20.8</td></tr></table>
+
+Table 3. Comparison of Word Error Rate (WER) for various self-supervised pre-training and supervised fine-tuning methods evaluated on the LRS2 [47] dataset. Our method achieves competitive performance with other methods, without the requirement of extensive pre-training, additional labelled data, or additional audio modality as in the best performing approaches. We are the only method to train on only the LRS2 train dataset.
+
+method provides a lower WER with the same amount, or less data, without the requirement of self-supervised pretraining as current approaches [4, 13, 24, 25, 27, 28, 42, 45]. This large increase in performance and generalisation can be explained by the powerful combination of the multi-task objectives and the ability of the LLM to reconstruct sentences accurately even when phonemes are predicted incorrectly. In Fig. 4, we find that the phonetic output of the visual encoder is sometimes incorrect or misses phonemes, but the LLM is able to still reconstruct the correct word (as in the example  $z \rightarrow s$ ). We also show how the model still struggles with tricky homophones such as your and you're. In Fig. 3, we show an additional example of the model's outputs at both phonetic and word levels, demonstrating how phonemes are correctly predicted and combined by the LLM for sentence reconstruction. For further examples please see the appendix.
+
+# 5.1. Ablations
+
+In this section, we aim to understand the performance contribution of each component in the network and their performance on visual  $\rightarrow$  phoneme prediction and phoneme  $\rightarrow$  word reconstruction, alongside ablations related to architectural decisions, and the limitations of the approach.
+
+Visual  $\rightarrow$  Phoneme: The confusion matrix in Fig. 5 shows phoneme prediction performance of the model on the LRS3 [3] dataset. The matrix indicates that some phonemes are misclassified more frequently than others, as seen by the off-diagonal elements, with the worst missed classifications being highlighted with a red bounding box. This bias could be due to the similarity in the visual representation of these sounds, making it challenging for the model to distinguish between them. For instance, /S/ and /Z/ may often be con-
+
+![](images/4e8fc1ec0b0814031a2d09ff8bf89e914bb5512b94eccdc80d0511cedb025fbc.jpg)  
+Figure 4. Comparison of the model's phonetic and sentence outputs with ground truth from a sample in the LRS3 [3] dataset. Red shows an incorrect prediction, M shows a missing prediction and green shows a correction from phonemes to words. In this example, even though the model incorrectly predicts certain phonemes, the LLM can correctly recreate the word but struggles to recreate homophones.
+
+<table><tr><td>Predicted 
+Phoneme</td><td>Y, UH, R</td><td>JH, AH, Z, T</td><td>W, AH, N</td><td>AH, V</td></tr><tr><td>Predicted 
+Word</td><td>Your</td><td>just</td><td>one</td><td>of</td></tr><tr><td>GT</td><td>You&#x27;re</td><td>just</td><td>one</td><td>of</td></tr><tr><td>Predicted 
+Phoneme</td><td>DH, AH</td><td>P, IY, P, &lt;M&gt;, L</td><td>HH, UW</td><td>HH, AE, S, N, T</td></tr><tr><td>Predicted 
+Word</td><td>the</td><td>people</td><td>who</td><td>hasn&#x27;t</td></tr><tr><td>GT</td><td>the</td><td>people</td><td>who</td><td>hasn&#x27;t</td></tr></table>
+
+fused due to similar lip movements, leading to miss classifications between these phonemes.
+
+![](images/f7d26b30e521eb6a97cd06f3b94f654e12b2cf2e2cd055439660d384e40abed9.jpg)  
+Figure 5. Confusion matrix showing the performance on isolated phonemes of the LRS3 [3] dataset. We observe a very high match rate between the predicted phonemes and the ground truth. In red, we show the most difficult phonemes for our model to identify.
+
+Phoneme  $\rightarrow$  Word: In Table 4 we take the top 5 words in alphabetical order that are misclassified when predicting words directly from phonemes after the LLM pretraining stage. We can observe that these are words that may occur infrequently in the training set and are missclassified due to some phonetic similarities between the words.
+
+Different LLMs: In this ablation study, we evaluate three different Language models on our model architecture to
+
+<table><tr><td>Phonemes</td><td>True Word</td><td>Predicted Word</td></tr><tr><td>/EH/ /R/ /AH/ /N/ /S/ /AH/ /N/</td><td>Aaronson</td><td>Erasmus</td></tr><tr><td>/AA/ /B/ /ER/ /G/</td><td>Aaberg</td><td>Braggartly</td></tr><tr><td>/AH/ /S/ /EH/ /R /AH/</td><td>Acerra</td><td>Esraa</td></tr><tr><td>/AE/ /S/ /AH/ /T/ /EY /T/</td><td>Acetate</td><td>Asquette</td></tr><tr><td>/AH/ /K/ /AH/ /S/ /T/ /AH/ /M/ /D/</td><td>Accustomed</td><td>Accosted</td></tr></table>
+
+understand the impact on word error rate (WER) performance. The first model, GPT 2 [43], serves as our baseline, consisting of the default GPT 2 [43] without fine-tuning on lip-reading datasets. This model achieves a WER of  $23.9\%$ , indicating the core components' foundational capabilities. The second model, Llama 3.2-1B [51] introduces fine-tuning on phoneme word pairs, enhancing the model's contextual understanding of phoneme sequences. With this addition, Llama 3.2-1B [51] has a substantial improvement, reducing the WER to  $22.8\%$ . Finally, Llama 3.2-3B [51] also incorporates fine-tuning for phoneme-to-word reconstruction. However, Llama 3.2-3B [51] is a larger model with more parameters than the Llama 3.2-1B [51] model, further lowering the WER to  $18.7\%$ . This progression highlights the significant contributions of fine-tuning for purpose-specific LLMs.
+
+Table 4. Examples of misclassification's between predicted words and ground truth based on phoneme sequences. Each row presents a sequence of phonemes, the intended (true) word, and the model's predicted word. Differences in predictions illustrate common misclassifications, often due to phonetic similarities between visually indistinct sounds.  
+
+<table><tr><td>Model</td><td>Parameters (B)</td><td>WER (%)</td></tr><tr><td>GPT-2 Small [43]</td><td>0.12</td><td>33.9</td></tr><tr><td>Llama 3.2-1B [51]</td><td>1.0</td><td>22.8</td></tr><tr><td>Llama 3.2-3B [51]</td><td>3.0</td><td>18.7</td></tr></table>
+
+Table 5. Comparison of Word Error Rate (WER) for different large language models (LLMs) tested with our model. The table includes each model's parameter count (in billions) and resulting WER on the LRS3 [3] dataset. GPT-2 [43] Small with 0.12B parameters has the highest WER at  $23.9\%$ . In contrast, Llama 3.2-3B [51] achieves the best WER of  $18.7\%$ , highlighting that increased model capacity improves phoneme-to-word reconstruction accuracy.
+
+Freezing the CTC Head: Inspired by works that directly map visual features to pretrained ASR networks [42] we investigate the effect of initialising the CTC head in our model with weights from a pre-trained Wav2Vec2 ASR model [6]. The CTC head plays a crucial role in mapping visual features to the phoneme sequences, and we hypothesized that initialising and freezing its parameters could reduce computational overhead while preserving feature alignment from audio pre-training. To test this hypothesis a baseline model
+
+with a non-frozen CTC head and a comparison model with a frozen CTC head were trained and a comparison of the results was made as shown in Table 6. Freezing the CTC head led to a slight increase in WER of  $0.4\%$ , indicating a decrease in performance. This result suggests that while the frozen CTC head still provided a foundation for phoneme alignment, we did not require extensive audio pre-training to obtain superior performance.
+
+Varying Sample Sizes: To understand the impact of additional data on performance, we vary the quantity of data used to train both the Video  $\rightarrow$  Phoneme network and the Phoneme  $\rightarrow$  Sentence LLM and investigate the effect this has on the WER when evaluated on the LRS3 [3] dataset. By varying the percentages of data used in training, we can see the effect on the model in Figure 6. Decreasing the amount of data used to train both models has the expected effect on the WER by increasing it. However, increasing the amount of training data for the Feature Extractor has a negligible effect after 30 hours of data. However, by increasing the amount of training data for the LLM we can also increase the effectiveness of the model and decrease the total WER down further to  $17.5\%$ . The additional data is randomly sampled from both AvSpeech [17] for the visual encoder, and a BookCorpus [63] replica to match the quantity of data in WikiText.
+
+![](images/9039c2c35aa9c82d656a47cc63694fed24bdefa3cff3157abb122551594fff59.jpg)  
+Figure 6. Graph showing the effect of varying the sample size of the training datasets for both the Video  $\rightarrow$  Phoneme network (in blue) and for the Phoneme  $\rightarrow$  Sentence LLM (in red). The graph shows that decreasing the amount of data negatively impacts the model for both networks, increasing the amount of training data for the Video  $\rightarrow$  Phoneme network has negligible effect and increasing the amount of training data for the Phoneme  $\rightarrow$  Sentence LLM has a noticeable positive effect.
+
+End-To-End: For our final study, we investigate the effect of removing the video  $\rightarrow$  phoneme stage and train the
+
+model to directly predict sentences from the video features, reproducing the method in [41] but with the same ViT encoder [44], LLama LLM [51], and limited data set. As shown in Table 6 we achieve aWER of  $25.6\%$  on LRS3, demonstrating that the more parameterised transformer encoder improves performance over the CNN encoder used in [41]. However, our two stage approach still performs significantly better.
+
+<table><tr><td>Model</td><td>WER (%)</td></tr><tr><td>Ours (Two Stage)</td><td>18.7</td></tr><tr><td>ViT + LLM (End to End)</td><td>25.6</td></tr><tr><td>Sub-Word (End to End) [41]</td><td>30.7</td></tr></table>
+
+Table 6. Comparison of Word Error Rate (WER) between using the intermediary Phonemes and skipping the CTC head. Removing the intermediary representation increases the WER from 18.7 to 25.6, showing the importance of the intermediary step and the effectiveness of the phoneme centric fine-tuning.
+
+# 6. Conclusion and Limitations
+
+We have introduced a two-stage, phoneme-centric framework for visual-only lip reading that first predicts phonemes from lip movements and then reconstructs coherent sentences via a fine-tuned LLM. This design sharply reduces word error rates on benchmark datasets (LRS2 [47] and LRS3 [3]), providing improvements of  $3.5\%$  and  $6.8\%$  WER, respectively, over existing state-of-the-art approaches that focus on end-to-end connectionist approaches.
+
+Despite these gains, the approach remains limited by the visual ambiguity of certain phonemes. Lip movements alone cannot always differentiate minimal pairs such as /s/ vs. /z/, and words that share near-identical phoneme sequences (e.g., "Hello" vs. "Hallow") can still result in misclassification. We plan to address these issues by refining the phoneme-to-word generation process, potentially through deeper linguistic context modeling and more specialized phoneme embeddings. Future work may also explore speaker-adaptive fine-tuning or multi-frame alignment strategies to better handle subtle visual distinctions.
+
+# Acknowledgements
+
+This work was supported by the SNSF project 'SMILE II' (CRSII5 193686), the Innosuisse IICT Flagship (PFFS-21-47), EPSRC grant APP24554 (SignGPT-EP/Z535370/1) and through funding from Google.org via the AI for Global Goals scheme. This work reflects only the author's views and the funders are not responsible for any use that may be made of the information it contains.
+
+# References
+
+[1] Emre Afacan and Gültekin Demirci. A survey on automatic lip-reading: Challenges and future directions. Multimedia Tools and Applications, 2019. 1, 3  
+[2] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 2018. 3, 6  
+[3] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 1, 3, 5, 6, 7, 8  
+[4] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2143-2147. IEEE, 2020. 3, 6  
+[5] Yannis M Assael, Brendan Shillingford, Shimon Whiteson, and Nando De Freitas. Lipnet: End-to-end sentence-level lipreading. arXiv preprint arXiv:1611.01599, 2016. 3  
+[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 2020. 7  
+[7] Helen L Bear, Richard W Harvey, Barry-John Theobald, and Yuxuan Lan. Which phoneme-to-viseme maps best improve visual-only computer lip-reading? In International symposium on visual computing. Springer, 2014. 1  
+[8] Umberto Cappellazzo, Minsu Kim, Honglie Chen, Pingchuan Ma, Stavros Petridis, Daniele Falavigna, Alessio Brutti, and Maja Pantic. Large language models are strong audio-visual speech recognition learners, 2025. 6  
+[9] Luca Cappelletta and Naomi Harte. Viseme definitions comparison for visual-only speech recognition. In 2011 19th European Signal Processing Conference. IEEE, 2011. 3  
+[10] Oscar Chang, Hank Liao, Dmitriy Serdyuk, Ankit Shahy, and Olivier Siohan. Conformer is all you need for visual speech recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024. 5, 6  
+[11] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian Conference on Computer Vision, 2016. 3, 5  
+[12] Martin Cooke, Jon Barker, Stuart Cunningham, and Xu Shao. Grid av speech corpus sample. The Journal of the Accoustical Society of America, 2013. 3  
+[13] Yasser Abdelaziz Dahou Djilali, Sanath Narayan, Haithem Boussaid, Ebtessam Almazrouei, and Merouane Debbah. Lip2vec: Efficient and robust visual speech recognition via latent-to-latent visual to audio representation mapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3, 6  
+[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3  
+[15] Randa El-Bialy, Daqing Chen, Souheil Fenghour, Walid Hussein, Perry Xiao, Omar H. Karam, and Bo Li. Developing phoneme-based lip-reading sentences system for silent
+
+speech recognition. CAAI Transactions on Intelligence Technology, 8(1):129-138, 2023. 3  
+[16] Randa El-Bialy, Daqing Chen, Souheil Fenghour, Walid Hussein, Perry Xiao, Omar H Karam, and Bo Li. Developing phoneme-based lip-reading sentences system for silent speech recognition. CAAI Transactions on Intelligence Technology, 2023. 3  
+[17] A. Ephrat, I. Mosseri, O. Lang, T. Dekel, K Wilson, A. Hassidim, W. T. Freeman, and M. Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 8  
+[18] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, 2006. 4  
+[19] Alexandros Haliassos, Pingchuan Ma, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Jointly learning visual and auditory speech representations from raw data. arXiv preprint arXiv:2212.06246, 2022. 6  
+[20] Alexandros Haliassos, Rodrigo Mira, Honglie Chen, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Unified speech recognition: A single model for auditory, visual, and audio-visual inputs. arXiv preprint arXiv:2411.02256, 2024. 6  
+[21] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 2021. 3  
+[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 2022. 4  
+[23] F. Jelinek, L. Bahl, and R. Mercer. Design of a linguistic statistical decoder for the recognition of continuous speech. IEEE Transactions on Information Theory, 21(3), 1975. 5  
+[24] Hendrik Laux, Emil Mededovic, Ahmed Hallawa, Lukas Martin, Arne Peine, and Anke Schmeink. Litevsr: Efficient visual speech recognition by learning from speech representations of unlabeled data. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024. 3, 6  
+[25] Xubo Liu, Egor Lakomkin, Konstantinos Vougioukas, Pingchuan Ma, Honglie Chen, Ruiming Xie, Morrie Doulaty, Niko Moritz, Jachym Kolar, Stavros Petridis, et al. Synthvsr: Scaling up visual speech recognition with synthetic supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 6  
+[26] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 4  
+[27] Pingchuan Ma and Alexandros et al. Haliassos. Auto-avsr: Audio-visual speech recognition with automatic labels. In
+
+IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023. 3, 6  
+[28] Pingchuan Ma and Rodrigo et al. Mira. Lira: Learning visual speech representations from audio through self-supervision. arXiv, 2021. 3, 6  
+[29] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with conformers. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021. 6  
+[30] Takaki Makino, Hank Liao, Yannis Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In IEEE automatic speech recognition and understanding workshop (ASRU). IEEE, 2019. 3, 6  
+[31] William D Marslen-Wilson. Functional parallelism in spoken word-recognition. Cognition, 25(1-2), 1987. 2  
+[32] William D Marslen-Wilson and Alan Welsh. Processing interactions and lexical access during word recognition in continuous speech. Cognitive psychology, 10(1), 1978. 2  
+[33] Harry McGurk and John MacDonald. Hearing lips and seeing voices. Nature, 1976. 1  
+[34] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. 4, 5  
+[35] Ara V Nefian, Luhong Liang, Xiaobo Pi, Liu Xiaoxiang, Crusoe Mao, and Kevin Murphy. A coupled hmm for audiovisual speech recognition. In 2002 IEEE International Conference on Acoustics, Speech, and Signal Processing. IEEE, 2002. 3  
+[36] Eng-Jon Ong and Richard Bowden. Learning temporal signatures for lip reading. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops). IEEE, 2011. 3  
+[37] Stavros Petridis and Maja Pantic. Deep complementary bottleneck features for visual speech recognition. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3  
+[38] Stavros Petridis, Themos Stafylakis, Pingchuan Ma, Georgios Tzimiropoulos, and Maja Pantic. Audio-visual speech recognition with a hybrid ctc/attention architecture. In 2018 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2018. 3  
+[39] Javad Peymanfard, Vahid Saeedi, Mohammad Reza Mohammadi, Hossein Zeinali, and Nasser Mozayani. Leveraging visemes for better visual speech representation and lip reading, 2023. 3  
+[40] Gerasimos Potamianos, Chalapathy Neti, Guillaume Gravier, Ashutosh Garg, and Andrew W Senior. Recent advances in the automatic recognition of audiovisual speech. Proceedings of the IEEE, 91(9), 2003. 3  
+[41] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 5, 6, 8  
+[42] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Speech recognition models are strong lip-readers. In Proc. Interspeech 2024, 2024. 3, 5, 6, 7
+
+[43] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. arXiv, 2018. 3, 7  
+[44] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition for single and multi-person video. arXiv preprint arXiv:2201.10439, 2022. 3, 4, 6, 8  
+[45] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. *ICLR*, 2022. 3, 6  
+[46] Brendan Shillingford, Yannis Assael, Matthew W Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, et al. Large-scale visual speech recognition. InterSpeech, 2018. 3, 6  
+[47] Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Lip reading sentences in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 3, 5, 6, 8  
+[48] Themos Stafylakis and Georgios Tzimiropoulos. Zero-shot keyword spotting for visual speech recognition in-the-wild. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-529, 2018. 3  
+[49] Kwanchiva Thangthai, Helen L Bear, and Richard Harvey. Comparing phonemes and visemes with dnn-based lipreading. arXiv preprint arXiv:1805.02924, 2018. 3  
+[50] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35, 2022. 3  
+[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3, 7, 8  
+[52] Carnegie Mellon University. The cmu pronouncing dictionary, 1993. http://www.speech.cs.cmu.edu/cgi-bin/cmudict.2,5  
+[53] 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, 2017. 3  
+[54] Michael Wand, Jan Koutnik, and Jürgen Schmidhuber. Lipreading with long short-term memory. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3  
+[55] Michael Wand, Jan Koutnik, and Jürgen Schmidhuber. Lipreading with long short-term memory. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3  
+[56] Xinshuo Weng and Kris Kitani. Learning spatio-temporal features with two-stream deep 3d cnn for lipreading. arXiv preprint arXiv:1905.02540, 2019. 3  
+[57] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, 2020. 3, 6  
+[58] Jeong Hun Yeo, Seunghee Han, Minsu Kim, and Yong Man Ro. Where visual speech meets language: Vsp-llm frame
+
+work for efficient and context-aware visual speech processing. arXiv preprint arXiv:2402.15151, 2024. 2, 3  
+[59] Jeong Hun Yeo, Chae Won Kim, Hyunjun Kim, Hyeongseop Rha, Seunghee Han, Wen-Huang Cheng, and Yong Man Ro. Personalized lip reading: Adapting to your unique lip movements with vision and language. arXiv preprint arXiv:2409.00986, 2024. 2, 3  
+[60] Xingxuan Zhang, Feng Cheng, and Shilin Wang. Spatiotemporal fusion based convolutional sequence learning for lip reading. In Proceedings of the IEEE/CVF International conference on Computer Vision, 2019. 3, 6  
+[61] Yuanhang Zhang, Shuang Yang, Shiguang Shan, and Xilin Chen. Es3: Evolving self-supervised learning of robust audio-visual speech representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 6  
+[62] Ziheng Zhou, Guoying Zhao, Xiaopeng Hong, and Matti Pietikainen. A review of recent advances in visual speech decoding. Image and vision computing, 32(9), 2014. 3  
+[63] Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, 2015. 8

2025/VALLR_ Visual ASR Language Model for Lip Reading/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0556c17ccf71be7e1012a391f3b4b84247ff5e661599d1da12f4b89a7c77124f
+size 540249

2025/VCA_ Video Curious Agent for Long Video Understanding/a73c253b-5afa-42a6-b79a-31559f118997_content_list.json

@@ -0,0 +1,1768 @@
+[
+    {
+        "type": "text",
+        "text": "VCA: Video Curious Agent for Long Video Understanding",
+        "text_level": 1,
+        "bbox": [
+            202,
+            130,
+            795,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zeyuan Yang†  \nUniversity of Massachusetts, Amherst  \nyangzeyuan@umass.edu",
+        "bbox": [
+            173,
+            179,
+            475,
+            233
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Xueyang Yu  \nUniversity of Massachusetts, Amherst  \nxueyangyu@umass.edu",
+        "bbox": [
+            173,
+            241,
+            475,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Delin Chen†  \nUniversity of Massachusetts, Amherst delinchen@umass.edu",
+        "bbox": [
+            521,
+            179,
+            823,
+            233
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Maohao Shen  \nMassachusetts Institute of Technology  \nmaohao@mit.edu",
+        "bbox": [
+            521,
+            241,
+            823,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Chuang Gan  \nUniversity of Massachusetts, Amherst  \nganchuang1990@gmail.com",
+        "bbox": [
+            346,
+            301,
+            650,
+            353
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            248,
+            388,
+            326,
+            404
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Long video understanding poses unique challenges due to their temporal complexity and low information density. Recent works address this task by sampling numerous frames or incorporating auxiliary tools using LLMs, both of which result in high computational costs. In this work, we introduce a curiosity-driven video agent with self-exploration capability, dubbed as \"VCA\". Built upon VLMs, VCA autonomously navigates video segments and efficiently builds a comprehensive understanding of complex video sequences. Instead of directly sampling frames, VCA employs a tree-search structure to explore video segments and collect frames. Rather than relying on external feedback or reward, VCA leverages VLM's self-generated intrinsic reward to guide its exploration, enabling it to capture the most crucial information for reasoning. Experimental results on multiple long video benchmarks demonstrate our approach's superior effectiveness and efficiency.",
+        "bbox": [
+            88,
+            421,
+            485,
+            679
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            89,
+            700,
+            220,
+            715
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "There is a growing demand for developing long-form video understanding techniques [48, 69] capable of extracting valuable information from extended visual narratives, such as surveillance footage, documentaries, and instructional videos. Analyzing these videos [53, 71], which can range from minutes to hours, requires models that can process multimodal data and perform reasoning over extremely long sequences [4, 31, 60].",
+        "bbox": [
+            89,
+            724,
+            482,
+            845
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Previous works commonly utilize two-stream networks to capture spatial and temporal information [18, 19] or in",
+        "bbox": [
+            89,
+            847,
+            482,
+            876
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/b3a34f2280701b001980d124b23ff96bfa7878b2f2423e84306fc895c68a5841.jpg",
+        "image_caption": [
+            "Figure 1. Accuracy v.s. Computation Cost and Number of Observed Frames. Compared to baselines, our agent demonstrates superior performance with higher efficiency (observing fewer frames and incurring comparable or lower computational cost) on EgoSchema, using both GPT-4o and Qwen2-VL-72B."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            535,
+            388,
+            710,
+            554
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/91ceaf3d095c177613ad02b2b753cad522a8335adc05cf71027f60062516008a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            712,
+            388,
+            888,
+            554
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "troduce 3D operations [8, 57]. Recent approach [32, 75] leverages the long-sequence reasoning capabilities of Large Language Models (LLMs) [1, 64] for long-form video question answering, often by converting videos into densely and uniformly sampled frames [33, 38, 55]. However, this approach is usually inefficient due to the inherent imbalance and varying information density across video segments [61, 78]. Some segments are rich in critical details necessary for answering questions, while others contain only trivial information [66]. Uniform sampling [49, 60], which might focus on redundant frames that increase the computational burden and may overlook frames with essential information, often fails to efficiently utilize the video's dynamic structure for question answering [17, 62].",
+        "bbox": [
+            511,
+            641,
+            906,
+            852
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address this limitation, recent approaches [29, 34, 36, 68] employ complex architectures [63, 85] or specialized preprocessing techniques [44, 65] to improve the efficiency",
+        "bbox": [
+            511,
+            854,
+            908,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV 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": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "† Equal Contribution",
+        "bbox": [
+            107,
+            886,
+            218,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "20168",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "of long video content processing. These methods [62, 66] often convert videos into text by captioning densely sampled frames or incorporating memory modules [5, 17, 39] to store essential information. However, these approaches still incur significant computational costs. For example, captioning an hour-long video with Qwen2-VL-72B at just 1 fps requires over 3 hours on a single H100. Additionally, these methods [12] identify key frames based on image similarity rather than a holistic understanding of the video content, which can result in missing relevant video features. This necessitates the need for an efficient and scalable solution for long-form video understanding.",
+        "bbox": [
+            89,
+            90,
+            483,
+            272
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this work, we introduce VCA, a curiosity-driven agent inspired by how humans tackle the long video understanding task. Given a question, human is usually intrinsically motivated to identify relevant information in the video without external knowledge. In practise, human tends to build a global understanding of the video, then curiously zoom in to explore specific sections in greater detail. Motivated by this process, we propose a novel video agent framework, dubbed as \"VCA\", to emulate such behavior. VCA treats the video as an environment to explore, actively navigating its contents and identifying segments that merit deeper analysis. VCA incorporates three essential techniques: (1) Tree-search Exploration: Rather than selecting specific frames [62], VCA uses a tree-search structure to adaptively explore the video segments in a structured way. (2) Reward Model: VCA leverages an intrinsic reward scoring mechanism to guide exploration toward the segments most relevant to the query. (3) Memory Management: VCA employs a fixed memory buffer to store key information and discard irrelevant frames, effectively reducing computational costs and avoiding memory overflow. The key difference between VCA and prior works is outlined in Fig. 2. Extensive experiments and superior performance across multiple benchmarks demonstrate that our proposed VCA framework significantly improves an agent's effectiveness and efficiency in reasoning over extended video content. More concretely, our contributions are threefold,",
+        "bbox": [
+            91,
+            313,
+            483,
+            720
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- VCA is a long video understanding agent that can self-exlore the most informative video segments based on its intrinsic reward, without relying on any auxiliary tools.",
+            "- VCA is easy to implement, training-free, and is flexible to integrate with any vision-language model.",
+            "- As highlighted in Fig. 1, VCA outperforms existing baselines while observing fewer frames and incurring comparable or lower computational cost, demonstrating both effectiveness and efficiency."
+        ],
+        "bbox": [
+            89,
+            761,
+            483,
+            898
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/86d97b7ae7cef89e8e72f2b8f1a26808351cdd81b23ff0ab9e2edc6c90c89ae7.jpg",
+        "image_caption": [
+            "Figure 2. Comparison of VCA with prior works. Difference-1: When answering a query, VCA uniformly samples (a few) frames, whereas previous methods (e.g., VideoTree) require captioning each frame for subsequent image retrieval, which can be computationally expensive for long videos. Difference-2: previous methods rely on image similarity to locate key frames, potentially losing important video features. In contrast, VCA uses segment-level exploration to identify the most informative segments."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            540,
+            92,
+            880,
+            258
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            513,
+            382,
+            655,
+            397
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Long-context Multi-modal Agents",
+        "text_level": 1,
+        "bbox": [
+            511,
+            407,
+            810,
+            422
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Large multi-modal models (LMMs) [2, 6, 9, 35, 54] have demonstrated promising performance in addressing complex multimodal tasks [20, 45, 49, 81]. However, these LMMs typically lack self-directed exploration capabilities required for handling more tricky tasks [24, 74]. To bridge this gap, multimodal agents are designed to enhance LMMs with the ability to autonomously make decisions and execute tasks within specific environments [15, 73], such as interpreting visual cues and text prompts to make real-time decisions within websites [28, 76].",
+        "bbox": [
+            511,
+            429,
+            906,
+            580
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "A key feature of multimodal agents is their long-context capability [84], i.e., the ability to process and utilize information across extended periods to make informed decisions [26]. An example of this is vision-language model (VLM) agents for video understanding tasks. Despite their advancements, even sophisticated VLM agents struggle to extract crucial information and perform effective reasoning over lengthy inputs [62, 79]. Therefore, efficiently extracting essential information and making informed decisions remains a primary challenge. This work aims to address this challenge by enabling agents to self-exlore the environment and self-gather the most relevant information to efficiently solve video-understanding tasks.",
+        "bbox": [
+            511,
+            580,
+            908,
+            777
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Long Video Understanding",
+        "text_level": 1,
+        "bbox": [
+            511,
+            787,
+            759,
+            803
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Early deep learning models for video understanding primarily followed two main approaches [52]. One uses two-stream networks [18, 19, 42, 58] to integrate spatial and motion information, while the second focuses on 3D CNNs [8, 57, 72] for spatial-temporal feature extraction. Vision Transformers [14] then introduce self-attention mechanisms",
+        "bbox": [
+            511,
+            810,
+            908,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "20169",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/36be31cb5baa1db2c03e215ae3ccb61924a453519badc0eb3c88049bd76108e3.jpg",
+        "image_caption": [
+            "Figure 3. Pipeline of VCA Framework. First, the reward model predicts the relevance scores of each segment. Next, the memory buffer is updated by adding new frames and discarding irrelevant ones. Finally, emulating human reasoning, the agent actively selects segments to explore until the correct answer is found."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            197,
+            80,
+            803,
+            435
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "to handle temporal modeling more effectively [3, 7, 16]. Methods with self-supervised pre-training [50, 56, 86] exhibit transferability across different tasks and achieve more robust video understanding performance.",
+        "bbox": [
+            89,
+            486,
+            482,
+            547
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Multimodal LLM [22, 59, 83] shows strong capabilities in video understanding tasks, but efficiently processing information within a prolonged video environment is the key challenge. Recent approaches often use frame sampling [32, 46, 63, 75, 77] and feature extraction [27, 49, 51, 70] to reduce computational costs. Frame sampling selects a subset of frames, which risks missing essential context in lengthy videos, where transitions and subtle cues may be easily neglected. Additionally, processing entire video sequences as tokens presents significant memory and computational challenges [11, 43, 64, 80].",
+        "bbox": [
+            89,
+            565,
+            482,
+            733
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Recent work has explored auxiliary techniques and modules to improve efficiency, such as pre-extracted content like captions [44, 62], multimodal summaries [40, 65], and combinations of various auxiliary tools [17, 30]. However, these tools introduce additional computational and memory costs. Moreover, some approaches still require traversing the entire video content [17]. Notably, VideoTree [66] requires captioning every frame in the video, making it prohibitively expensive. In contrast, our VCA can solve the task using a single VLM, without relying on any supplementary tools.",
+        "bbox": [
+            89,
+            750,
+            482,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            513,
+            484,
+            604,
+            500
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Inspired by human reasoning, we introduce a self-exploration agent that adaptively navigates video environments. Unlike approaches that depend on dense frame sampling or captioning [62], which require VLMs to process numerous frames, our agent analyzes videos iteratively within a limited context window in a process we call self-exploration. This approach allows essential information to be extracted from long videos in a coarse-to-fine manner.",
+        "bbox": [
+            511,
+            513,
+            906,
+            635
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we propose a segment-based tree-search exploration structure. Given a video  $v$  and a user query  $q$ , the agent  $\\pi$  begins by randomly sampling  $N$  frames  $\\{f_v^i\\}_{i=1}^N$  to obtain an initial overview of the content in each video segment  $v_i \\in \\{v_i\\}_{i=1}^{N+1}$ . The agent then decides whether additional exploration is needed to answer the query. If further investigation is necessary, the agent selects a specific segment  $s^* \\in \\{v_i\\}_{i=1}^{N+1}$  for deeper exploration until sufficient information is gathered to respond to the query.",
+        "bbox": [
+            511,
+            638,
+            908,
+            777
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Nevertheless, a one-hour movie typically contains approximately 86,400 frames, posing a challenge for current VLM-based agents to efficiently identify the most relevant segment [17, 62, 66]. Moreover, even with a small  $N$ , conducting multiple rounds of exploration accumulates a substantial number of frames, leading to high computational costs. To address these issues, we introduce a reward model  $R$  to guide the agent's exploration and propose a memory",
+        "bbox": [
+            511,
+            780,
+            908,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "20170",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "buffer  $M$  to reduce storage and inference costs.",
+        "bbox": [
+            89,
+            90,
+            403,
+            104
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this section, we first explain how the reward model assists the agent throughout the exploration process (Sec. 3.1). Then we introduce our memory management strategy in Sec. 3.2. Finally, we explain how VCA adaptively explores the video using a segment-based tree structure (Sec. 3.3).",
+        "bbox": [
+            89,
+            107,
+            483,
+            183
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1. Reward Model",
+        "text_level": 1,
+        "bbox": [
+            89,
+            195,
+            241,
+            210
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "With the segment-based tree-search exploration, the agent is required to select the most informative video segment for exploration at each step. However, we observe that current VLM-based agents struggle to independently identify optimal exploration paths (see Sec. 4.3). To this end, inspired by the selective attention mechanism in human cognition [13], we incorporate a reward model  $R$  to guide the agent's selection process, i.e., the agent will gain a higher reward if it plans to take the action to explore more relevant segment.",
+        "bbox": [
+            89,
+            218,
+            483,
+            354
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Specifically, during the exploration of a current selected segment  $s^*$ , the reward model observes a set of frames  $\\{f_{s^*}^i\\}_{i=1}^N$  sampled from  $s^*$ . These frames  $\\{f_{s^*}^i\\}_{i=1}^N$  then form finer sub-segments  $\\{s_i^* = [l_{s^*}^i, l_{s^*}^{i+1}]\\}_{i=1}^{N+1}$ , where  $l_{s^*}^i$  and  $l_{s^*}^{i+1}$  represent the indices of the beginning and end frame of the sub-segment, respectively. The reward model aims to generate a reward score for each sub-segment  $\\{s_i^*\\}_{i=1}^{N+1}$  based on its relevance to the user query  $q$ . We utilize chain-of-thoughts [67] technique to let the reward model provide a verbal explanation of each segment and then generate a relevant score, i.e.,  $\\{t_{s_i^*}; r_{s_i^*}\\}_{i=1}^{N+1}$ . Additionally, to maintain consistency across different exploration steps, we incorporate the reward history  $H_r$  from previous steps alongside the new segments during evaluation. A simplified prompt template for the reward model is presented here, with detailed prompts in Appendix ??",
+        "bbox": [
+            89,
+            356,
+            483,
+            612
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Reward Model Prompt (Simplified)",
+        "text_level": 1,
+        "bbox": [
+            117,
+            623,
+            352,
+            638
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Given:",
+        "bbox": [
+            116,
+            651,
+            165,
+            664
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "$H_{r}$ , reward score history in previous steps.",
+        "bbox": [
+            116,
+            665,
+            398,
+            681
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "$\\left\\{f_{s^*}^i\\right\\}_{i = 1}^N$  newly sampled frames from current segment  $s^{*}$",
+        "bbox": [
+            116,
+            680,
+            455,
+            712
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Please first explain each sub-segment  $\\{s_i^*\\}_{i = 1}^{N + 1}$  and then assign a relevance score of each sub-segment to the user query  $q$ .",
+        "bbox": [
+            114,
+            710,
+            457,
+            760
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In practice, rather than training a separate specialized model, we utilize a same VLM-based agent to function as the reward model. While the reward model  $R$  and exploration agent  $\\pi$  serve distinct purposes, both can be implemented using the same model. In this work, we employ a single VLM-based agent for both roles across different settings, providing a neat and efficient agent framework that is guided by its self-generated intrinsic reward and without",
+        "bbox": [
+            89,
+            780,
+            483,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "relying on auxiliary models or tools.",
+        "bbox": [
+            511,
+            90,
+            754,
+            106
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Memory Management",
+        "text_level": 1,
+        "bbox": [
+            511,
+            116,
+            720,
+            132
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As exploration proceeds over multiple rounds, the accumulation of frames can lead to significant computational costs for long videos, even when  $N$  is small. However, as illustrated in Fig. 2, they tend to gradually disregard irrelevant frames, especially from earlier steps, retaining only the most crucial information in their working memory for later analysis. Inspired by this, we introduce a fixed-size memory buffer  $M$  that stores only the most relevant frames throughout exploration.",
+        "bbox": [
+            511,
+            138,
+            903,
+            273
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Specifically, when the agent  $\\pi$  explores a video segment  $s^*$ , the associated frames  $\\{f_{s^*}^i\\}_{i=1}^N$  are added to the memory buffer. If the total number of frames exceeds the buffer limit, irrelevant frames are removed based on relevance scores  $r_{s_i^*}$  defined in Sec. 3.1. We discard the frames with the lowest scores and only retain the most relevant ones. While these less relevant frames are removed, their visual information is extracted and encoded as text descriptions  $\\{t_{s_i^*}\\}_{i=1}^N$  within the exploration history  $H_r$ , allowing the reward model to utilize. This memory management strategy enables the agent to conduct reasoning within a fixed resource limit, regardless of the exploration trajectory's length, thereby preventing excessive memory usage.",
+        "bbox": [
+            511,
+            275,
+            905,
+            476
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Tree-Search Exploration",
+        "text_level": 1,
+        "bbox": [
+            511,
+            484,
+            738,
+            502
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Inspired by humans adaptively focusing on different video segments during exploration rather than targeting specific frames directly, our proposed agent VCA treats the video as an unexplored environment, where each segment represents a node on an exploration tree. Given the entire video as the root node, the agent's goal is to efficiently identify an optimal short path leading to a leaf node (segment) containing the most informative content needed to answer the query.",
+        "bbox": [
+            511,
+            507,
+            905,
+            628
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "At each step, the agent  $\\pi$  receives both textual and visual guidance: a set of candidate segments  $S$ , which includes a collection of segments with associated reward scores, and a memory buffer  $M$  containing multiple frames. If the agent has gathered sufficient information to answer the user's query  $q$ , it directly generates a response  $a$ . Otherwise, it continues exploring by selecting a specific segment  $s^* \\in S$  to expand the tree based on reward scores.",
+        "bbox": [
+            511,
+            628,
+            905,
+            748
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "However, due to the limited capacity of current foundational models, these reward scores are not always accurate (see further analysis in Sec. 5.3). Selecting segments based on reward scores in a purely greedy manner can potentially lead to a suboptimal solution. Instead, we use the reward scores as guidance for agent  $\\pi$ , allowing it to make the final segment selection independently. In this case, the agent is not forced to select the segment with the highest reward, balancing the exploitation and exploration. More importantly, the candidate set  $S$  includes both unexplored",
+        "bbox": [
+            511,
+            750,
+            906,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "20171",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "coarse segments from prior steps and fine-grained segments from the current step, enabling the agent to backtrack and select earlier segments if it detects that the current exploration path may be suboptimal. Below are the simplified prompts. The full prompt is provided in Appendix ??",
+        "bbox": [
+            89,
+            90,
+            483,
+            167
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Exploration Agent Prompt (Simplified)",
+        "text_level": 1,
+        "bbox": [
+            116,
+            181,
+            377,
+            199
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Given:",
+        "bbox": [
+            116,
+            210,
+            165,
+            223
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Candidate segments set:  $S$",
+        "bbox": [
+            116,
+            224,
+            295,
+            239
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Frames in Memory buffer:  $M$",
+        "bbox": [
+            116,
+            241,
+            315,
+            255
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Please choose one segment to explore based on the query  $q$ . Output answer if you have gathered sufficient information.",
+        "bbox": [
+            116,
+            256,
+            457,
+            299
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The overall framework of our proposed method is summarized in Algo. 1. Leveraging a tree-search exploration technique boosted with a reward model and a memory buffer, our self-exploration agent adaptively and efficiently explores the video like a human. Empirical results in Sec. 4.3 further validate the effectiveness of our tree-search exploration over naive frame selection methods.",
+        "bbox": [
+            89,
+            321,
+            483,
+            426
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            430,
+            223,
+            446
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            89,
+            455,
+            297,
+            473
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Benchmark. We mainly evaluate VCA on two long-term video question-answering benchmarks. EgoSchema [41] is an egocentric video dataset sourced from Ego4D [21], with a duration of 180 seconds. Here we use the validation set of 500 QA pairs. LVBench [61] covers about 1,500 samples for evaluation of reasoning abilities from high-quality videos over 30 minutes in length. We also conduct experiments on other relevant benchmark datasets in Appendix ??",
+        "bbox": [
+            89,
+            481,
+            483,
+            617
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Baselines. We compare with most relevant baselines to our work, including state-of-the-art open-source Video-LLMs [10, 23, 37, 47, 59, 82], the close-source model GPT-4o [25], and agent-based systems [44, 62, 66]. For video models, we report the results obtained with uniform sampling frames, while for the agent-based systems, we follow their official implementation with GPT-4o [25], consistent with VCA's setup to ensure a fair comparison.",
+        "bbox": [
+            89,
+            637,
+            483,
+            758
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. In this work, we use the same model for the exploration agent and the reward model. Unless otherwise stated, all experiments were performed using the August 2024 version of `gpt-4` as the base model, leveraging the OpenAI LLM API service<sup>1</sup> with a temperature of 0.5. We set memory buffer sizes set to 8 and 16 frames for EgoSchema and LVBench, respectively. Further implementation details are provided in Appendix ??",
+        "bbox": [
+            89,
+            779,
+            483,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [
+            "Algorithm 1 VCA: Video Curious Agent",
+            "Ensure: Answer  $a$"
+        ],
+        "code_body": "Require: Video environment  $v$ , question  $q$ , agent  $\\pi$ , reward model  $R$ , memory buffer  $M$ , candidate segments set  $S$ , reward score history  $H_{r}$ , sampling frame number  $N$ .",
+        "guess_lang": "txt",
+        "bbox": [
+            514,
+            109,
+            906,
+            155
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1: Initialize selected segment  $s^* \\gets v$",
+            "2: Initialize candidate segments set  $S \\gets \\emptyset$",
+            "3: Initialize memory buffer  $M\\gets \\emptyset$",
+            "4: Initialize reward score history  $H_{r}\\gets \\emptyset$",
+            "5: continue  $\\leftarrow$  true",
+            "6: while continue do",
+            "7: # Sampling frames & form segments",
+            "8:  $\\left\\{f_{s^*}^*\\right\\}_{i=1}^{N}, \\left\\{s_i^*\\right\\}_{i=1}^{N+1} \\gets \\text{UniformSample}(s^*)$",
+            "9: # Scoring each segment",
+            "10:  $\\left\\{t_{s_i^*};r_{s_i^*}\\right\\}_{i = 1}^{N + 1}\\gets R(\\{s_i^*\\}_{i = 1}^{N + 1};q,\\{f_{s_*}^i\\}_{i = 1}^N,H_r)$",
+            "11: # Update candidate segment set",
+            "12:  $S\\gets S\\cup \\{s_i^*;r_{s_i^*}\\}_{i = 1}^{N + 1}$",
+            "13: # Update history with thoughts",
+            "14:  $H_{r}\\gets H_{r}\\cup \\left\\{t_{s_{i}^{*}};r_{s_{i}^{*}}\\right\\}_{i = 1}^{N + 1}$",
+            "15: # Update memory with frames",
+            "16:  $M\\gets \\mathrm{UpdateMemory}(M,\\{f_{s^{*}}^{i}\\}_{i = 1}^{N};S)$",
+            "17: # Output answer or select segment",
+            "18:  $a,s\\gets \\pi (q,S;M)$",
+            "19: if  $a = \\emptyset$  then",
+            "20: # Continue exploration",
+            "21:  $s^* \\gets s$",
+            "22: else",
+            "23: continue  $\\leftarrow$  false",
+            "24: end if",
+            "25: end while",
+            "26: return  $a$"
+        ],
+        "bbox": [
+            517,
+            170,
+            885,
+            570
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Experimental Results",
+        "text_level": 1,
+        "bbox": [
+            511,
+            599,
+            715,
+            614
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The experimental results on LVBench and EgoSchema are presented in Tab. 2 and Tab. 1, respectively. By adaptively focusing on critical segments rather than scanning the entire video, our agent not only achieves stronger performance but also incurs less computational cost. With less than  $30\\%$  of the frames, VCA achieves substantial improvements over directly feeding uniformly sampled frames into GPT-4o. Specifically, we observe performance gains of  $3.2\\%$  on the EgoSchema subset and  $6.6\\%$  on LVBench. Significant increases in the Event Understanding and Temporal Grounding domains further underscore our effectiveness in identifying crucial segments within long videos.",
+        "bbox": [
+            511,
+            621,
+            906,
+            801
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, we compare our framework with recently proposed SOTA video agent systems, VideoAgent [62] and VideoTree [66]. For a fair comparison, we re-implement both frameworks using GPT-4o for all reasoning components, consistent with our setup. However, captioning all",
+        "bbox": [
+            511,
+            801,
+            906,
+            878
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "1https://platform.openai.com/docs/models",
+        "bbox": [
+            531,
+            886,
+            846,
+            898
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "20172",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/cdc65838a4e11839a1a1465235c041052cb787d85754c744b501c8f8c7688785.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Frames</td><td>ER</td><td>EU</td><td>KIR</td><td>TG</td><td>Rea</td><td>Sum</td><td>Avg.</td></tr><tr><td>TimeChat* [47]</td><td>&gt;96</td><td>21.9</td><td>21.7</td><td>25.9</td><td>22.7</td><td>25.0</td><td>24.1</td><td>22.3</td></tr><tr><td>LLaVA-NeXT* [83]</td><td>32</td><td>30.1</td><td>31.2</td><td>34.1</td><td>31.4</td><td>35.0</td><td>27.6</td><td>32.2</td></tr><tr><td>InternVL2-40B* [10]</td><td>16</td><td>37.4</td><td>39.7</td><td>43.4</td><td>31.4</td><td>42.5</td><td>41.4</td><td>39.8</td></tr><tr><td>GLM4V-Plus* [23]</td><td>20</td><td>39.9</td><td>35.8</td><td>34.8</td><td>37.7</td><td>40.0</td><td>32.8</td><td>38.3</td></tr><tr><td>GPT-4o* [25]</td><td>64</td><td>35.9</td><td>30.8</td><td>35.5</td><td>28.3</td><td>33.5</td><td>34.5</td><td>34.7</td></tr><tr><td>VideoAgent† [62]</td><td>Avg. 25.5</td><td>28.0</td><td>30.3</td><td>28.0</td><td>29.3</td><td>28.0</td><td>36.4</td><td>29.3</td></tr><tr><td>VideoTree† [66]</td><td>Avg. 103.2</td><td>30.3</td><td>25.1</td><td>26.5</td><td>27.7</td><td>31.9</td><td>25.5</td><td>28.8</td></tr><tr><td>VCA (ours)</td><td>Avg. 20.0</td><td>43.7</td><td>40.7</td><td>37.8</td><td>38.0</td><td>46.2</td><td>27.3</td><td>41.3</td></tr></table>",
+        "bbox": [
+            207,
+            88,
+            787,
+            253
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/58c7945053aceb0fd40e6d71d931795c4e8194cd4f36892dd1f0f3f85477e13b.jpg",
+        "table_caption": [
+            "Table 1. Experimental Results on LVBench. Results of methods marked with * are sourced from LVBench [61]. Methods marked with † indicate implementations where GPT-4o serves as the main component. Our agent's memory buffer size is set to 16 frames. We report the average number of observed frames for agent-based methods, with the best performance highlighted in bold."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Frames</td><td>Subset</td></tr><tr><td rowspan=\"2\">GPT-4o [25]</td><td>8</td><td>69.0</td></tr><tr><td>32</td><td>70.4</td></tr><tr><td rowspan=\"2\">\\(VideoAgent^\\dagger\\) [62]</td><td>Avg. 8.1</td><td>63.2</td></tr><tr><td>Avg. 10.7</td><td>60.8</td></tr><tr><td>\\(VideoTree^\\dagger\\) [66]</td><td>Avg. 69.5</td><td>67.0</td></tr><tr><td>\\(LVNet^\\dagger\\) [44]</td><td>12</td><td>68.2</td></tr><tr><td>VCA (ours)</td><td>Avg. 7.2</td><td>73.6</td></tr></table>",
+        "bbox": [
+            151,
+            323,
+            421,
+            474
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "frames with GPT-4o is prohibitively expensive. Hence, following the baseline implementations, we use open-source VLMs for captioning. A comparison using the same model for all components is provided in Sec. 5.1. As shown in Tab. 1, VCA outperforms all baselines over  $5\\%$  on EgoSchema, while requiring processing fewer frames. These results suggest the limitations of dense captioning for video preprocessing tasks, which can incur more computational costs while introducing more redundant information. In contrast, our proposed memory management strategy effectively improves efficiency by only retaining the most crucial information for reasoning. Similar insights can be drawn from results on LVBench in Tab. 2. Additional results are provided in Appendix ??",
+        "bbox": [
+            88,
+            578,
+            482,
+            790
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Moreover, we observe that VCA explores approximately three times more frames on LVBench than on EgoSchema. Recall that EgoSchema videos have an average duration of 180 seconds, while videos in LVBench extend to at least 30 minutes. This phenomenon suggests that our proposed tree-search exploration technique actively explores the video to locate the most relevant information and adaptively in-",
+        "bbox": [
+            89,
+            795,
+            482,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/f6591466ea61c795db00ac6cff0bc93214d577866bef4bd46d75d4ca9daf4aaa.jpg",
+        "table_caption": [
+            "Table 2. Experimental Results on EgoSchema. Methods marked with  $\\dagger$  indicate implementations where GPT-4o is the main component. The results of LVNet are from [44]. Our agent's memory buffer size is set to 8 frames. We report the average number of observed frames, with the best performance highlighted in bold."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">EgoSchema</td><td colspan=\"2\">LVBench</td></tr><tr><td>Frames</td><td>Acc.</td><td>Frames</td><td>Acc.</td></tr><tr><td>VCA (ours)</td><td>7.2</td><td>73.6</td><td>20.0</td><td>41.3</td></tr><tr><td>- w/o Reward</td><td>7.5</td><td>72.6</td><td>22.2</td><td>39.5</td></tr><tr><td>- w/o Tree Search</td><td>10.2</td><td>69.8</td><td>39.9</td><td>36.2</td></tr></table>",
+        "bbox": [
+            540,
+            323,
+            879,
+            412
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/12eb3b064a7106c503cac75584b9acfb634f0a1eb1846cbde7e0a2a197d01119.jpg",
+        "table_caption": [
+            "Table 3. Ablation Study on EgoSchema and LVBench. Both the tree-search exploration and reward model contribute significantly to improvements in performance and efficiency. Detailed ablation study results on LVBench are included in Appendix ??"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Buffer Size</td><td>8</td><td>16</td><td>32</td></tr><tr><td>Acc.</td><td>38.9</td><td>41.3</td><td>43.2</td></tr></table>",
+        "bbox": [
+            602,
+            486,
+            818,
+            530
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Ablation Study of Buffer Size on LVBench. Increasing the buffer size allows the model to observe more frames during reasoning, leading to improved performance.",
+        "bbox": [
+            511,
+            541,
+            906,
+            583
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "creases the exploration budget based on task complexity. For example, VCA exhibits both exploitation and exploration: when the agent determines that the current exploration path is suboptimal, it backtracks to previous segments to explore a new path (see Sec. 5.2). These empirical observations further validate the robustness and scalability of our approach, especially in diverse long-video contexts.",
+        "bbox": [
+            511,
+            597,
+            905,
+            704
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            511,
+            712,
+            666,
+            729
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we conduct an ablation study to evaluate the components of our agent. Results across both benchmarks are presented in Tab. 3. As illustrated in the table, performance consistently declines when the reward model is removed, highlighting that the agent's inherent reasoning ability alone is insufficient for accurately identifying the most informative segments. This finding emphasizes the effectiveness of incorporating an intrinsic reward to guide the agent's segment selection process. A detailed analysis of our reward model's performance is provided in Sec. 5.3.",
+        "bbox": [
+            511,
+            734,
+            906,
+            883
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Additionally, we observe a significant performance drop",
+        "bbox": [
+            532,
+            886,
+            905,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "20173",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "Why does the mice run away?",
+            "(A) Because the black cat sheriff calls the police",
+            "(C) Because their boss is injured"
+        ],
+        "bbox": [
+            222,
+            95,
+            455,
+            127
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(B) Because they meet more powerful mice",
+            "(D) Because the owner of the house is back"
+        ],
+        "bbox": [
+            460,
+            107,
+            665,
+            126
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/494c2e39d33787b489c52c42d9cf3f62f97d802bf3cb5b76c77ffd4cb93b8d38.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            222,
+            132,
+            779,
+            183
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Reward Model",
+        "text_level": 1,
+        "bbox": [
+            223,
+            188,
+            310,
+            199
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "Segment1: Thought: The segment shows a black cat with a gun, possibly indicating a threat. This could suggest that the mice might run away due to the presence of the cat, which aligns with option (A). Score: 0.7",
+            "Segment4: Thought: The segment features the black cat with a communication device, possibly calling for help or backup. This aligns with option (A), suggesting the mice might run because the cat is calling the police. Score: 0.8"
+        ],
+        "bbox": [
+            223,
+            200,
+            602,
+            266
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/cab27bdb7ff0b0177e4565f738e98f8d8b8a7fec79b098210f2663bdfcdbd937.jpg",
+        "image_caption": [
+            "Agent Choice: (segment: 4)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            609,
+            210,
+            696,
+            250
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/db534f7962bd217c6a8212442a281185644644533499078be1979c927fe8315b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            210,
+            777,
+            250
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/9709989d9606dbfa36835e193981ef9fd1b3c4452fb7a2442047d274c61b2bbb.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            223,
+            272,
+            779,
+            323
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Reward Model",
+        "text_level": 1,
+        "bbox": [
+            223,
+            327,
+            308,
+            338
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Segment4.1: Thought: The segment shows a close-up of a document with red text and exclamation marks, indicating urgency or importance. This could imply a critical message or alert being communicated. Score: 0.7",
+        "bbox": [
+            223,
+            339,
+            601,
+            369
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "…",
+        "bbox": [
+            223,
+            369,
+            233,
+            378
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/0be7582fe7139255d5979ad1a0a4b2f8416fde86fabbd6434c0c69aee2357f20.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            223,
+            387,
+            777,
+            438
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Reward Model",
+        "text_level": 1,
+        "bbox": [
+            223,
+            443,
+            310,
+            453
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Segment1.2: Thought: The segment features mice characters in mid-air, possibly fleeing or being thrown. This suggests a direct reaction to a threat, supporting the idea that they are running away from the cat. Score: 0.8",
+        "bbox": [
+            223,
+            454,
+            584,
+            496
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e82b2c43037a928c41960aef4203fd12a1f1ce59e083cac00c6ba6a3ec87c9ad.jpg",
+        "image_caption": [
+            "Figure 4. Exploration Trajectory Example. The agent generally tends to select video segments with higher reward scores for exploration, yet it can adaptively redirect to earlier segments, effectively balancing exploitation and exploration."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            223,
+            502,
+            777,
+            539
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/5386d87e480fee48366b58408db92687c9c944cdedb4c56e93e21002e2bd9557.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            573,
+            542,
+            602,
+            558
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Agent: Answer: A",
+        "bbox": [
+            640,
+            544,
+            746,
+            556
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "when the tree-search exploration technique is ablated, i.e., the agent directly selects frame indices without the segment-based tree structure. Interestingly, this modification also results in a substantial increase in the average number of frames acquired, especially on LVBench. This indicates that the agent struggles to locate the most crucial information and consequently produces a longer but redundant trajectory. These ablation study results validate the effectiveness of the tree-search exploration framework assisted by the reward model. Additionally, this underscores VCA's potential to achieve greater efficiency and improved performance with more advanced VLMs and reward models.",
+        "bbox": [
+            89,
+            611,
+            482,
+            791
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Moreover, as shown in Tab. 4, increasing the memory buffer size yields consistently improved performance for VCA, demonstrating that our framework benefits from a broader reasoning window while maintaining effectiveness with fewer frames. Due to limited computation resources, we only tested up to 32 frames; however, the upward trend indicates the potential for gains with additional frames.",
+        "bbox": [
+            89,
+            795,
+            482,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5. Analysis",
+        "text_level": 1,
+        "bbox": [
+            513,
+            609,
+            609,
+            626
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1. Can VCA generalize to Open-source Models?",
+        "text_level": 1,
+        "bbox": [
+            511,
+            635,
+            897,
+            651
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "While VCA demonstrates strong performance using GPT4o as the base model, we also analyze whether these positive results can be extrapolated to open-source models. We implement the VCA framework with Qwen2-VL-72B-Instruct-GPTQ-Int4 [59] as both the exploration agent and reward model, as well as the captioner for baselines. The results in Tab. 5 reveal that, compared to directly feeding uniformly sampled frames into Qwen2-VL, our framework achieves higher accuracy on the EgoSchema benchmark despite using only about  $60\\%$  of the frames. In contrast, VideoAgent performs significantly worse. Furthermore, our approach outperforms VideoTree with far fewer frames. Notably, VideoTree employs clustering to build a caption similarity-based tree structure for frame selection, requiring caption generation for each frame. This process is computationally expensive, demanding over 60 GPU hours",
+        "bbox": [
+            511,
+            657,
+            906,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "20174",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/de5d40b1cd7f8593e0c298b92c7fc57eb36751dc7abde60dd39b7f4435f5e892.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Frames</td><td>Subset</td></tr><tr><td rowspan=\"2\">Qwen2-VL</td><td>8</td><td>70.2</td></tr><tr><td>32</td><td>74.2</td></tr><tr><td>VideoAgent</td><td>Avg. 8.6</td><td>65.2</td></tr><tr><td>VideoTree</td><td>Avg. 85.4</td><td>71.0</td></tr><tr><td>VCA (ours)</td><td>Avg. 16.9</td><td>75.2</td></tr><tr><td>- w/ Image Reward</td><td>Avg. 20.9</td><td>74.0</td></tr></table>",
+        "bbox": [
+            148,
+            88,
+            426,
+            223
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5. Performance of Using Open-source Model on EgoSchema. Our approach achieves higher accuracy with fewer observed frames, showing consistency improvements over baseline methods. Memory buffer size is set to 16 frames.",
+        "bbox": [
+            89,
+            233,
+            482,
+            291
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "on a single H100, as shown in Fig. 1. These results further underscore the robustness of VCA.",
+        "bbox": [
+            89,
+            321,
+            482,
+            352
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2. What's the Exploration Behavior of VCA?",
+        "text_level": 1,
+        "bbox": [
+            89,
+            357,
+            452,
+            373
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Here we present a real exploration trajectory from LVBench to illustrate our agent's decision-making process. As shown in Fig. 4, given a user query, the reward model initially evaluates each segment's relevance based on uniformly sampled frames, and the agent selects the fourth segment, which has the highest relevance score. However, upon further investigation of segment 4, both the reward model and the agent find it relatively irrelevant to the query. Rather than continuing with it, the agent redirects its exploration to segment 1 in the last round. After selecting segment 1, the agent chooses to investigate the third segment, despite the second segment having the highest reward score. This behavior reflects VCA's ability to adaptively adjust its exploration trajectory based on visual context rather than purely following the reward score, resulting in a more robust system.",
+        "bbox": [
+            89,
+            382,
+            482,
+            608
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We also examine the limitations of VCA by analyzing its failure cases, which fall into the following scenarios: The most common failure occurs when the agent overlooks subtle clues, resulting in an inaccurate response to the user query even after prolonged exploration, particularly for queries requiring specific visual details. Other failure cases include instances where the agent is misled by inaccurate reward scores or fails to provide the correct answer despite observing key frames. Detailed cases and further discussion are provided in the Appendix ??",
+        "bbox": [
+            89,
+            611,
+            482,
+            762
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In addition, unlike previous methods, VCA employs a segment-level exploration strategy instead of direct image similarity to select key frames. To further investigate this difference, we evaluate the performance of selecting the segment whose start and end images have the highest visual relevance. As shown in Tab. 5, relying solely on image relevance instead of inferring segment semantics results in a performance drop on EgoSchema, confirming the effectiveness of our segment-level exploration.",
+        "bbox": [
+            89,
+            763,
+            482,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/fca8834cee31104ddaa6b77c3ab66dbfffdf08de055321278fd4a45f2a317be3.jpg",
+        "image_caption": [
+            "Figure 5. Segment Selection Accuracy on LVBench. The y-axis shows segment distance from ground truth; x-axis shows count ratio. Red curves are fitted normal distributions. The reward-guided agent selects segments closer to ground truth than the reward model."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            540,
+            89,
+            879,
+            321
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3. How Reliable is the Reward Model?",
+        "text_level": 1,
+        "bbox": [
+            511,
+            421,
+            823,
+            436
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To further assess the capability of our reward model, we visualize the matching accuracy between segments with the highest reward scores and ground truth time references from LVBench. The results are shown at the bottom of Fig. 5. We observe that segment distances follow an approximately normal distribution with a mean of 1.28, indicating that the segment with the highest reward score often closely aligns with the ground truth segment. This validates the reward model's effectiveness in providing meaningful guidance.",
+        "bbox": [
+            511,
+            443,
+            905,
+            580
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Additionally, as discussed in Sec 3.3, we encourage the agent to make independent decisions when selecting segments rather than blindly following reward scores in a purely greedy manner. The benefit of this design choice is demonstrated at the top of Fig. 5, where we present the matching accuracy of the agent's choices. Compared to the greedy approach, the agent's selections result in a sharper distribution of segment distances, with an average distance gap of  $-0.60$ , indicating closer alignment with the ground truth. In summary, our reward model assigns reliable relevance scores to segments, while the agent further refines these choices to achieve even more accurate selections.",
+        "bbox": [
+            511,
+            580,
+            905,
+            762
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            511,
+            777,
+            633,
+            792
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this work, we propose a curiosity-driven video agent framework for long video understanding tasks with self-exploration capability. VCA achieves state-of-the-art performance and efficiency without relying on auxiliary tools. Appendix ?? shows that improved reward guidance could further boost results, highlighting the potential of synthetic data and specialized reward models for future work.",
+        "bbox": [
+            511,
+            801,
+            905,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "20175",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            91,
+            118,
+            187,
+            133
+        ],
+        "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] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Devendra Chaplot, Jessica Chudnovsky, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amélie Héliou, Paul Jacob, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 2",
+            "[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6816-6826, 2021. 3",
+            "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1",
+            "[5] Ivana Balazevic, Yuge Shi, Pinelopi Papalampidi, Rahma Chaabouni, Skanda Koppula, and Olivier J Henaff. Memory consolidation enables long-context video understanding. In *Forty-first International Conference on Machine Learning*, 2024. 2",
+            "[6] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Fuyu-8b: A multimodal architecture for ai agents, 2023. 2",
+            "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding?, 2021. 3",
+            "[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset, 2018. 1, 2",
+            "[9] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 2",
+            "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 5, 6",
+            "[11] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 3",
+            "[12] Jiwan Chung and Youngjae Yu. Long story short: a summarize-then-search method for long video question answering. arXiv preprint arXiv:2311.01233, 2023. 2",
+            "[13] Robert Desimone, John Duncan, et al. Neural mechanisms of selective visual attention. Annual review of neuroscience, 18(1):193-222, 1995. 4"
+        ],
+        "bbox": [
+            93,
+            143,
+            480,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[14] 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, 2021. 2",
+            "[15] Zane Durante, Qiuyuan Huang, Naoki Wake, Ran Gong, Jae Sung Park, Bidipta Sarkar, Rohan Taori, Yusuke Noda, Demetri Terzopoulos, Yejin Choi, Katsushi Ikeuchi, Hoi Vo, Li Fei-Fei, and Jianfeng Gao. Agent ai: Surveying the horizons of multimodal interaction, 2024. 2",
+            "[16] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers, 2021. 3",
+            "[17] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2025. 1, 2, 3",
+            "[18] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition, 2016. 1, 2",
+            "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 2",
+            "[20] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2",
+            "[21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 5",
+            "[22] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, Lei Zhao, Zhuoyi Yang, Xiaotao Gu, Xiaohan Zhang, Guanyu Feng, Da Yin, Zihan Wang, Ji Qi, Xixuan Song, Peng Zhang, Debing Liu, Bin Xu, Juanzi Li, Yuxiao Dong, and Jie Tang. CogvIm2: Visual language models for image and video understanding, 2024. 3",
+            "[23] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 5, 6",
+            "[24] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 2",
+            "[25] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "20176",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 5, 6",
+            "[26] Lawrence Jang, Yinheng Li, Charles Ding, Justin Lin, Paul Pu Liang, Dan Zhao, Rogerio Bonatti, and Kazuhito Koishida. Videowebarena: Evaluating long context multimodal agents with video understanding web tasks, 2024. 2",
+            "[27] Peng Jin, Ryuichi Takanobu, Caiwan Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding, 2023. 3",
+            "[28] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks, 2024. 2",
+            "[29] Bruno Korbar, Yongqin Xian, Alessio Tonioni, Andrew Zisserman, and Federico Tombari. Text-conditioned resampler for long form video understanding. In European Conference on Computer Vision, pages 271-288. Springer, 2025. 1",
+            "[30] Somnath Kumar, Yash Gadhia, Tanuja Ganu, and Akshay Nambi. Mmctagent: Multi-modal critical thinking agent framework for complex visual reasoning, 2024. 3",
+            "[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1",
+            "[32] Yunxin Li, Xinyu Chen, Baotain Hu, and Min Zhang. Llms meet long video: Advancing long video comprehension with an interactive visual adapter in llms, 2024. 1, 3",
+            "[33] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2025. 1",
+            "[34] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1",
+            "[35] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 2",
+            "[36] Ming C Lin and Shan Yang. Vila: Efficient video-language alignment for video question answering. 1",
+            "[37] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5",
+            "[38] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv preprint arXiv:2402.08268, 2024. 1",
+            "[39] Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. Drvideo: Document retrieval based long video understanding. arXiv preprint arXiv:2406.12846, 2024. 2",
+            "[40] Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. Drvideo: Document retrieval based long video understanding, 2024. 3"
+        ],
+        "bbox": [
+            91,
+            90,
+            482,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[41] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. arXiv preprint arXiv:2308.09126, 2023. 5",
+            "[42] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification, 2015. 2",
+            "[43] Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14386-14397, 2024. 3",
+            "[44] Jongwoo Park, Kanchana Ranasinghe, Kumara Kahatapitiya, Wonjeong Ryoo, Donghyun Kim, and Michael S Ryoo. Too many frames, not all useful: Efficient strategies for long-form video qa. arXiv preprint arXiv:2406.09396, 2024. 1, 3, 5, 6",
+            "[45] Ruchit Rawal, Khalid Saifullah, Ronen Basri, David Jacobs, Gowthami Somepalli, and Tom Goldstein. Cinepile: A long video question answering dataset and benchmark. In Synthetic Data for Computer Vision Workshop @ CVPR 2024. 2",
+            "[46] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. arXiv preprint arXiv:2312.02051, 2023. 3",
+            "[47] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 5, 6",
+            "[48] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5026-5035, 2022. 1",
+            "[49] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023. 1, 2, 3",
+            "[50] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning, 2019. 3",
+            "[51] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. Advances in neural information processing systems, 35:38032-38045, 2022. 3",
+            "[52] Yunlong Tang, Jing Bi, Sitting Xu, Luchuan Song, Susan Liang, Teng Wang, Daoan Zhang, Jie An, Jingyang Lin, Rongyi Zhu, Ali Vosoughi, Chao Huang, Zeliang Zhang, Pinxin Liu, Mingqian Feng, Feng Zheng, Jianguo Zhang,"
+        ],
+        "bbox": [
+            514,
+            92,
+            903,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "20177",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ping Luo, Jiebo Luo, and Chenliang Xu. Video understanding with large language models: A survey, 2024. 2",
+            "[53] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4631–4640, 2016. 1",
+            "[54] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2",
+            "[55] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1",
+            "[56] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training, 2022. 3",
+            "[57] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks, 2015. 1, 2",
+            "[58] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition, 2016. 2",
+            "[59] 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. arXiv preprint arXiv:2409.12191, 2024. 3, 5, 7",
+            "[60] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1",
+            "[61] Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Shiyu Huang, Bin Xu, Yuxiao Dong, Ming Ding, and Jie Tang. Lvbench: An extreme long video understanding benchmark, 2024. 1, 5, 6",
+            "[62] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. European Conference on Computer Vision (ECCV), 2024. 1, 2, 3, 5, 6",
+            "[63] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 1, 3",
+            "[64] Yuxuan Wang, Cihang Xie, Yang Liu, and Zilong Zheng. Videollamb: Long-context video understanding with recurrent memory bridges. arXiv preprint arXiv:2409.01071, 2024. 1, 3",
+            "[65] Ying WANG, Yanlai YANG, and Mengye REN. Lifelong-memory: Leveraging llms for answering queries in long"
+        ],
+        "bbox": [
+            91,
+            90,
+            482,
+            901
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "form egocentric videos. arXiv preprint arXiv:2312.05269, 2024. 1, 3",
+            "[66] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos, 2024. 1, 2, 3, 5, 6",
+            "[67] 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. Advances in neural information processing systems, 35:24824-24837, 2022. 4",
+            "[68] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. In European Conference on Computer Vision, pages 453-470. Springer, 2025. 1",
+            "[69] Chao-Yuan Wu and Philipp Krahenbuhl. Towards long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1884-1894, 2021. 1",
+            "[70] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 3",
+            "[71] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. CoRR, 2024. 1",
+            "[72] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification, 2018. 2",
+            "[73] Tianbao Xie, Fan Zhou, Zhoujun Cheng, Peng Shi, Luoxuan Weng, Yitao Liu, Toh Jing Hua, Junning Zhao, Qian Liu, Che Liu, et al. Openagents: An open platform for language agents in the wild. In ICLR 2024 Workshop on Large Language Model (LLM) Agents. 2",
+            "[74] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. arXiv preprint arXiv:2404.07972, 2024. 2",
+            "[75] Jiaqi Xu, Cuiling Lan, Wenxuan Xie, Xuejin Chen, and Yan Lu. Retrieval-based video language model for efficient long video question answering, 2023. 1, 3",
+            "[76] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents, 2023. 2",
+            "[77] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. NeurIPS, 2023. 3",
+            "[78] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. Advances in Neural Information Processing Systems, 36, 2024. 1",
+            "[79] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple lmm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 2"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            900
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "20178",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[80] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, 2023. 3",
+            "[81] Hongjie Zhang, Yi Liu, Lu Dong, Yifei Huang, Zhen-Hua Ling, Yali Wang, Limin Wang, and Yu Qiao. Movqa: A benchmark of versatile question-answering for long-form movie understanding. arXiv preprint arXiv:2312.04817, 2023. 2",
+            "[82] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 5",
+            "[83] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 3, 6",
+            "[84] Jun Zhao, Can Zu, Hao Xu, Yi Lu, Wei He, Yiwen Ding, Tao Gui, Qi Zhang, and Xuanjing Huang. Longagent: Scaling language models to 128k context through multi-agent collaboration. arXiv preprint arXiv:2402.11550, 2024. 2",
+            "[85] Long Zhao, Nitesh Bharadwaj Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. In *Forty-first International Conference on Machine Learning*. 1",
+            "[86] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations, 2020. 3"
+        ],
+        "bbox": [
+            91,
+            90,
+            480,
+            529
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "20179",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 11
+    }
+]

2025/VCA_ Video Curious Agent for Long Video Understanding/a73c253b-5afa-42a6-b79a-31559f118997_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b2d12f8092f3508d39a82b284d887acba4d240d7e9bea4e4a37aecc06ab16f8
+size 5127347

2025/VCA_ Video Curious Agent for Long Video Understanding/full.md

@@ -0,0 +1,383 @@
+# VCA: Video Curious Agent for Long Video Understanding
+
+Zeyuan Yang†  
+University of Massachusetts, Amherst  
+yangzeyuan@umass.edu
+
+Xueyang Yu  
+University of Massachusetts, Amherst  
+xueyangyu@umass.edu
+
+Delin Chen†  
+University of Massachusetts, Amherst delinchen@umass.edu
+
+Maohao Shen  
+Massachusetts Institute of Technology  
+maohao@mit.edu
+
+Chuang Gan  
+University of Massachusetts, Amherst  
+ganchuang1990@gmail.com
+
+# Abstract
+
+Long video understanding poses unique challenges due to their temporal complexity and low information density. Recent works address this task by sampling numerous frames or incorporating auxiliary tools using LLMs, both of which result in high computational costs. In this work, we introduce a curiosity-driven video agent with self-exploration capability, dubbed as "VCA". Built upon VLMs, VCA autonomously navigates video segments and efficiently builds a comprehensive understanding of complex video sequences. Instead of directly sampling frames, VCA employs a tree-search structure to explore video segments and collect frames. Rather than relying on external feedback or reward, VCA leverages VLM's self-generated intrinsic reward to guide its exploration, enabling it to capture the most crucial information for reasoning. Experimental results on multiple long video benchmarks demonstrate our approach's superior effectiveness and efficiency.
+
+# 1. Introduction
+
+There is a growing demand for developing long-form video understanding techniques [48, 69] capable of extracting valuable information from extended visual narratives, such as surveillance footage, documentaries, and instructional videos. Analyzing these videos [53, 71], which can range from minutes to hours, requires models that can process multimodal data and perform reasoning over extremely long sequences [4, 31, 60].
+
+Previous works commonly utilize two-stream networks to capture spatial and temporal information [18, 19] or in
+
+![](images/b3a34f2280701b001980d124b23ff96bfa7878b2f2423e84306fc895c68a5841.jpg)  
+Figure 1. Accuracy v.s. Computation Cost and Number of Observed Frames. Compared to baselines, our agent demonstrates superior performance with higher efficiency (observing fewer frames and incurring comparable or lower computational cost) on EgoSchema, using both GPT-4o and Qwen2-VL-72B.
+
+![](images/91ceaf3d095c177613ad02b2b753cad522a8335adc05cf71027f60062516008a.jpg)
+
+troduce 3D operations [8, 57]. Recent approach [32, 75] leverages the long-sequence reasoning capabilities of Large Language Models (LLMs) [1, 64] for long-form video question answering, often by converting videos into densely and uniformly sampled frames [33, 38, 55]. However, this approach is usually inefficient due to the inherent imbalance and varying information density across video segments [61, 78]. Some segments are rich in critical details necessary for answering questions, while others contain only trivial information [66]. Uniform sampling [49, 60], which might focus on redundant frames that increase the computational burden and may overlook frames with essential information, often fails to efficiently utilize the video's dynamic structure for question answering [17, 62].
+
+To address this limitation, recent approaches [29, 34, 36, 68] employ complex architectures [63, 85] or specialized preprocessing techniques [44, 65] to improve the efficiency
+
+of long video content processing. These methods [62, 66] often convert videos into text by captioning densely sampled frames or incorporating memory modules [5, 17, 39] to store essential information. However, these approaches still incur significant computational costs. For example, captioning an hour-long video with Qwen2-VL-72B at just 1 fps requires over 3 hours on a single H100. Additionally, these methods [12] identify key frames based on image similarity rather than a holistic understanding of the video content, which can result in missing relevant video features. This necessitates the need for an efficient and scalable solution for long-form video understanding.
+
+In this work, we introduce VCA, a curiosity-driven agent inspired by how humans tackle the long video understanding task. Given a question, human is usually intrinsically motivated to identify relevant information in the video without external knowledge. In practise, human tends to build a global understanding of the video, then curiously zoom in to explore specific sections in greater detail. Motivated by this process, we propose a novel video agent framework, dubbed as "VCA", to emulate such behavior. VCA treats the video as an environment to explore, actively navigating its contents and identifying segments that merit deeper analysis. VCA incorporates three essential techniques: (1) Tree-search Exploration: Rather than selecting specific frames [62], VCA uses a tree-search structure to adaptively explore the video segments in a structured way. (2) Reward Model: VCA leverages an intrinsic reward scoring mechanism to guide exploration toward the segments most relevant to the query. (3) Memory Management: VCA employs a fixed memory buffer to store key information and discard irrelevant frames, effectively reducing computational costs and avoiding memory overflow. The key difference between VCA and prior works is outlined in Fig. 2. Extensive experiments and superior performance across multiple benchmarks demonstrate that our proposed VCA framework significantly improves an agent's effectiveness and efficiency in reasoning over extended video content. More concretely, our contributions are threefold,
+
+- VCA is a long video understanding agent that can self-exlore the most informative video segments based on its intrinsic reward, without relying on any auxiliary tools.  
+- VCA is easy to implement, training-free, and is flexible to integrate with any vision-language model.  
+- As highlighted in Fig. 1, VCA outperforms existing baselines while observing fewer frames and incurring comparable or lower computational cost, demonstrating both effectiveness and efficiency.
+
+![](images/86d97b7ae7cef89e8e72f2b8f1a26808351cdd81b23ff0ab9e2edc6c90c89ae7.jpg)  
+Figure 2. Comparison of VCA with prior works. Difference-1: When answering a query, VCA uniformly samples (a few) frames, whereas previous methods (e.g., VideoTree) require captioning each frame for subsequent image retrieval, which can be computationally expensive for long videos. Difference-2: previous methods rely on image similarity to locate key frames, potentially losing important video features. In contrast, VCA uses segment-level exploration to identify the most informative segments.
+
+# 2. Related Work
+
+# 2.1. Long-context Multi-modal Agents
+
+Large multi-modal models (LMMs) [2, 6, 9, 35, 54] have demonstrated promising performance in addressing complex multimodal tasks [20, 45, 49, 81]. However, these LMMs typically lack self-directed exploration capabilities required for handling more tricky tasks [24, 74]. To bridge this gap, multimodal agents are designed to enhance LMMs with the ability to autonomously make decisions and execute tasks within specific environments [15, 73], such as interpreting visual cues and text prompts to make real-time decisions within websites [28, 76].
+
+A key feature of multimodal agents is their long-context capability [84], i.e., the ability to process and utilize information across extended periods to make informed decisions [26]. An example of this is vision-language model (VLM) agents for video understanding tasks. Despite their advancements, even sophisticated VLM agents struggle to extract crucial information and perform effective reasoning over lengthy inputs [62, 79]. Therefore, efficiently extracting essential information and making informed decisions remains a primary challenge. This work aims to address this challenge by enabling agents to self-exlore the environment and self-gather the most relevant information to efficiently solve video-understanding tasks.
+
+# 2.2. Long Video Understanding
+
+Early deep learning models for video understanding primarily followed two main approaches [52]. One uses two-stream networks [18, 19, 42, 58] to integrate spatial and motion information, while the second focuses on 3D CNNs [8, 57, 72] for spatial-temporal feature extraction. Vision Transformers [14] then introduce self-attention mechanisms
+
+![](images/36be31cb5baa1db2c03e215ae3ccb61924a453519badc0eb3c88049bd76108e3.jpg)  
+Figure 3. Pipeline of VCA Framework. First, the reward model predicts the relevance scores of each segment. Next, the memory buffer is updated by adding new frames and discarding irrelevant ones. Finally, emulating human reasoning, the agent actively selects segments to explore until the correct answer is found.
+
+to handle temporal modeling more effectively [3, 7, 16]. Methods with self-supervised pre-training [50, 56, 86] exhibit transferability across different tasks and achieve more robust video understanding performance.
+
+Multimodal LLM [22, 59, 83] shows strong capabilities in video understanding tasks, but efficiently processing information within a prolonged video environment is the key challenge. Recent approaches often use frame sampling [32, 46, 63, 75, 77] and feature extraction [27, 49, 51, 70] to reduce computational costs. Frame sampling selects a subset of frames, which risks missing essential context in lengthy videos, where transitions and subtle cues may be easily neglected. Additionally, processing entire video sequences as tokens presents significant memory and computational challenges [11, 43, 64, 80].
+
+Recent work has explored auxiliary techniques and modules to improve efficiency, such as pre-extracted content like captions [44, 62], multimodal summaries [40, 65], and combinations of various auxiliary tools [17, 30]. However, these tools introduce additional computational and memory costs. Moreover, some approaches still require traversing the entire video content [17]. Notably, VideoTree [66] requires captioning every frame in the video, making it prohibitively expensive. In contrast, our VCA can solve the task using a single VLM, without relying on any supplementary tools.
+
+# 3. Method
+
+Inspired by human reasoning, we introduce a self-exploration agent that adaptively navigates video environments. Unlike approaches that depend on dense frame sampling or captioning [62], which require VLMs to process numerous frames, our agent analyzes videos iteratively within a limited context window in a process we call self-exploration. This approach allows essential information to be extracted from long videos in a coarse-to-fine manner.
+
+Specifically, we propose a segment-based tree-search exploration structure. Given a video  $v$  and a user query  $q$ , the agent  $\pi$  begins by randomly sampling  $N$  frames  $\{f_v^i\}_{i=1}^N$  to obtain an initial overview of the content in each video segment  $v_i \in \{v_i\}_{i=1}^{N+1}$ . The agent then decides whether additional exploration is needed to answer the query. If further investigation is necessary, the agent selects a specific segment  $s^* \in \{v_i\}_{i=1}^{N+1}$  for deeper exploration until sufficient information is gathered to respond to the query.
+
+Nevertheless, a one-hour movie typically contains approximately 86,400 frames, posing a challenge for current VLM-based agents to efficiently identify the most relevant segment [17, 62, 66]. Moreover, even with a small  $N$ , conducting multiple rounds of exploration accumulates a substantial number of frames, leading to high computational costs. To address these issues, we introduce a reward model  $R$  to guide the agent's exploration and propose a memory
+
+buffer  $M$  to reduce storage and inference costs.
+
+In this section, we first explain how the reward model assists the agent throughout the exploration process (Sec. 3.1). Then we introduce our memory management strategy in Sec. 3.2. Finally, we explain how VCA adaptively explores the video using a segment-based tree structure (Sec. 3.3).
+
+# 3.1. Reward Model
+
+With the segment-based tree-search exploration, the agent is required to select the most informative video segment for exploration at each step. However, we observe that current VLM-based agents struggle to independently identify optimal exploration paths (see Sec. 4.3). To this end, inspired by the selective attention mechanism in human cognition [13], we incorporate a reward model  $R$  to guide the agent's selection process, i.e., the agent will gain a higher reward if it plans to take the action to explore more relevant segment.
+
+Specifically, during the exploration of a current selected segment  $s^*$ , the reward model observes a set of frames  $\{f_{s^*}^i\}_{i=1}^N$  sampled from  $s^*$ . These frames  $\{f_{s^*}^i\}_{i=1}^N$  then form finer sub-segments  $\{s_i^* = [l_{s^*}^i, l_{s^*}^{i+1}]\}_{i=1}^{N+1}$ , where  $l_{s^*}^i$  and  $l_{s^*}^{i+1}$  represent the indices of the beginning and end frame of the sub-segment, respectively. The reward model aims to generate a reward score for each sub-segment  $\{s_i^*\}_{i=1}^{N+1}$  based on its relevance to the user query  $q$ . We utilize chain-of-thoughts [67] technique to let the reward model provide a verbal explanation of each segment and then generate a relevant score, i.e.,  $\{t_{s_i^*}; r_{s_i^*}\}_{i=1}^{N+1}$ . Additionally, to maintain consistency across different exploration steps, we incorporate the reward history  $H_r$  from previous steps alongside the new segments during evaluation. A simplified prompt template for the reward model is presented here, with detailed prompts in Appendix ??
+
+# Reward Model Prompt (Simplified)
+
+Given:
+
+$H_{r}$ , reward score history in previous steps.
+
+$\left\{f_{s^*}^i\right\}_{i = 1}^N$  newly sampled frames from current segment  $s^{*}$
+
+Please first explain each sub-segment  $\{s_i^*\}_{i = 1}^{N + 1}$  and then assign a relevance score of each sub-segment to the user query  $q$ .
+
+In practice, rather than training a separate specialized model, we utilize a same VLM-based agent to function as the reward model. While the reward model  $R$  and exploration agent  $\pi$  serve distinct purposes, both can be implemented using the same model. In this work, we employ a single VLM-based agent for both roles across different settings, providing a neat and efficient agent framework that is guided by its self-generated intrinsic reward and without
+
+relying on auxiliary models or tools.
+
+# 3.2. Memory Management
+
+As exploration proceeds over multiple rounds, the accumulation of frames can lead to significant computational costs for long videos, even when  $N$  is small. However, as illustrated in Fig. 2, they tend to gradually disregard irrelevant frames, especially from earlier steps, retaining only the most crucial information in their working memory for later analysis. Inspired by this, we introduce a fixed-size memory buffer  $M$  that stores only the most relevant frames throughout exploration.
+
+Specifically, when the agent  $\pi$  explores a video segment  $s^*$ , the associated frames  $\{f_{s^*}^i\}_{i=1}^N$  are added to the memory buffer. If the total number of frames exceeds the buffer limit, irrelevant frames are removed based on relevance scores  $r_{s_i^*}$  defined in Sec. 3.1. We discard the frames with the lowest scores and only retain the most relevant ones. While these less relevant frames are removed, their visual information is extracted and encoded as text descriptions  $\{t_{s_i^*}\}_{i=1}^N$  within the exploration history  $H_r$ , allowing the reward model to utilize. This memory management strategy enables the agent to conduct reasoning within a fixed resource limit, regardless of the exploration trajectory's length, thereby preventing excessive memory usage.
+
+# 3.3. Tree-Search Exploration
+
+Inspired by humans adaptively focusing on different video segments during exploration rather than targeting specific frames directly, our proposed agent VCA treats the video as an unexplored environment, where each segment represents a node on an exploration tree. Given the entire video as the root node, the agent's goal is to efficiently identify an optimal short path leading to a leaf node (segment) containing the most informative content needed to answer the query.
+
+At each step, the agent  $\pi$  receives both textual and visual guidance: a set of candidate segments  $S$ , which includes a collection of segments with associated reward scores, and a memory buffer  $M$  containing multiple frames. If the agent has gathered sufficient information to answer the user's query  $q$ , it directly generates a response  $a$ . Otherwise, it continues exploring by selecting a specific segment  $s^* \in S$  to expand the tree based on reward scores.
+
+However, due to the limited capacity of current foundational models, these reward scores are not always accurate (see further analysis in Sec. 5.3). Selecting segments based on reward scores in a purely greedy manner can potentially lead to a suboptimal solution. Instead, we use the reward scores as guidance for agent  $\pi$ , allowing it to make the final segment selection independently. In this case, the agent is not forced to select the segment with the highest reward, balancing the exploitation and exploration. More importantly, the candidate set  $S$  includes both unexplored
+
+coarse segments from prior steps and fine-grained segments from the current step, enabling the agent to backtrack and select earlier segments if it detects that the current exploration path may be suboptimal. Below are the simplified prompts. The full prompt is provided in Appendix ??
+
+# Exploration Agent Prompt (Simplified)
+
+Given:
+
+Candidate segments set:  $S$
+
+Frames in Memory buffer:  $M$
+
+Please choose one segment to explore based on the query  $q$ . Output answer if you have gathered sufficient information.
+
+The overall framework of our proposed method is summarized in Algo. 1. Leveraging a tree-search exploration technique boosted with a reward model and a memory buffer, our self-exploration agent adaptively and efficiently explores the video like a human. Empirical results in Sec. 4.3 further validate the effectiveness of our tree-search exploration over naive frame selection methods.
+
+# 4. Experiments
+
+# 4.1. Experimental Settings
+
+Benchmark. We mainly evaluate VCA on two long-term video question-answering benchmarks. EgoSchema [41] is an egocentric video dataset sourced from Ego4D [21], with a duration of 180 seconds. Here we use the validation set of 500 QA pairs. LVBench [61] covers about 1,500 samples for evaluation of reasoning abilities from high-quality videos over 30 minutes in length. We also conduct experiments on other relevant benchmark datasets in Appendix ??
+
+Baselines. We compare with most relevant baselines to our work, including state-of-the-art open-source Video-LLMs [10, 23, 37, 47, 59, 82], the close-source model GPT-4o [25], and agent-based systems [44, 62, 66]. For video models, we report the results obtained with uniform sampling frames, while for the agent-based systems, we follow their official implementation with GPT-4o [25], consistent with VCA's setup to ensure a fair comparison.
+
+Implementation Details. In this work, we use the same model for the exploration agent and the reward model. Unless otherwise stated, all experiments were performed using the August 2024 version of `gpt-4` as the base model, leveraging the OpenAI LLM API service<sup>1</sup> with a temperature of 0.5. We set memory buffer sizes set to 8 and 16 frames for EgoSchema and LVBench, respectively. Further implementation details are provided in Appendix ??
+
+Algorithm 1 VCA: Video Curious Agent  
+Ensure: Answer  $a$  
+```txt
+Require: Video environment  $v$ , question  $q$ , agent  $\pi$ , reward model  $R$ , memory buffer  $M$ , candidate segments set  $S$ , reward score history  $H_{r}$ , sampling frame number  $N$ .
+```
+
+1: Initialize selected segment  $s^* \gets v$  
+2: Initialize candidate segments set  $S \gets \emptyset$  
+3: Initialize memory buffer  $M\gets \emptyset$  
+4: Initialize reward score history  $H_{r}\gets \emptyset$  
+5: continue  $\leftarrow$  true  
+6: while continue do  
+7: # Sampling frames & form segments  
+8:  $\left\{f_{s^*}^*\right\}_{i=1}^{N}, \left\{s_i^*\right\}_{i=1}^{N+1} \gets \text{UniformSample}(s^*)$  
+9: # Scoring each segment  
+10:  $\left\{t_{s_i^*};r_{s_i^*}\right\}_{i = 1}^{N + 1}\gets R(\{s_i^*\}_{i = 1}^{N + 1};q,\{f_{s_*}^i\}_{i = 1}^N,H_r)$  
+11: # Update candidate segment set  
+12:  $S\gets S\cup \{s_i^*;r_{s_i^*}\}_{i = 1}^{N + 1}$  
+13: # Update history with thoughts  
+14:  $H_{r}\gets H_{r}\cup \left\{t_{s_{i}^{*}};r_{s_{i}^{*}}\right\}_{i = 1}^{N + 1}$  
+15: # Update memory with frames  
+16:  $M\gets \mathrm{UpdateMemory}(M,\{f_{s^{*}}^{i}\}_{i = 1}^{N};S)$  
+17: # Output answer or select segment  
+18:  $a,s\gets \pi (q,S;M)$  
+19: if  $a = \emptyset$  then  
+20: # Continue exploration  
+21:  $s^* \gets s$  
+22: else  
+23: continue  $\leftarrow$  false  
+24: end if  
+25: end while  
+26: return  $a$
+
+# 4.2. Experimental Results
+
+The experimental results on LVBench and EgoSchema are presented in Tab. 2 and Tab. 1, respectively. By adaptively focusing on critical segments rather than scanning the entire video, our agent not only achieves stronger performance but also incurs less computational cost. With less than  $30\%$  of the frames, VCA achieves substantial improvements over directly feeding uniformly sampled frames into GPT-4o. Specifically, we observe performance gains of  $3.2\%$  on the EgoSchema subset and  $6.6\%$  on LVBench. Significant increases in the Event Understanding and Temporal Grounding domains further underscore our effectiveness in identifying crucial segments within long videos.
+
+Furthermore, we compare our framework with recently proposed SOTA video agent systems, VideoAgent [62] and VideoTree [66]. For a fair comparison, we re-implement both frameworks using GPT-4o for all reasoning components, consistent with our setup. However, captioning all
+
+<table><tr><td>Method</td><td>Frames</td><td>ER</td><td>EU</td><td>KIR</td><td>TG</td><td>Rea</td><td>Sum</td><td>Avg.</td></tr><tr><td>TimeChat* [47]</td><td>&gt;96</td><td>21.9</td><td>21.7</td><td>25.9</td><td>22.7</td><td>25.0</td><td>24.1</td><td>22.3</td></tr><tr><td>LLaVA-NeXT* [83]</td><td>32</td><td>30.1</td><td>31.2</td><td>34.1</td><td>31.4</td><td>35.0</td><td>27.6</td><td>32.2</td></tr><tr><td>InternVL2-40B* [10]</td><td>16</td><td>37.4</td><td>39.7</td><td>43.4</td><td>31.4</td><td>42.5</td><td>41.4</td><td>39.8</td></tr><tr><td>GLM4V-Plus* [23]</td><td>20</td><td>39.9</td><td>35.8</td><td>34.8</td><td>37.7</td><td>40.0</td><td>32.8</td><td>38.3</td></tr><tr><td>GPT-4o* [25]</td><td>64</td><td>35.9</td><td>30.8</td><td>35.5</td><td>28.3</td><td>33.5</td><td>34.5</td><td>34.7</td></tr><tr><td>VideoAgent† [62]</td><td>Avg. 25.5</td><td>28.0</td><td>30.3</td><td>28.0</td><td>29.3</td><td>28.0</td><td>36.4</td><td>29.3</td></tr><tr><td>VideoTree† [66]</td><td>Avg. 103.2</td><td>30.3</td><td>25.1</td><td>26.5</td><td>27.7</td><td>31.9</td><td>25.5</td><td>28.8</td></tr><tr><td>VCA (ours)</td><td>Avg. 20.0</td><td>43.7</td><td>40.7</td><td>37.8</td><td>38.0</td><td>46.2</td><td>27.3</td><td>41.3</td></tr></table>
+
+Table 1. Experimental Results on LVBench. Results of methods marked with * are sourced from LVBench [61]. Methods marked with † indicate implementations where GPT-4o serves as the main component. Our agent's memory buffer size is set to 16 frames. We report the average number of observed frames for agent-based methods, with the best performance highlighted in bold.  
+
+<table><tr><td>Method</td><td>Frames</td><td>Subset</td></tr><tr><td rowspan="2">GPT-4o [25]</td><td>8</td><td>69.0</td></tr><tr><td>32</td><td>70.4</td></tr><tr><td rowspan="2">\(VideoAgent^\dagger\) [62]</td><td>Avg. 8.1</td><td>63.2</td></tr><tr><td>Avg. 10.7</td><td>60.8</td></tr><tr><td>\(VideoTree^\dagger\) [66]</td><td>Avg. 69.5</td><td>67.0</td></tr><tr><td>\(LVNet^\dagger\) [44]</td><td>12</td><td>68.2</td></tr><tr><td>VCA (ours)</td><td>Avg. 7.2</td><td>73.6</td></tr></table>
+
+frames with GPT-4o is prohibitively expensive. Hence, following the baseline implementations, we use open-source VLMs for captioning. A comparison using the same model for all components is provided in Sec. 5.1. As shown in Tab. 1, VCA outperforms all baselines over  $5\%$  on EgoSchema, while requiring processing fewer frames. These results suggest the limitations of dense captioning for video preprocessing tasks, which can incur more computational costs while introducing more redundant information. In contrast, our proposed memory management strategy effectively improves efficiency by only retaining the most crucial information for reasoning. Similar insights can be drawn from results on LVBench in Tab. 2. Additional results are provided in Appendix ??
+
+Moreover, we observe that VCA explores approximately three times more frames on LVBench than on EgoSchema. Recall that EgoSchema videos have an average duration of 180 seconds, while videos in LVBench extend to at least 30 minutes. This phenomenon suggests that our proposed tree-search exploration technique actively explores the video to locate the most relevant information and adaptively in-
+
+Table 2. Experimental Results on EgoSchema. Methods marked with  $\dagger$  indicate implementations where GPT-4o is the main component. The results of LVNet are from [44]. Our agent's memory buffer size is set to 8 frames. We report the average number of observed frames, with the best performance highlighted in bold.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">EgoSchema</td><td colspan="2">LVBench</td></tr><tr><td>Frames</td><td>Acc.</td><td>Frames</td><td>Acc.</td></tr><tr><td>VCA (ours)</td><td>7.2</td><td>73.6</td><td>20.0</td><td>41.3</td></tr><tr><td>- w/o Reward</td><td>7.5</td><td>72.6</td><td>22.2</td><td>39.5</td></tr><tr><td>- w/o Tree Search</td><td>10.2</td><td>69.8</td><td>39.9</td><td>36.2</td></tr></table>
+
+Table 3. Ablation Study on EgoSchema and LVBench. Both the tree-search exploration and reward model contribute significantly to improvements in performance and efficiency. Detailed ablation study results on LVBench are included in Appendix ??  
+
+<table><tr><td>Buffer Size</td><td>8</td><td>16</td><td>32</td></tr><tr><td>Acc.</td><td>38.9</td><td>41.3</td><td>43.2</td></tr></table>
+
+Table 4. Ablation Study of Buffer Size on LVBench. Increasing the buffer size allows the model to observe more frames during reasoning, leading to improved performance.
+
+creases the exploration budget based on task complexity. For example, VCA exhibits both exploitation and exploration: when the agent determines that the current exploration path is suboptimal, it backtracks to previous segments to explore a new path (see Sec. 5.2). These empirical observations further validate the robustness and scalability of our approach, especially in diverse long-video contexts.
+
+# 4.3. Ablation Study
+
+In this section, we conduct an ablation study to evaluate the components of our agent. Results across both benchmarks are presented in Tab. 3. As illustrated in the table, performance consistently declines when the reward model is removed, highlighting that the agent's inherent reasoning ability alone is insufficient for accurately identifying the most informative segments. This finding emphasizes the effectiveness of incorporating an intrinsic reward to guide the agent's segment selection process. A detailed analysis of our reward model's performance is provided in Sec. 5.3.
+
+Additionally, we observe a significant performance drop
+
+Why does the mice run away?  
+(A) Because the black cat sheriff calls the police  
+(C) Because their boss is injured
+
+(B) Because they meet more powerful mice  
+(D) Because the owner of the house is back
+
+![](images/494c2e39d33787b489c52c42d9cf3f62f97d802bf3cb5b76c77ffd4cb93b8d38.jpg)
+
+# Reward Model
+
+Segment1: Thought: The segment shows a black cat with a gun, possibly indicating a threat. This could suggest that the mice might run away due to the presence of the cat, which aligns with option (A). Score: 0.7  
+Segment4: Thought: The segment features the black cat with a communication device, possibly calling for help or backup. This aligns with option (A), suggesting the mice might run because the cat is calling the police. Score: 0.8
+
+![](images/cab27bdb7ff0b0177e4565f738e98f8d8b8a7fec79b098210f2663bdfcdbd937.jpg)  
+Agent Choice: (segment: 4)
+
+![](images/db534f7962bd217c6a8212442a281185644644533499078be1979c927fe8315b.jpg)
+
+![](images/9709989d9606dbfa36835e193981ef9fd1b3c4452fb7a2442047d274c61b2bbb.jpg)
+
+# Reward Model
+
+Segment4.1: Thought: The segment shows a close-up of a document with red text and exclamation marks, indicating urgency or importance. This could imply a critical message or alert being communicated. Score: 0.7
+
+…
+
+![](images/0be7582fe7139255d5979ad1a0a4b2f8416fde86fabbd6434c0c69aee2357f20.jpg)
+
+# Reward Model
+
+Segment1.2: Thought: The segment features mice characters in mid-air, possibly fleeing or being thrown. This suggests a direct reaction to a threat, supporting the idea that they are running away from the cat. Score: 0.8
+
+![](images/e82b2c43037a928c41960aef4203fd12a1f1ce59e083cac00c6ba6a3ec87c9ad.jpg)  
+Figure 4. Exploration Trajectory Example. The agent generally tends to select video segments with higher reward scores for exploration, yet it can adaptively redirect to earlier segments, effectively balancing exploitation and exploration.
+
+![](images/5386d87e480fee48366b58408db92687c9c944cdedb4c56e93e21002e2bd9557.jpg)
+
+Agent: Answer: A
+
+when the tree-search exploration technique is ablated, i.e., the agent directly selects frame indices without the segment-based tree structure. Interestingly, this modification also results in a substantial increase in the average number of frames acquired, especially on LVBench. This indicates that the agent struggles to locate the most crucial information and consequently produces a longer but redundant trajectory. These ablation study results validate the effectiveness of the tree-search exploration framework assisted by the reward model. Additionally, this underscores VCA's potential to achieve greater efficiency and improved performance with more advanced VLMs and reward models.
+
+Moreover, as shown in Tab. 4, increasing the memory buffer size yields consistently improved performance for VCA, demonstrating that our framework benefits from a broader reasoning window while maintaining effectiveness with fewer frames. Due to limited computation resources, we only tested up to 32 frames; however, the upward trend indicates the potential for gains with additional frames.
+
+# 5. Analysis
+
+# 5.1. Can VCA generalize to Open-source Models?
+
+While VCA demonstrates strong performance using GPT4o as the base model, we also analyze whether these positive results can be extrapolated to open-source models. We implement the VCA framework with Qwen2-VL-72B-Instruct-GPTQ-Int4 [59] as both the exploration agent and reward model, as well as the captioner for baselines. The results in Tab. 5 reveal that, compared to directly feeding uniformly sampled frames into Qwen2-VL, our framework achieves higher accuracy on the EgoSchema benchmark despite using only about  $60\%$  of the frames. In contrast, VideoAgent performs significantly worse. Furthermore, our approach outperforms VideoTree with far fewer frames. Notably, VideoTree employs clustering to build a caption similarity-based tree structure for frame selection, requiring caption generation for each frame. This process is computationally expensive, demanding over 60 GPU hours
+
+<table><tr><td>Method</td><td>Frames</td><td>Subset</td></tr><tr><td rowspan="2">Qwen2-VL</td><td>8</td><td>70.2</td></tr><tr><td>32</td><td>74.2</td></tr><tr><td>VideoAgent</td><td>Avg. 8.6</td><td>65.2</td></tr><tr><td>VideoTree</td><td>Avg. 85.4</td><td>71.0</td></tr><tr><td>VCA (ours)</td><td>Avg. 16.9</td><td>75.2</td></tr><tr><td>- w/ Image Reward</td><td>Avg. 20.9</td><td>74.0</td></tr></table>
+
+Table 5. Performance of Using Open-source Model on EgoSchema. Our approach achieves higher accuracy with fewer observed frames, showing consistency improvements over baseline methods. Memory buffer size is set to 16 frames.
+
+on a single H100, as shown in Fig. 1. These results further underscore the robustness of VCA.
+
+# 5.2. What's the Exploration Behavior of VCA?
+
+Here we present a real exploration trajectory from LVBench to illustrate our agent's decision-making process. As shown in Fig. 4, given a user query, the reward model initially evaluates each segment's relevance based on uniformly sampled frames, and the agent selects the fourth segment, which has the highest relevance score. However, upon further investigation of segment 4, both the reward model and the agent find it relatively irrelevant to the query. Rather than continuing with it, the agent redirects its exploration to segment 1 in the last round. After selecting segment 1, the agent chooses to investigate the third segment, despite the second segment having the highest reward score. This behavior reflects VCA's ability to adaptively adjust its exploration trajectory based on visual context rather than purely following the reward score, resulting in a more robust system.
+
+We also examine the limitations of VCA by analyzing its failure cases, which fall into the following scenarios: The most common failure occurs when the agent overlooks subtle clues, resulting in an inaccurate response to the user query even after prolonged exploration, particularly for queries requiring specific visual details. Other failure cases include instances where the agent is misled by inaccurate reward scores or fails to provide the correct answer despite observing key frames. Detailed cases and further discussion are provided in the Appendix ??
+
+In addition, unlike previous methods, VCA employs a segment-level exploration strategy instead of direct image similarity to select key frames. To further investigate this difference, we evaluate the performance of selecting the segment whose start and end images have the highest visual relevance. As shown in Tab. 5, relying solely on image relevance instead of inferring segment semantics results in a performance drop on EgoSchema, confirming the effectiveness of our segment-level exploration.
+
+![](images/fca8834cee31104ddaa6b77c3ab66dbfffdf08de055321278fd4a45f2a317be3.jpg)  
+Figure 5. Segment Selection Accuracy on LVBench. The y-axis shows segment distance from ground truth; x-axis shows count ratio. Red curves are fitted normal distributions. The reward-guided agent selects segments closer to ground truth than the reward model.
+
+# 5.3. How Reliable is the Reward Model?
+
+To further assess the capability of our reward model, we visualize the matching accuracy between segments with the highest reward scores and ground truth time references from LVBench. The results are shown at the bottom of Fig. 5. We observe that segment distances follow an approximately normal distribution with a mean of 1.28, indicating that the segment with the highest reward score often closely aligns with the ground truth segment. This validates the reward model's effectiveness in providing meaningful guidance.
+
+Additionally, as discussed in Sec 3.3, we encourage the agent to make independent decisions when selecting segments rather than blindly following reward scores in a purely greedy manner. The benefit of this design choice is demonstrated at the top of Fig. 5, where we present the matching accuracy of the agent's choices. Compared to the greedy approach, the agent's selections result in a sharper distribution of segment distances, with an average distance gap of  $-0.60$ , indicating closer alignment with the ground truth. In summary, our reward model assigns reliable relevance scores to segments, while the agent further refines these choices to achieve even more accurate selections.
+
+# 6. Conclusion
+
+In this work, we propose a curiosity-driven video agent framework for long video understanding tasks with self-exploration capability. VCA achieves state-of-the-art performance and efficiency without relying on auxiliary tools. Appendix ?? shows that improved reward guidance could further boost results, highlighting the potential of synthetic data and specialized reward models for future work.
+
+# 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] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Devendra Chaplot, Jessica Chudnovsky, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amélie Héliou, Paul Jacob, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 2  
+[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6816-6826, 2021. 3  
+[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1  
+[5] Ivana Balazevic, Yuge Shi, Pinelopi Papalampidi, Rahma Chaabouni, Skanda Koppula, and Olivier J Henaff. Memory consolidation enables long-context video understanding. In *Forty-first International Conference on Machine Learning*, 2024. 2  
+[6] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Fuyu-8b: A multimodal architecture for ai agents, 2023. 2  
+[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding?, 2021. 3  
+[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset, 2018. 1, 2  
+[9] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 2  
+[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 5, 6  
+[11] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 3  
+[12] Jiwan Chung and Youngjae Yu. Long story short: a summarize-then-search method for long video question answering. arXiv preprint arXiv:2311.01233, 2023. 2  
+[13] Robert Desimone, John Duncan, et al. Neural mechanisms of selective visual attention. Annual review of neuroscience, 18(1):193-222, 1995. 4
+
+[14] 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, 2021. 2  
+[15] Zane Durante, Qiuyuan Huang, Naoki Wake, Ran Gong, Jae Sung Park, Bidipta Sarkar, Rohan Taori, Yusuke Noda, Demetri Terzopoulos, Yejin Choi, Katsushi Ikeuchi, Hoi Vo, Li Fei-Fei, and Jianfeng Gao. Agent ai: Surveying the horizons of multimodal interaction, 2024. 2  
+[16] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers, 2021. 3  
+[17] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2025. 1, 2, 3  
+[18] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition, 2016. 1, 2  
+[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 2  
+[20] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2  
+[21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 5  
+[22] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, Lei Zhao, Zhuoyi Yang, Xiaotao Gu, Xiaohan Zhang, Guanyu Feng, Da Yin, Zihan Wang, Ji Qi, Xixuan Song, Peng Zhang, Debing Liu, Bin Xu, Juanzi Li, Yuxiao Dong, and Jie Tang. CogvIm2: Visual language models for image and video understanding, 2024. 3  
+[23] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 5, 6  
+[24] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 2  
+[25] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli
+
+hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 5, 6  
+[26] Lawrence Jang, Yinheng Li, Charles Ding, Justin Lin, Paul Pu Liang, Dan Zhao, Rogerio Bonatti, and Kazuhito Koishida. Videowebarena: Evaluating long context multimodal agents with video understanding web tasks, 2024. 2  
+[27] Peng Jin, Ryuichi Takanobu, Caiwan Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding, 2023. 3  
+[28] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks, 2024. 2  
+[29] Bruno Korbar, Yongqin Xian, Alessio Tonioni, Andrew Zisserman, and Federico Tombari. Text-conditioned resampler for long form video understanding. In European Conference on Computer Vision, pages 271-288. Springer, 2025. 1  
+[30] Somnath Kumar, Yash Gadhia, Tanuja Ganu, and Akshay Nambi. Mmctagent: Multi-modal critical thinking agent framework for complex visual reasoning, 2024. 3  
+[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1  
+[32] Yunxin Li, Xinyu Chen, Baotain Hu, and Min Zhang. Llms meet long video: Advancing long video comprehension with an interactive visual adapter in llms, 2024. 1, 3  
+[33] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2025. 1  
+[34] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1  
+[35] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 2  
+[36] Ming C Lin and Shan Yang. Vila: Efficient video-language alignment for video question answering. 1  
+[37] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5  
+[38] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv preprint arXiv:2402.08268, 2024. 1  
+[39] Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. Drvideo: Document retrieval based long video understanding. arXiv preprint arXiv:2406.12846, 2024. 2  
+[40] Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. Drvideo: Document retrieval based long video understanding, 2024. 3
+
+[41] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. arXiv preprint arXiv:2308.09126, 2023. 5  
+[42] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification, 2015. 2  
+[43] Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14386-14397, 2024. 3  
+[44] Jongwoo Park, Kanchana Ranasinghe, Kumara Kahatapitiya, Wonjeong Ryoo, Donghyun Kim, and Michael S Ryoo. Too many frames, not all useful: Efficient strategies for long-form video qa. arXiv preprint arXiv:2406.09396, 2024. 1, 3, 5, 6  
+[45] Ruchit Rawal, Khalid Saifullah, Ronen Basri, David Jacobs, Gowthami Somepalli, and Tom Goldstein. Cinepile: A long video question answering dataset and benchmark. In Synthetic Data for Computer Vision Workshop @ CVPR 2024. 2  
+[46] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. arXiv preprint arXiv:2312.02051, 2023. 3  
+[47] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 5, 6  
+[48] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5026-5035, 2022. 1  
+[49] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023. 1, 2, 3  
+[50] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning, 2019. 3  
+[51] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. Advances in neural information processing systems, 35:38032-38045, 2022. 3  
+[52] Yunlong Tang, Jing Bi, Sitting Xu, Luchuan Song, Susan Liang, Teng Wang, Daoan Zhang, Jie An, Jingyang Lin, Rongyi Zhu, Ali Vosoughi, Chao Huang, Zeliang Zhang, Pinxin Liu, Mingqian Feng, Feng Zheng, Jianguo Zhang,
+
+Ping Luo, Jiebo Luo, and Chenliang Xu. Video understanding with large language models: A survey, 2024. 2  
+[53] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4631–4640, 2016. 1  
+[54] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2  
+[55] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1  
+[56] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training, 2022. 3  
+[57] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks, 2015. 1, 2  
+[58] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition, 2016. 2  
+[59] 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. arXiv preprint arXiv:2409.12191, 2024. 3, 5, 7  
+[60] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1  
+[61] Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Shiyu Huang, Bin Xu, Yuxiao Dong, Ming Ding, and Jie Tang. Lvbench: An extreme long video understanding benchmark, 2024. 1, 5, 6  
+[62] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. European Conference on Computer Vision (ECCV), 2024. 1, 2, 3, 5, 6  
+[63] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 1, 3  
+[64] Yuxuan Wang, Cihang Xie, Yang Liu, and Zilong Zheng. Videollamb: Long-context video understanding with recurrent memory bridges. arXiv preprint arXiv:2409.01071, 2024. 1, 3  
+[65] Ying WANG, Yanlai YANG, and Mengye REN. Lifelong-memory: Leveraging llms for answering queries in long
+
+form egocentric videos. arXiv preprint arXiv:2312.05269, 2024. 1, 3  
+[66] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos, 2024. 1, 2, 3, 5, 6  
+[67] 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. Advances in neural information processing systems, 35:24824-24837, 2022. 4  
+[68] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. In European Conference on Computer Vision, pages 453-470. Springer, 2025. 1  
+[69] Chao-Yuan Wu and Philipp Krahenbuhl. Towards long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1884-1894, 2021. 1  
+[70] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 3  
+[71] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. CoRR, 2024. 1  
+[72] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification, 2018. 2  
+[73] Tianbao Xie, Fan Zhou, Zhoujun Cheng, Peng Shi, Luoxuan Weng, Yitao Liu, Toh Jing Hua, Junning Zhao, Qian Liu, Che Liu, et al. Openagents: An open platform for language agents in the wild. In ICLR 2024 Workshop on Large Language Model (LLM) Agents. 2  
+[74] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. arXiv preprint arXiv:2404.07972, 2024. 2  
+[75] Jiaqi Xu, Cuiling Lan, Wenxuan Xie, Xuejin Chen, and Yan Lu. Retrieval-based video language model for efficient long video question answering, 2023. 1, 3  
+[76] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents, 2023. 2  
+[77] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. NeurIPS, 2023. 3  
+[78] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. Advances in Neural Information Processing Systems, 36, 2024. 1  
+[79] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple lmm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 2
+
+[80] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, 2023. 3  
+[81] Hongjie Zhang, Yi Liu, Lu Dong, Yifei Huang, Zhen-Hua Ling, Yali Wang, Limin Wang, and Yu Qiao. Movqa: A benchmark of versatile question-answering for long-form movie understanding. arXiv preprint arXiv:2312.04817, 2023. 2  
+[82] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 5  
+[83] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 3, 6  
+[84] Jun Zhao, Can Zu, Hao Xu, Yi Lu, Wei He, Yiwen Ding, Tao Gui, Qi Zhang, and Xuanjing Huang. Longagent: Scaling language models to 128k context through multi-agent collaboration. arXiv preprint arXiv:2402.11550, 2024. 2  
+[85] Long Zhao, Nitesh Bharadwaj Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. In *Forty-first International Conference on Machine Learning*. 1  
+[86] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations, 2020. 3

2025/VCA_ Video Curious Agent for Long Video Understanding/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f79d775bae197443e1f123dc4992a2b2bf38d4f1358bd21c18627b16e731cc65
+size 439132

2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/d727aeb7-1f89-498a-ac85-9bcc05bddf8b_content_list.json

@@ -0,0 +1,1315 @@
+[
+    {
+        "type": "text",
+        "text": "VEGGIE: Instructional Editing and Reasoning Video Concepts with Grounded Generation",
+        "text_level": 1,
+        "bbox": [
+            250,
+            128,
+            748,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Shoubin Yu $^{1,3*}$  Difan Liu $^{1*}$  Ziqiao Ma $^{1,2*}$  Yicong Hong $^{1}$  Yang Zhou $^{1}$  Hao Tan $^{1}$  Joyce Chai $^{2}$  Mohit Bansal $^{3}$",
+        "bbox": [
+            202,
+            202,
+            794,
+            243
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Adobe Research  $^{2}$ University of Michigan  $^{3}$ UNC Chapel Hill",
+        "bbox": [
+            235,
+            255,
+            761,
+            273
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "https://veggie-gen.github.io/",
+        "bbox": [
+            351,
+            276,
+            640,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/6de461b48922321aa8d110e2c86339f81fccd8681d7979d3daf8439499923d75.jpg",
+        "image_caption": [
+            "Figure 1. We propose VEGGIE, a unified and versatile video generative model that handles various tasks for both video concept grounding and editing according to user instructions. With VEGGIE, users can locate, add, delete, and change concepts in a given video through diverse instruction formats (direct referring instruction or reasoning-demanding questions). Users can also edit videos with multimodal instruction empowered by MLLM, enabling applications like video editing from a reference image."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            133,
+            309,
+            857,
+            555
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            635,
+            326,
+            651
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "While recent video diffusion models enable video editing, unifying diverse instructional editing tasks (e.g., add, remove, modify) under a single framework remains a significant challenge. In this paper, we introduce VEGGIE, a Video Editor with Grounded Generation from Instructions, a simple end-to-end framework that unifies video concept editing, grounding, and reasoning based on diverse user instructions. Specifically, given a video and text query, VEGGIE first utilizes an MLLM to interpret user intentions in instructions and ground them to the video contexts, generating framespecific grounded task queries for pixel-space responses. A diffusion model then renders these plans and generates edited videos that align with user intent. To support diverse",
+        "bbox": [
+            88,
+            667,
+            485,
+            864
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tasks and complex instructions, we employ a curriculum learning strategy: first aligning the MLLM and video diffusion model with large-scale instructional image editing data, followed by end-to-end fine-tuning on high-quality multitask video data. Additionally, we introduce a novel data synthesis pipeline to generate paired instructional video editing data for model training. It transforms static image data into diverse, high-quality video editing samples by leveraging Image-to-Video models to inject dynamics. VEGGIE shows strong performance in instructional video editing with different editing skills, outperforming the best instructional baseline as a versatile model, while other models struggle with multi-tasking. VEGGIE also excels in video object grounding and reasoning segmentation, where other baselines fail. We further reveal how the multiple tasks help each other and highlight promising applications like zero-shot multimodal instructional and in-context video editing.",
+        "bbox": [
+            511,
+            637,
+            908,
+            893
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV 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": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Equal contribution.",
+        "bbox": [
+            107,
+            875,
+            218,
+            887
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "$\\dagger$  Work done during internship at Adobe Research.",
+        "bbox": [
+            109,
+            887,
+            372,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "15147",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            91,
+            89,
+            222,
+            106
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Building on the advances in Video Diffusion Models (VIDDMs) [2, 4, 24, 25, 74], video editing methods have emerged as video design tools, allowing users to manipulate video concepts such as adding, removing, altering objects and style translation [20, 53, 62, 79, 85]. To enhance user experiences, instructional video editing methods [54, 86] have been developed, using triples of text prompts, source videos, and target videos for training. Due to their limited performance in understanding user intent and multimodal semantics [17], several methods have incorporated multimodal large language models (MLLMs) to handle complex instructions/reasoning [17, 29, 75].",
+        "bbox": [
+            89,
+            114,
+            485,
+            296
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "However, existing methods fall short of the goal of a simple, versatile video concept editor, facing three primary challenges. First, most methods are not end-to-end, requiring intermediate layout/mask/human or model caption guidance [35, 53, 63, 75], which adds workload on users and disrupts a seamless editing experience. Second, existing pipelines connecting MLLMs to VidDMs require multiple training objectives beyond simple pixel-space diffusion loss, such as language loss [75] or mask losses [66]. This increases optimization difficulty and often requires additional hyperparameter tuning or annotations. Third, existing video editing models, both instructional and non-instructional, struggle with handling other diverse editing tasks, ranging from addition, and deletion to stylization. For example, LGVI [66] fails in global edits such as stylization and color change, while VidToMe [41] struggles with local edits such as adding or removing objects. These methods also struggle with input videos that contain multiple objects or when user instructions require complex reasoning.",
+        "bbox": [
+            89,
+            296,
+            485,
+            583
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "These challenges result from two limitations: First, there is a lack of multitasking fine-tuning on well-curated instructional video editing datasets that span a broad range of skills. Second, models often lack two critical capabilities needed to interpret user intentions and accurately locate concepts: multimodal reasoning to infer the intended modification from the user's instruction; and grounding language to the input video to precisely identify the region or object to be edited. For example, in Figure 1, one can effortlessly locate the girl given \"identify the little girl.\" When asked to \"add a hat to the little girl,\" we intuitively imagine the hat placed on her head from commonsense, even without seeing an actual hat.",
+        "bbox": [
+            89,
+            584,
+            485,
+            763
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address these challenges, we introduce VEGGIE, a Video Editor with Grounded Generation from Instructions. VEGGIE unifies video concept grounding and editing without relying on additional layout, mask guidance, or intermediate caption [35, 39, 42, 75, 78, 79]. Instead, we formulate the problem as end-to-end grounded generation in pixel space, using only a diffusion loss. Specifically, given a video and a text query, VEGGIE first leverages an MLLM to interpret complex instructions, generating frame-wise con",
+        "bbox": [
+            89,
+            765,
+            485,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "ditions. Unlike prior methods [43, 75] that use discrete text tokens as conditions, which disconnect the pipeline and block gradient propagation, VEGGIE employs continuous, learnable task query embeddings per frame. This enables end-to-end training and effectively captures grounded task representations for diffusion model conditioning. To handle diverse tasks and accurately interpret complex queries, we employ a curriculum learning strategy that begins by aligning MLLMs with diffusion models using massive paired instructional image editing data and then fine-tuning the model end-to-end on high-quality multitask video data to adapt video. Unlike tool-use methods [16, 35, 75], VEGGIE formulates both video grounding and instructional editing in the same video-to-video task formulation, enabling efficient handling through a unified single model. To further support end-to-end training, we introduce a novel automatic instructional video data generation pipeline that lifts high-quality instructional image editing data into the video domain using image-to-video and video evaluation tools.",
+        "bbox": [
+            511,
+            90,
+            906,
+            378
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Existing video editing benchmarks do not provide wide and uniform coverage of diverse editing skills [59, 68]. To address this gap, we contribute VEG-Bench, an instructional video editing benchmark that spans 8 editing skills: concept addition, removal, object changing, environment background changing, visual feature changing, stylization, object grounding, and reasoning segmentation. Each skill is evaluated using a dedicated suite of metrics. We assess the proposed VEGGIE alongside 6 baselines on VEG-Bench. VEGGIE demonstrates strong performance across diverse editing skills, outperforming the best instructional baseline as a versatile, all-in-one model, while other models struggle with multi-tasking. Additionally, VEGGIE excels in video object grounding and reasoning segmentation tasks, where other baselines fall short. We show further analysis of how multi-task learning enhances our framework and highlight applications such as zero-shot multimodal instructional following (Fig. 2) and few-shot in-context editing (Fig. 3). Our contributions are summarized as follows:",
+        "bbox": [
+            511,
+            381,
+            908,
+            667
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose VEGGIE, an end-to-end model that integrates an MLLM and a VidDM. VEGGIE is a versatile framework that handles diverse instructional requests for editing and grounding various video concepts. Unlike existing work that achieves multitasking via tool use, VEGGIE unifies diverse tasks in a single model, thus simplifying the training with only diffusion loss.",
+            "- We propose a data synthesis pipeline, scaling high-quality instructional video editing data for future work.",
+            "- We propose VEG-Bench, an instructional video editing benchmark that spans 8 editing skills with dedicated metrics for each skill.",
+            "- VEGGIE achieves strong performance across diverse editing skills compared with SoTA methods, and shows potentials for multimodal instruction and in-context following."
+        ],
+        "bbox": [
+            514,
+            674,
+            908,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "15148",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/e0d151b937adc7407e14b52cd875ca0bfb673a4ac4838074cdae58f7cce26412.jpg",
+        "image_caption": [
+            "Reference"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            161,
+            89,
+            243,
+            142
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/4540fb3623e1042b72cdc51d40fae52db70bda03718617114a98d44bd5cf175c.jpg",
+        "image_caption": [
+            "Transfer the style in the reference image.",
+            "Add the object in the reference image on the women."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            248,
+            90,
+            493,
+            218
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/a83a9ab1de823a8e4032e1dedd9635d6aa3ee84abd0b1589dedcb0a1c5522e76.jpg",
+        "image_caption": [
+            "Reference"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            101,
+            581,
+            142
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/22d931f8d1d6d4cd46b69a5e91dbd7bdc650d603c3c3e56bcf3baecacfd0144e.jpg",
+        "image_caption": [
+            "Figure 2. Multimodal instruction following emerges in VEGGIE, allowing for style transfer or object addition from reference images."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            586,
+            90,
+            831,
+            217
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/b5f1369b96b80acff89b8832473d68767bd364283daa8b18fad10e9742759d4e.jpg",
+        "image_caption": [
+            "Figure 3. In-context editing emerges in VEGGIE, allowing for few-shot learning of editing tasks with paired image demonstrations."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            161,
+            262,
+            831,
+            393
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            89,
+            448,
+            232,
+            463
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Instructional Video Editing Video Diffusion Models (VidDMs) [2, 4, 6, 7, 15, 26, 49, 67, 71], enable high-quality video generation across a wide range of video concepts. Building on these advances, video editing methods have emerged as tools for video design, allowing users to manipulate video concepts such as adding, removing, altering objects and style translation [20, 53, 62, 79, 85]. To enhance user experiences, instructional video editing methods [54, 70, 86] have been developed, using triples of video instructions, source videos, and target videos for training. These methods demonstrate limited performance when complex multimodal reasoning is required, as noted by previous research on instructional image editing [17]. Moreover, they struggle with diverse editing tasks, from addition and deletion to stylization. For example, LGVI [66] is primarily designed for removal tasks, while TokenFlow [20] struggles with local edits such as adding, removing, or changing objects. We address this limitation with pixel-level multitasking fine-tuning on well-curated instructional video editing datasets covering various grounding and editing skills.",
+        "bbox": [
+            89,
+            474,
+            483,
+            777
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Video Grounding and Segmentation Visual grounding requires models to connect language to its corresponding visual concept in the visual context [40, 47]. This is commonly evaluated via the language-guided semantic localization tasks, ranging from simple referring expressions in RefCOCO series [48, 77] and their generalized variant [44] that takes no-target and multi-target into account. Recently, grounded multimodal large lan",
+        "bbox": [
+            89,
+            779,
+            483,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "guage models (MLLMs) are trained for object grounding to bounding boxes [8, 51, 52, 76, 82, 87] and segmentation masks [56, 69, 81, 84] using text-image pairs with fine-grained annotations linking phrases to entities. These models unlock the potential of reasoning segmentation [11, 36], bringing language-informed reasoning into semantic segmentation. Instead of using dedicated object detection or segmentation modules, we achieve video grounding through end-to-end training with only diffusion loss.",
+        "bbox": [
+            511,
+            450,
+            906,
+            585
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Our Method: VEGGIE",
+        "text_level": 1,
+        "bbox": [
+            511,
+            599,
+            725,
+            614
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we introduce a Video Editor with Grounded Generation from Instructions (VEGGIE), a unified and versatile generative video model. It combines the complex instruction understanding and reasoning capabilities of MLLMs with the generative capacity of VidDMs. The model is trained end-to-end with diffusion loss only. VEGGIE efficiently handles diverse user inputs, including direct instructions, complex questions requiring in-depth reasoning, and multimodal conditioning. It performs various pixel-level manipulations, enabling tasks such as video concept addition, removal, changing, stylization, grounding, and reasoning segmentation based on user instructions. We elaborate on the model design (Sec. 3.1), training and inference process (Sec. 3.2), and data curation (Sec. 3.3).",
+        "bbox": [
+            511,
+            626,
+            908,
+            837
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Model Architecture",
+        "text_level": 1,
+        "bbox": [
+            511,
+            848,
+            700,
+            862
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "VEGGIE consists of four main components (see Fig. 4): (1) a multimodal large language model, (2) a set of learnable",
+        "bbox": [
+            511,
+            869,
+            906,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "15149",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/419c74734a879ff33d66c796ab67f7d5591f5073932ab3c43c6916f3395226b5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            200,
+            92,
+            225,
+            109
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/4614bdaab383a2d7be34477abb6dd43be30e9fc13b804982c124645ab6ae1578.jpg",
+        "image_caption": [
+            "Figure 4. Overview of our proposed end-to-end VEGGIE framework. Our Multimodal Large Language Model first understands input video frames and diverse user instructions, then it generates frame-wise reasoning queries that maintain per-frame editing conditions for the video diffusion model. The video diffusion model will render the MLLM-generated conditions to the pixel space for diverse tasks, including video editing, video grounding, and video reasoning segmentation with questions. We only apply diffusion loss for the whole pipeline training."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            112,
+            118,
+            883,
+            354
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "grounded task queries, (3) an alignment network (single-layer MLP) that projects the MLLM output into the condition space of the diffusion model, and (4) a video diffusion model initialized from an instructional image editing model [83]. Our model first generates latent conditions for target video frames by querying multimodal context using an MLLM, then renders these conditions at the pixel level through a video diffusion model, as detailed below.",
+        "bbox": [
+            88,
+            443,
+            485,
+            563
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "MLLM for Generating Grounded Task Guidance. As illustrated in the left of Fig. 4, given a video consisting of a sequence of frames  $V = [f_1, \\dots, f_n]$ , where  $n$  is the frame number of the given video, a user instruction/question  $I$ , our goal is to obtain the response  $\\widehat{V} = [\\widehat{f}_1, \\dots, \\widehat{f}_n]$  at pixel space that faithfully reflects user instruction about the given video. The MLLM module processes both the input video  $V$  and a user instruction  $I$  to generate a sequence of grounded task tokens per frame:  $C = [c_1, \\dots, c_n]$ , which are input and output in parallel. These tokens serve as task guidance and implicitly encode the target manipulation, such as object attributes, spatial relationships, or style transfer parameters. The MLLM ensures the model captures both explicit user instructions and implicit reasoning needs.",
+        "bbox": [
+            88,
+            565,
+            482,
+            777
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "VidDM for Rendering MLLM Guidance at Pixel Space. As illustrated in the right of Table 4, the VidDM takes the original video  $V$  and the grounded task tokens  $C$  as conditions to synthesize the target video  $\\widehat{V}$ . The original video is concatenated with the noise volume, and the task tokens are input to the cross-attention. With grounded task guidance in denoising steps, the generation process ensures that the output faithfully follows user instructions while preserving the",
+        "bbox": [
+            88,
+            779,
+            482,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "video's structure and motion dynamics. Through iterative denoising, it refines each frame while maintaining temporal consistency, applying pixel modifications coherently for a smooth and visually consistent output video  $\\widehat{V}$ .",
+        "bbox": [
+            511,
+            443,
+            906,
+            503
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Curriculum Learning from Images to Videos",
+        "text_level": 1,
+        "bbox": [
+            511,
+            515,
+            893,
+            531
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Training the model directly on video tasks presents two key challenges: (1) misalignment between MLLM and diffusion model representations, making it difficult for the diffusion model to interpret MLLM-generated task queries with limited fine-tuning data, and (2) the diffusion model's lack of multitasking capability, even for image tasks, due to insufficient training on diverse tasks. Our initial experiments also found the model collapsed when the whole pipeline was directly trained with all data. These challenges/observations underscore the need for pre-alignment between MLLM and the diffusion model to enable seamless adaptation from language-space task queries to pixel-space modifications. To this end, we adopt a two-stage curriculum learning strategy for the proposed VEGGIE framework.",
+        "bbox": [
+            511,
+            537,
+            906,
+            748
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Stage 1: Aligning Diffusion and Language Spaces. In the first stage, we align the diffusion model with the MLLM using large-scale image-level instructional editing data. The MLLM remains frozen while we update the alignment network, grounded task queries, and diffusion UNet. This process fine-tunes the diffusion model weights to align with the language space, enabling the model to interpret MLLM-generated guidance and translate user instructions into pixel-level edits while preserving the MLLM's strong ability to understand instructions and user intentions.",
+        "bbox": [
+            511,
+            750,
+            908,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "header",
+        "text": "VEGGIE: A Unified and Versatile Instructional Video Generative Model",
+        "bbox": [
+            225,
+            92,
+            761,
+            108
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "15150",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/c6400e6808829343ebe8bff1b8ea156dad0faa011c49ed265b4bc83db3d484f0.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 1. Summary of our data for training. R.: Reasoning, E.: Editing, G.: Grounding. #Img/Vid: the number of images/videos, and #Ins.: the number of instruction-image/video pairs."
+        ],
+        "table_body": "<table><tr><td>Type</td><td>Source</td><td>R.</td><td>E.</td><td>G.</td><td># Img./Vid.</td><td># Ins.</td></tr><tr><td rowspan=\"7\">Video</td><td>ROVI [66]</td><td>✓</td><td>✓</td><td>✓</td><td>4.3K</td><td>27.4K</td></tr><tr><td>VPLM [75]</td><td></td><td>✓</td><td></td><td>4.3K</td><td>5.5K</td></tr><tr><td>GroundMoRe [12]</td><td>✓</td><td></td><td>✓</td><td>1.3K</td><td>5.5K</td></tr><tr><td>RVoS [60]</td><td></td><td></td><td>✓</td><td>1.9K</td><td>6.1K</td></tr><tr><td>MeViS [14]</td><td></td><td></td><td>✓</td><td>1.5K</td><td>17.1K</td></tr><tr><td>InstructV2V [9]</td><td></td><td>✓</td><td></td><td>68.3K</td><td>68.3K</td></tr><tr><td>VEG-Edit (Ours)</td><td></td><td>✓</td><td></td><td>4.0K</td><td>6.2K</td></tr><tr><td colspan=\"6\">Total</td><td>136.1K</td></tr><tr><td rowspan=\"7\">Image</td><td>Seed-Data-Edit [18]</td><td></td><td>✓</td><td></td><td>3M</td><td>3M</td></tr><tr><td>LISA [37]</td><td>✓</td><td></td><td>✓</td><td>0.2K</td><td>1.3K</td></tr><tr><td>gRefCoCo [44]</td><td></td><td></td><td>✓</td><td>13.6K</td><td>73.4K</td></tr><tr><td>PhraseCut [65]</td><td></td><td></td><td>✓</td><td>310.8K</td><td>310.8K</td></tr><tr><td>EraseDraw [5]</td><td></td><td>✓</td><td></td><td>64.9K</td><td>42.4K</td></tr><tr><td>MagicBrush [83]</td><td></td><td>✓</td><td></td><td>9.3K</td><td>9.3K</td></tr><tr><td>SmartEdit [29]</td><td></td><td>✓</td><td></td><td>0.5K</td><td>0.9K</td></tr><tr><td colspan=\"6\">Total</td><td>3438.1K</td></tr></table>",
+        "bbox": [
+            93,
+            88,
+            478,
+            311
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Stage 2: Enhancing Temporal Consistency and Dynamics. With the MLLM and diffusion model aligned, fine-tuning diverse instructional video editing data becomes more effective for improved instruction following at pixel-space including temporal consistency, dynamic coherence, and editing faithfulness. In this stage, we fine-tune the framework with the MLLM, including the alignment network, grounded task queries, and all 3 dimensions in diffusion UNet, end-to-end with carefully curated multitasking instructional video editing data. Following prior work [22, 67], we inflated the 2D UNet from Stage 1 with temporal attention layers for video adaptation. For both stages 1 and 2, we optimize the framework with a single diffusion loss, enabling unified learning for improved instructional video editing performance while maintaining simplicity and efficiency.",
+        "bbox": [
+            88,
+            383,
+            485,
+            609
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Classifier-Free Guidance during Testing. We employ classifier-free guidance to balance quality and diversity in diffusion-generated samples. Following prior work [3, 17], we apply classifier-free guidance to instructional visual editing considering two conditions: the grounded task tokens and the original video. To obtain unconditional guidance, we set null values  $(\\varnothing)$  for both task tokens and input video. In this case, our score estimate is:",
+        "bbox": [
+            88,
+            614,
+            486,
+            734
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\tilde {e _ {\\theta}} (z _ {t}, c _ {T}, c _ {V}) = e _ {\\theta} (z _ {t}, \\varnothing , \\varnothing) \\\\ + g _ {T} \\cdot \\left(e _ {\\theta} \\left(z _ {t}, c _ {V}, c _ {T}\\right) - e _ {\\theta} \\left(z _ {t}, c _ {V}, \\varnothing\\right)\\right) \\\\ + g _ {V} \\cdot \\left(e _ {\\theta} \\left(z _ {t}, c _ {V}, \\varnothing\\right) - e _ {\\theta} \\left(z _ {t}, \\varnothing , \\varnothing\\right)\\right), \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            99,
+            750,
+            472,
+            805
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\theta$  represents the model parameters,  $C_T$  and  $C_V$  denote the task tokens and video conditions,  $\\varnothing$  is the null value,  $z_t$  is the noised latent at timestamp  $t$ , and  $g_T$  and  $g_V$  are the task guidance and video guidance scales, respectively. More training details are included later in Appendix.",
+        "bbox": [
+            88,
+            824,
+            483,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3. Data Curation Pipeline",
+        "text_level": 1,
+        "bbox": [
+            513,
+            90,
+            728,
+            107
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Existing video editing models, both instructional and non-instructional, struggle with diverse editing skills due to the lack of high-quality multitasking fine-tuning data. In this section, we introduce our data curation strategy to support VEGGIE in achieving versatile video editing skills. As listed in Tab. 1, we collect 3.4M image and 133.9K video data from diverse sources to support our VEGGIE curriculum learning as discussed in Sec. 3.2. We create our training dataset from two sources: (1) collecting existing image and video data and converting it into an instructional editing format, and (2) synthesizing new instructional video editing samples using existing datasets and generative models.",
+        "bbox": [
+            509,
+            112,
+            906,
+            294
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Collecting Diverse Multitask Image and Video Data. We bring together instructional editing data from both image (Seed-Data-Edit [18], MagicBrush [83], EraseDraw [5]) and video (InstructV2V [9], VPLM [75]) sources. These datasets provide pairs of original and edited visual contents with user instructions. The tasks include adding, removing, and changing objects, stylizing, and performing global/local edits. Beyond editing datasets, we incorporate segmentation data at both the image level (gRefCoCo [44] and Phrase-Cut [65]) and the video level (RVoS and MeViS). These segmentation tasks are reformulated as color-filling challenges, which guide the model in learning referring grounding (i.e., understanding which object or region to edit) and strengthen its conceptual learning. To further unlock complex instruction understanding via MLLM, we include data that requires more advanced reasoning and implicit referencing. Specifically, we include: reasoning segmentation (LISA [37]), reasoning editing (SmartEdit [29]), interactive video inpainting (LGVI [66]), and motion-grounded video reasoning (GroundMoRe [12]). These tasks help VEGGIE learn implicit references and reasoning.",
+        "bbox": [
+            509,
+            295,
+            908,
+            613
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Synthesizing Instructional Video Editing Data via Image-to-Video Animation. Recent methods [3, 54] generate synthetic instructional video-editing data by first creating text instructions with LLM, then getting edited videos via T2V models and prompt-to-prompt editing [23]. While these methods adapt image-based editing pipelines [9] to videos, the generated data suffer from temporal-consistency issues. To address this gap, we propose a novel image-to-video animation strategy that leverages the abundance of high-quality image-level instructional editing datasets [64, 83], which provide well-annotated instructions, paired edited images, and well-organized editing skill categories. As illustrated in Fig. 5, given an original image  $I$ , an edited image  $\\bar{I}$ , and an instruction from an instructional image editing dataset [64], our approach involves three key steps. First, we use an offline MLLM [61, 72] to generate an image caption and an animation prompt that describes plausible motion within the image. Next, an image-to-video (I2V) model animates the image into a video  $V$ . Finally,",
+        "bbox": [
+            509,
+            613,
+            908,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "15151",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/519aa3cbb726c3f46600125ec3400403ec4db3182b9494a57667d8cbec97125a.jpg",
+        "image_caption": [
+            "Figure 5. Our data generation pipeline for synthetic instructional video editing data. It injects dynamics into well-constructed instructional image editing datasets via the Image-to-Video (I2V) Model, and generates paired video data for instruction editing."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            87,
+            450,
+            356
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "we generate the corresponding edited video  $\\bar{V}$  using a first-frame-conditioned video editing model [35], leveraging  $\\bar{I}$  as a strong prior to ensure consistent edits across frames. Finally, to ensure data quality, we evaluate each original-edited video pair with automatic video quality evaluation metrics [30], which assess the generated videos from diverse dimensions, e.g., motion smoothness, image quality, and background consistency. This pipeline transforms carefully curated image-based datasets into instructional video-editing resources while preserving the precision of the original edits. As a result, our data method expands the availability of high-quality synthetic video-editing data, supporting a wider range of editing tasks in our end-to-end unified framework. More details on data generation, prompting, examples, and pre/post-processing are in the Appendix.",
+        "bbox": [
+            88,
+            430,
+            483,
+            657
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            672,
+            223,
+            690
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We first introduce the VEG-Bench Benchmark and then demonstrate the superiority of VEGGIE across diverse video instructional editing skills. More experiments, visualization, and implementation details are in the Appendix.",
+        "bbox": [
+            89,
+            699,
+            485,
+            761
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. VEG-Bench and Metrics",
+        "text_level": 1,
+        "bbox": [
+            89,
+            772,
+            313,
+            787
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As no existing benchmark is designed for fine-grained instructional video editing skills, we manually collect and annotate VEG-Bench, containing 132 video-instruction pairs that balanced cover 8 different video generative skills (15-20 for each). Beyond standard metrics, including text-to-video alignment (CLIP-Text [55]), video smoothness (CLIP-F [55]), and image quality (MUSIQ [32]), we also",
+        "bbox": [
+            89,
+            794,
+            485,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "first introduce MLLM-as-a-Judge to give a holistic evaluation score according to the given original video, edited video, and user instruction. It is achieved by prompting GPT4o [21] to evaluate whether the requested semantic change has been fulfilled, using a scale from 1 to 10. For addition and removal, we also introduce an object detector (GroundingDiNo [45]) to detect if the object is added/removed faithfully. For grounding and reasoning segmentation, we following video grounding tasks [12, 14, 33, 58] and adopt the Jaccard index  $(\\mathcal{I})$  [31], F-measure  $(\\mathcal{F})$  [13], and their mean  $(\\mathcal{J} \\& \\mathcal{F})$ . We also compute SSIM between the generated video and the original video masked with GT masks. More evaluation/metrics details are included in Appendix.",
+        "bbox": [
+            511,
+            90,
+            908,
+            287
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Experimental Results",
+        "text_level": 1,
+        "bbox": [
+            511,
+            300,
+            718,
+            316
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Instructional Video Editing over Diverse Skills. As shown in Tab. 2, we evaluate 7 different models on VEG-Bench across 8 distinct editing skills. Overall, VEGGIE demonstrates the best performance among instructional video editing models. Compared to VEGGIE, non-instructional models often struggle with concept removal and addition. This limitation arises because these models rely on attention control or additional conditions (e.g., depth maps) that impose strong priors, constraining the model and making object addition or removal challenging. We also observe that InsV2V achieves high scores in quality and smoothness metrics, but underperforms in alignment and MLLM judgment, which demand faithful semantic changes. Qualitative examples in Fig. 6 illustrate that InsV2V often makes minimal changes to the input video, resulting in high video quality but unfaithful outputs. In contrast, VEGGIE strikes a better balance, delivering both high-quality visuals and accurate semantic alignment with the intended edits.",
+        "bbox": [
+            511,
+            323,
+            908,
+            595
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Can Multi-Task Help Each Other? To test the previous hypothesis, we train our model on the VPLM [75] dataset, which includes paired grounding and removal tasks (approximately 5.5K samples for each task). We focus on these tasks as representative examples due to their straightforward evaluation against ground truth. As shown in Table 3, multitask training yields a lower FVD score and a higher SSIM score, demonstrating that learning to locate and remove a video concept can mutually reinforce performance. We show an example in Fig. 7. However, this conclusion only holds with a balanced data combination. We also observe that an excessive amount of grounding data can introduce more artifacts and negatively impact visual editing skills.",
+        "bbox": [
+            511,
+            597,
+            908,
+            792
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Emergent Zero-shot Multimodal Instruction Following. We also highlight the emergent behavior of VEGGIE on multi-modal instruction following, even without dedicated training data for this specific editing instruction. Notably, VEGGIE demonstrates the ability to perform zero-shot multimodal instructional video editing. As illustrated in Fig. 2, VEGGIE can transfer styles or add objects from a reference",
+        "bbox": [
+            511,
+            795,
+            908,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "15152",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/448541813648206c696b8060be6836952b32cd5e03312d900e585ca11af2f430.jpg",
+        "image_caption": [
+            "[Addition] Please add a ball in the given video frames."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            99,
+            367,
+            339
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/67afd19caa3fbcb6b392c5ad751d2b0f2e3161441709a94940152b6d86723f7e.jpg",
+        "image_caption": [
+            "[Removal] Please remove the man in black in given video frames."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            370,
+            99,
+            625,
+            339
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/373ec45a7f7d566ba4d3a78d8e06b09d7bad2b84179f4cdf62f91a27bbbf530a.jpg",
+        "image_caption": [
+            "[Swap] Replace golden building with a white mountain."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            625,
+            99,
+            880,
+            339
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/41198c72d6faf22421eb58cefcc01dfd03a7dbb2d35994a01bd56dc226535d45.jpg",
+        "image_caption": [
+            "[Environment] Make it on the beach."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            349,
+            367,
+            589
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2aaed0e07fec7c8843e49de92130946b5b89bbd966dcb8e209bd72feffe8789e.jpg",
+        "image_caption": [
+            "[Color] Make the swan white."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            370,
+            349,
+            625,
+            589
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/a238051c85810f38dffc43dcef81eb1d2021aded79c1a77a3508a3bafc592ab6.jpg",
+        "image_caption": [
+            "[Texture] Make the rhinocero furry."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            625,
+            349,
+            880,
+            589
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/a41b7f9d6e96f7f28878ef89d70835e2d12c0c4a44228a64db4c7f7f34d6c0d5.jpg",
+        "image_caption": [
+            "[Style] Make it chinese ink style.",
+            "Figure 6. Qualitative comparison of editing results across 8 different abilities (splitting visual features into color and texture). We provide zoom-in details for a more detailed comparison. Best viewed in color. More in Appendix."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            606,
+            367,
+            845
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/6fecf891dfe80f439d7ac4732ec66bb78527fab1d6f6d18a6235978f3c780c18.jpg",
+        "image_caption": [
+            "[Grounding] Could you locate the knife"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            370,
+            599,
+            625,
+            845
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/b7c113e68f1b140f6b2fd89134382a5e2e6aa73949f6da185815fa51ef01fdde.jpg",
+        "image_caption": [
+            "[Reasoning] What can be used for heating food?"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            625,
+            599,
+            880,
+            845
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "15153",
+        "bbox": [
+            480,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/24831cb68b0c27cf147e918fb58dc3a60803aa169edf93235ef0f987dc167bca.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods\nGenerator</td><td colspan=\"3\">Non-Instructional Editing Model</td><td colspan=\"4\">Instructional Editing Model</td></tr><tr><td>VidToMe [41] (SD1.5)</td><td>TokenFlow [20] (SD2.1)</td><td>Flatten [10] (SD2.1)</td><td>InstructDiff [19] (SD1.5)</td><td>LGVI [66] (SD1.5)</td><td>InsV2V [9] (SD1.5)</td><td>VEGGIE (Ours) (SD1.5)</td></tr><tr><td colspan=\"8\">Concept Addition</td></tr><tr><td>MLLM-Judge (↑)</td><td>5.00</td><td>5.80</td><td>6.62</td><td>7.26</td><td>2.73</td><td>5.69</td><td>7.44</td></tr><tr><td>Alignment (↑)</td><td>27.80</td><td>29.30</td><td>28.22</td><td>28.10</td><td>25.06</td><td>28.27</td><td>29.27</td></tr><tr><td>Smoothness (↑)</td><td>96.79</td><td>97.26</td><td>95.74</td><td>93.66</td><td>96.09</td><td>96.94</td><td>94.93</td></tr><tr><td>Quality (↑)</td><td>62.81</td><td>65.62</td><td>55.45</td><td>54.08</td><td>41.59</td><td>56.24</td><td>61.31</td></tr><tr><td>Detection (↑)</td><td>47.98</td><td>49.53</td><td>49.74</td><td>55.36</td><td>14.42</td><td>48.01</td><td>57.96</td></tr><tr><td colspan=\"8\">Concept Removal</td></tr><tr><td>MLLM-Judge (↑)</td><td>2.60</td><td>3.73</td><td>4.46</td><td>6.12</td><td>6.59</td><td>2.78</td><td>5.07</td></tr><tr><td>Alignment (↑)</td><td>75.01</td><td>75.99</td><td>78.40</td><td>75.51</td><td>75.67</td><td>74.41</td><td>75.63</td></tr><tr><td>Smoothness (↑)</td><td>96.13</td><td>96.47</td><td>95.82</td><td>91.83</td><td>97.03</td><td>96.99</td><td>95.04</td></tr><tr><td>Quality (↑)</td><td>66.32</td><td>71.52</td><td>50.77</td><td>55.08</td><td>42.31</td><td>58.79</td><td>50.99</td></tr><tr><td>Detection (↑)</td><td>34.31</td><td>55.16</td><td>70.91</td><td>64.81</td><td>78.40</td><td>25.64</td><td>70.22</td></tr><tr><td colspan=\"8\">Object Changing</td></tr><tr><td>MLLM-Judge (↑)</td><td>5.00</td><td>6.53</td><td>7.37</td><td>7.00</td><td>2.06</td><td>6.60</td><td>6.63</td></tr><tr><td>Alignment (↑)</td><td>25.69</td><td>28.76</td><td>27.06</td><td>27.36</td><td>22.17</td><td>26.60</td><td>27.77</td></tr><tr><td>Smoothness (↑)</td><td>96.23</td><td>97.21</td><td>96.13</td><td>92.07</td><td>95.66</td><td>96.74</td><td>95.44</td></tr><tr><td>Quality (↑)</td><td>64.06</td><td>69.97</td><td>59.37</td><td>55.01</td><td>38.20</td><td>60.90</td><td>58.15</td></tr><tr><td colspan=\"8\">Environment &amp; Background Changing</td></tr><tr><td>MLLM-Judge (↑)</td><td>5.81</td><td>7.35</td><td>7.37</td><td>6.05</td><td>2.37</td><td>6.60</td><td>7.18</td></tr><tr><td>Alignment (↑)</td><td>28.17</td><td>30.00</td><td>30.04</td><td>28.03</td><td>21.94</td><td>28.27</td><td>29.15</td></tr><tr><td>Smoothness (↑)</td><td>95.76</td><td>96.96</td><td>95.90</td><td>89.85</td><td>95.66</td><td>96.03</td><td>94.58</td></tr><tr><td>Quality (↑)</td><td>61.95</td><td>67.06</td><td>54.58</td><td>53.06</td><td>38.97</td><td>54.94</td><td>54.25</td></tr><tr><td colspan=\"8\">Visual Feature Changing (Color &amp; Texture)</td></tr><tr><td>MLLM-Judge (↑)</td><td>5.86</td><td>6.85</td><td>6.60</td><td>6.43</td><td>2.14</td><td>7.53</td><td>7.33</td></tr><tr><td>Alignment (↑)</td><td>27.99</td><td>29.25</td><td>29.46</td><td>27.54</td><td>23.18</td><td>28.88</td><td>28.69</td></tr><tr><td>Smoothness (↑)</td><td>95.93</td><td>97.10</td><td>95.83</td><td>91.71</td><td>94.75</td><td>96.66</td><td>94.52</td></tr><tr><td>Quality (↑)</td><td>65.80</td><td>69.31</td><td>53.32</td><td>58.29</td><td>36.27</td><td>59.36</td><td>57.91</td></tr><tr><td colspan=\"8\">Stylization</td></tr><tr><td>MLLM-Judge (↑)</td><td>7.23</td><td>7.62</td><td>8.31</td><td>7.41</td><td>3.71</td><td>8.07</td><td>8.26</td></tr><tr><td>Alignment (↑)</td><td>29.84</td><td>30.25</td><td>29.00</td><td>27.74</td><td>22.80</td><td>29.14</td><td>29.38</td></tr><tr><td>Smoothness (↑)</td><td>96.31</td><td>97.23</td><td>96.71</td><td>88.97</td><td>95.62</td><td>96.50</td><td>95.69</td></tr><tr><td>Quality (↑)</td><td>64.05</td><td>68.22</td><td>53.18</td><td>54.15</td><td>35.76</td><td>62.59</td><td>57.00</td></tr><tr><td colspan=\"8\">Object Grounding</td></tr><tr><td>SSIM (↑)</td><td>40.47</td><td>50.46</td><td>47.21</td><td>37.98</td><td>66.84</td><td>49.65</td><td>70.90</td></tr><tr><td>Jaccard Index J (↑)</td><td>13.85</td><td>19.29</td><td>25.62</td><td>19.88</td><td>1.52</td><td>13.89</td><td>37.74</td></tr><tr><td>F-measure F (↑)</td><td>15.50</td><td>16.86</td><td>17.60</td><td>12.81</td><td>3.07</td><td>17.37</td><td>21.83</td></tr><tr><td colspan=\"8\">Reasoning Segmentation</td></tr><tr><td>SSIM (↑)</td><td>-</td><td>-</td><td>-</td><td>32.39</td><td>44.47</td><td>59.86</td><td>68.41</td></tr><tr><td>Jaccard Index J (↑)</td><td>-</td><td>-</td><td>-</td><td>14.02</td><td>10.12</td><td>16.89</td><td>22.53</td></tr><tr><td>F-measure F (↑)</td><td>-</td><td>-</td><td>-</td><td>8.07</td><td>9.06</td><td>10.45</td><td>15.97</td></tr><tr><td>Avg. Ranking</td><td>2.61</td><td>1.41</td><td>1.96</td><td>3.00</td><td>3.21</td><td>2.00</td><td>1.78</td></tr></table>",
+        "bbox": [
+            148,
+            88,
+            841,
+            512
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/9b625bcc9addcaa022aa6854cf74a6df0f31f087476d9e9b0190a15cb855a15d.jpg",
+        "table_caption": [
+            "Table 2. Comparison of video editing task with instructional / non-instructional models on VEG-Bench. -: the task is not capable of non-instructional models. We gray out numbers of non-instructional models that are in different categories."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Settings</td><td>Removal (FVD ↓)</td><td>Grounding (SSIM ↑)</td></tr><tr><td>Grd.-only</td><td>-</td><td>52.34</td></tr><tr><td>Rmv.-only</td><td>1098.52</td><td>-</td></tr><tr><td>Mixed</td><td>987.80</td><td>55.21</td></tr></table>",
+        "bbox": [
+            117,
+            556,
+            464,
+            621
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3. An ablation study on whether multi-task learning provides transferable benefits that enhance performance across tasks. We focus on removal and grounding tasks as representative examples.",
+        "bbox": [
+            89,
+            632,
+            482,
+            674
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/16dd7320589fd8681a64e1346642e6554bace0aed5cea65f14131a25d7187071.jpg",
+        "image_caption": [
+            "Figure 7. Comparison between single- and multi-skill models with different data training. We find tasks can help each other."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            686,
+            460,
+            828
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "image into the input video based on instructions.",
+        "bbox": [
+            511,
+            559,
+            834,
+            573
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Emergent Few-shot In-Context Editing. As shown in Fig. 3, VEGGIE can effectively utilize a few example image pairs to transfer the intended editing changes seamlessly to the input video. We observe that VEGGIE exhibits in-context learning for image editing without the need for language instructions. Instead, it uses image pairs as examples to infer and apply the desired editing intention directly.",
+        "bbox": [
+            511,
+            574,
+            906,
+            680
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            511,
+            694,
+            633,
+            709
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We present VEGGIE, a unified end-to-end model for instructional video editing that handles diverse pixel-level tasks. VEGGIE leverages MLLM for robust instruction understanding and employs a video diffusion model to execute pixel-level edits. Our framework uses a single diffusion loss for end-to-end optimization across varied tasks/skills. We also introduce a novel synthetic data generation pipeline and VEG-Bench, a benchmark that assesses a broad range of editing skills. Our VEGGIE outperforms previous methods as a versatile, all-in-one solution. We hope our model, data, and benchmark to advance research on instructional generative video models.",
+        "bbox": [
+            511,
+            719,
+            908,
+            887
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "15154",
+        "bbox": [
+            480,
+            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] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 2",
+            "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3",
+            "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 5",
+            "[4] 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. OpenAI, https://openai.com/research/video-generation-models-as-world-simulators, 2024.2.3",
+            "[5] Alper Canberk, Maksym Bondarenko, Ege Ozguroglu, Ruoshi Liu, and Carl Vondrick. Erasedraw: Learning to insert objects by erasing them from images. arXiv preprint arXiv:2409.00522, 2024. 5",
+            "[6] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23206-23217, 2023. 3",
+            "[7] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stablevideo: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 3",
+            "[8] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3",
+            "[9] Jiaxin Cheng, Tianjun Xiao, and Tong He. Consistent video-to-video transfer using synthetic dataset. In The Twelfth International Conference on Learning Representations, 2024. 5, 8, 1, 2",
+            "[10] Yuren Cong, Mengmeng Xu, christian simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: optical FLOW-guided ATTENtion for consistent text-to-video editing. In The Twelfth International Conference on Learning Representations, 2024. 8, 1",
+            "[11] Andong Deng, Tongjia Chen, Shoubin Yu, Taojiannan Yang, Lincoln Spencer, Yapeng Tian, Ajmal Saeed Mian, Mohit Bansal, and Chen Chen. Motion-grounded video reasoning: Understanding and perceiving motion at pixel level. arXiv preprint arXiv:2411.09921, 2024. 3",
+            "[12] Andong Deng, Tongjia Chen, Shoubin Yu, Taojiannan Yang, Lincoln Spencer, Yapeng Tian, Ajmal Saeed Mian, Bansal Mohit, and Chen. Chen. Motion-grounded video reasoning: Understanding and perceiving motion at pixel level. 2024. 5, 6, 2"
+        ],
+        "bbox": [
+            93,
+            114,
+            485,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[13] Lee R Dice. Measures of the amount of ecologic association between species. Ecology, 26(3):297-302, 1945. 6",
+            "[14] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 5, 6",
+            "[15] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In CVPR, pages 7346-7356, 2023. 3",
+            "[16] Hao Fei, Shengqiong Wu, Hanwang Zhang, Tat-Seng Chua, and Shuicheng Yan. Vitron: A unified pixel-level vision llm for understanding, generating, segmenting, editing, 2024. 2",
+            "[17] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 2, 3, 5",
+            "[18] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. Seed-data-edit technical report: A hybrid dataset for instructional image editing. arXiv preprint arXiv:2405.04007, 2024. 5",
+            "[19] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024. 8, 1, 2",
+            "[20] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 2, 3, 8, 1",
+            "[21] gpt 4o. https://openai.com/index/hello-gpt-4o/.2024.6",
+            "[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 5, 1",
+            "[23] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022.5",
+            "[24] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2",
+            "[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 2022. 2",
+            "[26] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In The Eleventh International Conference on Learning Representations, 2023. 3",
+            "[27] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 1"
+        ],
+        "bbox": [
+            516,
+            92,
+            906,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "15155",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[28] Jiahao Hu, Tianxiong Zhong, Xuebo Wang, Boyuan Jiang, Xingye Tian, Fei Yang, Pengfei Wan, and Di Zhang. Vivid-10m: A dataset and baseline for versatile and interactive video local editing. arXiv preprint arXiv:2411.15260, 2024. 2",
+            "[29] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024. 2, 5",
+            "[30] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6, 2",
+            "[31] Paul Jaccard. The distribution of the flora in the alpine zone. 1. New phytologist, 11(2):37-50, 1912. 6",
+            "[32] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6",
+            "[33] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 6",
+            "[34] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 2",
+            "[35] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv preprint arXiv:2403.14468, 2024. 2, 6",
+            "[36] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 3",
+            "[37] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 5",
+            "[38] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1",
+            "[39] Jialu Li, Shoubin Yu, Han Lin, Jaemin Cho, Jaehong Yoon, and Mohit Bansal. Training-free guidance in text-to-video generation via multimodal planning and structured noise initialization. arXiv preprint arXiv:2504.08641, 2025. 2",
+            "[40] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language"
+        ],
+        "bbox": [
+            91,
+            90,
+            485,
+            901
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 3",
+            "[41] Xirui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidthome: Video token merging for zero-shot video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7486-7495, 2024. 2, 8, 1",
+            "[42] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 2",
+            "[43] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 2",
+            "[44] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23592-23601, 2023. 3, 5",
+            "[45] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6",
+            "[46] Shaoteng Liu, Tianyu Wang, Jui-Hsien Wang, Qing Liu, Zhifei Zhang, Joon-Young Lee, Yijun Li, Bei Yu, Zhe Lin, Soo Ye Kim, et al. Generative video propagation. arXiv preprint arXiv:2412.19761, 2024. 2",
+            "[47] Ziqiao Ma, Jiayi Pan, and Joyce Chai. World-to-words: Grounded open vocabulary acquisition through fast mapping in vision-language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 524–544, 2023. 3",
+            "[48] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 3",
+            "[49] Kangfu Mei and Vishal Patel. Vidm: Video implicit diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 9117-9125, 2023. 3",
+            "[50] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, Mubarak Shah, and Ajmal Mian. Towards temporally consistent referring video object segmentation. https://arxiv.org/abs/2403.19407, 2024. 2",
+            "[51] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. Grounding multimodal large language models to the world. In The Twelfth International Conference on Learning Representations, 2024. 3",
+            "[52] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, et al. Detgpt: Detect what you need via reasoning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14172-14189, 2023. 3",
+            "[53] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing"
+        ],
+        "bbox": [
+            516,
+            90,
+            906,
+            901
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "15156",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            " attentions for zero-shot text-based video editing. In ICCV, pages 15932-15942, 2023. 2, 3",
+            "[54] Bosheng Qin, Juncheng Li, Siliang Tang, Tat-Seng Chua, and Yueting Zhuang. Instructvid2vid: Controllable video editing with natural language instructions. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2, 3, 5",
+            "[55] 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. 6",
+            "[56] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3",
+            "[57] 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",
+            "[58] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 6",
+            "[59] Shangkun Sun, Xiaoyu Liang, Songlin Fan, Wenxu Gao, and Wei Gao. Ve-bench: Subjective-aligned benchmark suite for text-driven video editing quality assessment. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2",
+            "[60] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019. 5",
+            "[61] 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. arXiv preprint arXiv:2409.12191, 2024. 5, 2",
+            "[62] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 2, 3",
+            "[63] Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal IIm as an agent for unified image generation and editing. arXiv preprint arXiv:2407.05600, 2024. 2",
+            "[64] Cong Wei, Zheyang Xiong, Weiming Ren, Xinrun Du, Ge Zhang, and Wenhu Chen. Omniedit: Building image edit"
+        ],
+        "bbox": [
+            91,
+            90,
+            483,
+            900
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "ing generalist models through specialist supervision. arXiv preprint arXiv:2411.07199, 2024. 5, 2",
+            "[65] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 5",
+            "[66] Jianzong Wu, Xiangtai Li, Chenyang Si, Shangchen Zhou, Jingkang Yang, Jiangning Zhang, Yining Li, Kai Chen, Yunhai Tong, Ziwei Liu, et al. Towards language-driven video inpainting via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12501-12511, 2024. 2, 3, 5, 8, 1",
+            "[67] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7623-7633, 2023. 3, 5",
+            "[68] Jay Zhangjie Wu, Xiuyu Li, Difei Gao, Zhen Dong, Jinbin Bai, Aishani Singh, Xiaoyu Xiang, Youzeng Li, Zuwei Huang, Yuanxi Sun, et al. Cvpr 2023 text guided video editing competition. arXiv preprint arXiv:2310.16003, 2023. 2",
+            "[69] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3",
+            "[70] Zhen Xing, Qi Dai, Zihao Zhang, Hui Zhang, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Vidiff: Translating videos via multi-modal instructions with diffusion models. arXiv preprint arXiv:2311.18837, 2023. 3",
+            "[71] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3",
+            "[72] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 5",
+            "[73] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1",
+            "[74] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2"
+        ],
+        "bbox": [
+            516,
+            90,
+            906,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "15157",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[75] Jaehong Yoon, Shoubin Yu, and Mohit Bansal. Raccoon: Remove, add, and change video content with auto-generated narratives. arXiv preprint arXiv:2405.18406, 2024. 2, 5, 6",
+            "[76] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. In The Twelfth International Conference on Learning Representations, 2023. 3",
+            "[77] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 3",
+            "[78] Shoubin Yu, Jacob Zhiyuan Fang, Jian Zheng, Gunnar Sigurdsson, Vicente Ordonez, Robinson Piramuthu, and Mohit Bansal. Zero-shot controllable image-to-video animation via motion decomposition. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3332-3341, 2024. 2",
+            "[79] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2, 3",
+            "[80] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975-11986, 2023. 1",
+            "[81] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Lei Zhang, Chunyuan Li, et al. Llava-grounding: Grounded visual chat with large multimodal models. arXiv preprint arXiv:2312.02949, 2023. 3",
+            "[82] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferret-v2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 3",
+            "[83] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 4, 5, 1",
+            "[84] Yichi Zhang, Ziqiao Ma, Xiaofeng Gao, Suhaila Shakiah, Qiaozi Gao, and Joyce Chai. Groundhog: Grounding large language models to holistic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3",
+            "[85] 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. arXiv preprint arXiv:2312.03816, 2023. 2, 3",
+            "[86] Zhenghao Zhang, Zuozhuo Dai, Long Qin, and Weizhi Wang. Effived: Efficient video editing via text-instruction diffusion models. arXiv preprint arXiv:2403.11568, 2024. 2, 3",
+            "[87] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. 3"
+        ],
+        "bbox": [
+            91,
+            92,
+            483,
+            897
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "15158",
+        "bbox": [
+            480,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 11
+    }
+]

2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/d727aeb7-1f89-498a-ac85-9bcc05bddf8b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cb04cfcd242d4fca0335d8c3779b2a738b08925e46803f0c40b19d7eec2207ab
+size 29056000

2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/full.md

@@ -0,0 +1,278 @@
+# VEGGIE: Instructional Editing and Reasoning Video Concepts with Grounded Generation
+
+Shoubin Yu $^{1,3*}$  Difan Liu $^{1*}$  Ziqiao Ma $^{1,2*}$  Yicong Hong $^{1}$  Yang Zhou $^{1}$  Hao Tan $^{1}$  Joyce Chai $^{2}$  Mohit Bansal $^{3}$
+
+$^{1}$ Adobe Research  $^{2}$ University of Michigan  $^{3}$ UNC Chapel Hill
+
+https://veggie-gen.github.io/
+
+![](images/6de461b48922321aa8d110e2c86339f81fccd8681d7979d3daf8439499923d75.jpg)  
+Figure 1. We propose VEGGIE, a unified and versatile video generative model that handles various tasks for both video concept grounding and editing according to user instructions. With VEGGIE, users can locate, add, delete, and change concepts in a given video through diverse instruction formats (direct referring instruction or reasoning-demanding questions). Users can also edit videos with multimodal instruction empowered by MLLM, enabling applications like video editing from a reference image.
+
+# Abstract
+
+While recent video diffusion models enable video editing, unifying diverse instructional editing tasks (e.g., add, remove, modify) under a single framework remains a significant challenge. In this paper, we introduce VEGGIE, a Video Editor with Grounded Generation from Instructions, a simple end-to-end framework that unifies video concept editing, grounding, and reasoning based on diverse user instructions. Specifically, given a video and text query, VEGGIE first utilizes an MLLM to interpret user intentions in instructions and ground them to the video contexts, generating framespecific grounded task queries for pixel-space responses. A diffusion model then renders these plans and generates edited videos that align with user intent. To support diverse
+
+tasks and complex instructions, we employ a curriculum learning strategy: first aligning the MLLM and video diffusion model with large-scale instructional image editing data, followed by end-to-end fine-tuning on high-quality multitask video data. Additionally, we introduce a novel data synthesis pipeline to generate paired instructional video editing data for model training. It transforms static image data into diverse, high-quality video editing samples by leveraging Image-to-Video models to inject dynamics. VEGGIE shows strong performance in instructional video editing with different editing skills, outperforming the best instructional baseline as a versatile model, while other models struggle with multi-tasking. VEGGIE also excels in video object grounding and reasoning segmentation, where other baselines fail. We further reveal how the multiple tasks help each other and highlight promising applications like zero-shot multimodal instructional and in-context video editing.
+
+# 1. Introduction
+
+Building on the advances in Video Diffusion Models (VIDDMs) [2, 4, 24, 25, 74], video editing methods have emerged as video design tools, allowing users to manipulate video concepts such as adding, removing, altering objects and style translation [20, 53, 62, 79, 85]. To enhance user experiences, instructional video editing methods [54, 86] have been developed, using triples of text prompts, source videos, and target videos for training. Due to their limited performance in understanding user intent and multimodal semantics [17], several methods have incorporated multimodal large language models (MLLMs) to handle complex instructions/reasoning [17, 29, 75].
+
+However, existing methods fall short of the goal of a simple, versatile video concept editor, facing three primary challenges. First, most methods are not end-to-end, requiring intermediate layout/mask/human or model caption guidance [35, 53, 63, 75], which adds workload on users and disrupts a seamless editing experience. Second, existing pipelines connecting MLLMs to VidDMs require multiple training objectives beyond simple pixel-space diffusion loss, such as language loss [75] or mask losses [66]. This increases optimization difficulty and often requires additional hyperparameter tuning or annotations. Third, existing video editing models, both instructional and non-instructional, struggle with handling other diverse editing tasks, ranging from addition, and deletion to stylization. For example, LGVI [66] fails in global edits such as stylization and color change, while VidToMe [41] struggles with local edits such as adding or removing objects. These methods also struggle with input videos that contain multiple objects or when user instructions require complex reasoning.
+
+These challenges result from two limitations: First, there is a lack of multitasking fine-tuning on well-curated instructional video editing datasets that span a broad range of skills. Second, models often lack two critical capabilities needed to interpret user intentions and accurately locate concepts: multimodal reasoning to infer the intended modification from the user's instruction; and grounding language to the input video to precisely identify the region or object to be edited. For example, in Figure 1, one can effortlessly locate the girl given "identify the little girl." When asked to "add a hat to the little girl," we intuitively imagine the hat placed on her head from commonsense, even without seeing an actual hat.
+
+To address these challenges, we introduce VEGGIE, a Video Editor with Grounded Generation from Instructions. VEGGIE unifies video concept grounding and editing without relying on additional layout, mask guidance, or intermediate caption [35, 39, 42, 75, 78, 79]. Instead, we formulate the problem as end-to-end grounded generation in pixel space, using only a diffusion loss. Specifically, given a video and a text query, VEGGIE first leverages an MLLM to interpret complex instructions, generating frame-wise con
+
+ditions. Unlike prior methods [43, 75] that use discrete text tokens as conditions, which disconnect the pipeline and block gradient propagation, VEGGIE employs continuous, learnable task query embeddings per frame. This enables end-to-end training and effectively captures grounded task representations for diffusion model conditioning. To handle diverse tasks and accurately interpret complex queries, we employ a curriculum learning strategy that begins by aligning MLLMs with diffusion models using massive paired instructional image editing data and then fine-tuning the model end-to-end on high-quality multitask video data to adapt video. Unlike tool-use methods [16, 35, 75], VEGGIE formulates both video grounding and instructional editing in the same video-to-video task formulation, enabling efficient handling through a unified single model. To further support end-to-end training, we introduce a novel automatic instructional video data generation pipeline that lifts high-quality instructional image editing data into the video domain using image-to-video and video evaluation tools.
+
+Existing video editing benchmarks do not provide wide and uniform coverage of diverse editing skills [59, 68]. To address this gap, we contribute VEG-Bench, an instructional video editing benchmark that spans 8 editing skills: concept addition, removal, object changing, environment background changing, visual feature changing, stylization, object grounding, and reasoning segmentation. Each skill is evaluated using a dedicated suite of metrics. We assess the proposed VEGGIE alongside 6 baselines on VEG-Bench. VEGGIE demonstrates strong performance across diverse editing skills, outperforming the best instructional baseline as a versatile, all-in-one model, while other models struggle with multi-tasking. Additionally, VEGGIE excels in video object grounding and reasoning segmentation tasks, where other baselines fall short. We show further analysis of how multi-task learning enhances our framework and highlight applications such as zero-shot multimodal instructional following (Fig. 2) and few-shot in-context editing (Fig. 3). Our contributions are summarized as follows:
+
+- We propose VEGGIE, an end-to-end model that integrates an MLLM and a VidDM. VEGGIE is a versatile framework that handles diverse instructional requests for editing and grounding various video concepts. Unlike existing work that achieves multitasking via tool use, VEGGIE unifies diverse tasks in a single model, thus simplifying the training with only diffusion loss.  
+- We propose a data synthesis pipeline, scaling high-quality instructional video editing data for future work.  
+- We propose VEG-Bench, an instructional video editing benchmark that spans 8 editing skills with dedicated metrics for each skill.  
+- VEGGIE achieves strong performance across diverse editing skills compared with SoTA methods, and shows potentials for multimodal instruction and in-context following.
+
+![](images/e0d151b937adc7407e14b52cd875ca0bfb673a4ac4838074cdae58f7cce26412.jpg)  
+Reference
+
+![](images/4540fb3623e1042b72cdc51d40fae52db70bda03718617114a98d44bd5cf175c.jpg)  
+Transfer the style in the reference image.  
+Add the object in the reference image on the women.
+
+![](images/a83a9ab1de823a8e4032e1dedd9635d6aa3ee84abd0b1589dedcb0a1c5522e76.jpg)  
+Reference
+
+![](images/22d931f8d1d6d4cd46b69a5e91dbd7bdc650d603c3c3e56bcf3baecacfd0144e.jpg)  
+Figure 2. Multimodal instruction following emerges in VEGGIE, allowing for style transfer or object addition from reference images.
+
+![](images/b5f1369b96b80acff89b8832473d68767bd364283daa8b18fad10e9742759d4e.jpg)  
+Figure 3. In-context editing emerges in VEGGIE, allowing for few-shot learning of editing tasks with paired image demonstrations.
+
+# 2. Related Work
+
+Instructional Video Editing Video Diffusion Models (VidDMs) [2, 4, 6, 7, 15, 26, 49, 67, 71], enable high-quality video generation across a wide range of video concepts. Building on these advances, video editing methods have emerged as tools for video design, allowing users to manipulate video concepts such as adding, removing, altering objects and style translation [20, 53, 62, 79, 85]. To enhance user experiences, instructional video editing methods [54, 70, 86] have been developed, using triples of video instructions, source videos, and target videos for training. These methods demonstrate limited performance when complex multimodal reasoning is required, as noted by previous research on instructional image editing [17]. Moreover, they struggle with diverse editing tasks, from addition and deletion to stylization. For example, LGVI [66] is primarily designed for removal tasks, while TokenFlow [20] struggles with local edits such as adding, removing, or changing objects. We address this limitation with pixel-level multitasking fine-tuning on well-curated instructional video editing datasets covering various grounding and editing skills.
+
+Video Grounding and Segmentation Visual grounding requires models to connect language to its corresponding visual concept in the visual context [40, 47]. This is commonly evaluated via the language-guided semantic localization tasks, ranging from simple referring expressions in RefCOCO series [48, 77] and their generalized variant [44] that takes no-target and multi-target into account. Recently, grounded multimodal large lan
+
+guage models (MLLMs) are trained for object grounding to bounding boxes [8, 51, 52, 76, 82, 87] and segmentation masks [56, 69, 81, 84] using text-image pairs with fine-grained annotations linking phrases to entities. These models unlock the potential of reasoning segmentation [11, 36], bringing language-informed reasoning into semantic segmentation. Instead of using dedicated object detection or segmentation modules, we achieve video grounding through end-to-end training with only diffusion loss.
+
+# 3. Our Method: VEGGIE
+
+In this paper, we introduce a Video Editor with Grounded Generation from Instructions (VEGGIE), a unified and versatile generative video model. It combines the complex instruction understanding and reasoning capabilities of MLLMs with the generative capacity of VidDMs. The model is trained end-to-end with diffusion loss only. VEGGIE efficiently handles diverse user inputs, including direct instructions, complex questions requiring in-depth reasoning, and multimodal conditioning. It performs various pixel-level manipulations, enabling tasks such as video concept addition, removal, changing, stylization, grounding, and reasoning segmentation based on user instructions. We elaborate on the model design (Sec. 3.1), training and inference process (Sec. 3.2), and data curation (Sec. 3.3).
+
+# 3.1. Model Architecture
+
+VEGGIE consists of four main components (see Fig. 4): (1) a multimodal large language model, (2) a set of learnable
+
+![](images/419c74734a879ff33d66c796ab67f7d5591f5073932ab3c43c6916f3395226b5.jpg)
+
+![](images/4614bdaab383a2d7be34477abb6dd43be30e9fc13b804982c124645ab6ae1578.jpg)  
+Figure 4. Overview of our proposed end-to-end VEGGIE framework. Our Multimodal Large Language Model first understands input video frames and diverse user instructions, then it generates frame-wise reasoning queries that maintain per-frame editing conditions for the video diffusion model. The video diffusion model will render the MLLM-generated conditions to the pixel space for diverse tasks, including video editing, video grounding, and video reasoning segmentation with questions. We only apply diffusion loss for the whole pipeline training.
+
+grounded task queries, (3) an alignment network (single-layer MLP) that projects the MLLM output into the condition space of the diffusion model, and (4) a video diffusion model initialized from an instructional image editing model [83]. Our model first generates latent conditions for target video frames by querying multimodal context using an MLLM, then renders these conditions at the pixel level through a video diffusion model, as detailed below.
+
+MLLM for Generating Grounded Task Guidance. As illustrated in the left of Fig. 4, given a video consisting of a sequence of frames  $V = [f_1, \dots, f_n]$ , where  $n$  is the frame number of the given video, a user instruction/question  $I$ , our goal is to obtain the response  $\widehat{V} = [\widehat{f}_1, \dots, \widehat{f}_n]$  at pixel space that faithfully reflects user instruction about the given video. The MLLM module processes both the input video  $V$  and a user instruction  $I$  to generate a sequence of grounded task tokens per frame:  $C = [c_1, \dots, c_n]$ , which are input and output in parallel. These tokens serve as task guidance and implicitly encode the target manipulation, such as object attributes, spatial relationships, or style transfer parameters. The MLLM ensures the model captures both explicit user instructions and implicit reasoning needs.
+
+VidDM for Rendering MLLM Guidance at Pixel Space. As illustrated in the right of Table 4, the VidDM takes the original video  $V$  and the grounded task tokens  $C$  as conditions to synthesize the target video  $\widehat{V}$ . The original video is concatenated with the noise volume, and the task tokens are input to the cross-attention. With grounded task guidance in denoising steps, the generation process ensures that the output faithfully follows user instructions while preserving the
+
+video's structure and motion dynamics. Through iterative denoising, it refines each frame while maintaining temporal consistency, applying pixel modifications coherently for a smooth and visually consistent output video  $\widehat{V}$ .
+
+# 3.2. Curriculum Learning from Images to Videos
+
+Training the model directly on video tasks presents two key challenges: (1) misalignment between MLLM and diffusion model representations, making it difficult for the diffusion model to interpret MLLM-generated task queries with limited fine-tuning data, and (2) the diffusion model's lack of multitasking capability, even for image tasks, due to insufficient training on diverse tasks. Our initial experiments also found the model collapsed when the whole pipeline was directly trained with all data. These challenges/observations underscore the need for pre-alignment between MLLM and the diffusion model to enable seamless adaptation from language-space task queries to pixel-space modifications. To this end, we adopt a two-stage curriculum learning strategy for the proposed VEGGIE framework.
+
+Stage 1: Aligning Diffusion and Language Spaces. In the first stage, we align the diffusion model with the MLLM using large-scale image-level instructional editing data. The MLLM remains frozen while we update the alignment network, grounded task queries, and diffusion UNet. This process fine-tunes the diffusion model weights to align with the language space, enabling the model to interpret MLLM-generated guidance and translate user instructions into pixel-level edits while preserving the MLLM's strong ability to understand instructions and user intentions.
+
+<table><tr><td>Type</td><td>Source</td><td>R.</td><td>E.</td><td>G.</td><td># Img./Vid.</td><td># Ins.</td></tr><tr><td rowspan="7">Video</td><td>ROVI [66]</td><td>✓</td><td>✓</td><td>✓</td><td>4.3K</td><td>27.4K</td></tr><tr><td>VPLM [75]</td><td></td><td>✓</td><td></td><td>4.3K</td><td>5.5K</td></tr><tr><td>GroundMoRe [12]</td><td>✓</td><td></td><td>✓</td><td>1.3K</td><td>5.5K</td></tr><tr><td>RVoS [60]</td><td></td><td></td><td>✓</td><td>1.9K</td><td>6.1K</td></tr><tr><td>MeViS [14]</td><td></td><td></td><td>✓</td><td>1.5K</td><td>17.1K</td></tr><tr><td>InstructV2V [9]</td><td></td><td>✓</td><td></td><td>68.3K</td><td>68.3K</td></tr><tr><td>VEG-Edit (Ours)</td><td></td><td>✓</td><td></td><td>4.0K</td><td>6.2K</td></tr><tr><td colspan="6">Total</td><td>136.1K</td></tr><tr><td rowspan="7">Image</td><td>Seed-Data-Edit [18]</td><td></td><td>✓</td><td></td><td>3M</td><td>3M</td></tr><tr><td>LISA [37]</td><td>✓</td><td></td><td>✓</td><td>0.2K</td><td>1.3K</td></tr><tr><td>gRefCoCo [44]</td><td></td><td></td><td>✓</td><td>13.6K</td><td>73.4K</td></tr><tr><td>PhraseCut [65]</td><td></td><td></td><td>✓</td><td>310.8K</td><td>310.8K</td></tr><tr><td>EraseDraw [5]</td><td></td><td>✓</td><td></td><td>64.9K</td><td>42.4K</td></tr><tr><td>MagicBrush [83]</td><td></td><td>✓</td><td></td><td>9.3K</td><td>9.3K</td></tr><tr><td>SmartEdit [29]</td><td></td><td>✓</td><td></td><td>0.5K</td><td>0.9K</td></tr><tr><td colspan="6">Total</td><td>3438.1K</td></tr></table>
+
+Table 1. Summary of our data for training. R.: Reasoning, E.: Editing, G.: Grounding. #Img/Vid: the number of images/videos, and #Ins.: the number of instruction-image/video pairs.
+
+Stage 2: Enhancing Temporal Consistency and Dynamics. With the MLLM and diffusion model aligned, fine-tuning diverse instructional video editing data becomes more effective for improved instruction following at pixel-space including temporal consistency, dynamic coherence, and editing faithfulness. In this stage, we fine-tune the framework with the MLLM, including the alignment network, grounded task queries, and all 3 dimensions in diffusion UNet, end-to-end with carefully curated multitasking instructional video editing data. Following prior work [22, 67], we inflated the 2D UNet from Stage 1 with temporal attention layers for video adaptation. For both stages 1 and 2, we optimize the framework with a single diffusion loss, enabling unified learning for improved instructional video editing performance while maintaining simplicity and efficiency.
+
+Classifier-Free Guidance during Testing. We employ classifier-free guidance to balance quality and diversity in diffusion-generated samples. Following prior work [3, 17], we apply classifier-free guidance to instructional visual editing considering two conditions: the grounded task tokens and the original video. To obtain unconditional guidance, we set null values  $(\varnothing)$  for both task tokens and input video. In this case, our score estimate is:
+
+$$
+\begin{array}{l} \tilde {e _ {\theta}} (z _ {t}, c _ {T}, c _ {V}) = e _ {\theta} (z _ {t}, \varnothing , \varnothing) \\ + g _ {T} \cdot \left(e _ {\theta} \left(z _ {t}, c _ {V}, c _ {T}\right) - e _ {\theta} \left(z _ {t}, c _ {V}, \varnothing\right)\right) \\ + g _ {V} \cdot \left(e _ {\theta} \left(z _ {t}, c _ {V}, \varnothing\right) - e _ {\theta} \left(z _ {t}, \varnothing , \varnothing\right)\right), \\ \end{array}
+$$
+
+where  $\theta$  represents the model parameters,  $C_T$  and  $C_V$  denote the task tokens and video conditions,  $\varnothing$  is the null value,  $z_t$  is the noised latent at timestamp  $t$ , and  $g_T$  and  $g_V$  are the task guidance and video guidance scales, respectively. More training details are included later in Appendix.
+
+# 3.3. Data Curation Pipeline
+
+Existing video editing models, both instructional and non-instructional, struggle with diverse editing skills due to the lack of high-quality multitasking fine-tuning data. In this section, we introduce our data curation strategy to support VEGGIE in achieving versatile video editing skills. As listed in Tab. 1, we collect 3.4M image and 133.9K video data from diverse sources to support our VEGGIE curriculum learning as discussed in Sec. 3.2. We create our training dataset from two sources: (1) collecting existing image and video data and converting it into an instructional editing format, and (2) synthesizing new instructional video editing samples using existing datasets and generative models.
+
+Collecting Diverse Multitask Image and Video Data. We bring together instructional editing data from both image (Seed-Data-Edit [18], MagicBrush [83], EraseDraw [5]) and video (InstructV2V [9], VPLM [75]) sources. These datasets provide pairs of original and edited visual contents with user instructions. The tasks include adding, removing, and changing objects, stylizing, and performing global/local edits. Beyond editing datasets, we incorporate segmentation data at both the image level (gRefCoCo [44] and Phrase-Cut [65]) and the video level (RVoS and MeViS). These segmentation tasks are reformulated as color-filling challenges, which guide the model in learning referring grounding (i.e., understanding which object or region to edit) and strengthen its conceptual learning. To further unlock complex instruction understanding via MLLM, we include data that requires more advanced reasoning and implicit referencing. Specifically, we include: reasoning segmentation (LISA [37]), reasoning editing (SmartEdit [29]), interactive video inpainting (LGVI [66]), and motion-grounded video reasoning (GroundMoRe [12]). These tasks help VEGGIE learn implicit references and reasoning.
+
+Synthesizing Instructional Video Editing Data via Image-to-Video Animation. Recent methods [3, 54] generate synthetic instructional video-editing data by first creating text instructions with LLM, then getting edited videos via T2V models and prompt-to-prompt editing [23]. While these methods adapt image-based editing pipelines [9] to videos, the generated data suffer from temporal-consistency issues. To address this gap, we propose a novel image-to-video animation strategy that leverages the abundance of high-quality image-level instructional editing datasets [64, 83], which provide well-annotated instructions, paired edited images, and well-organized editing skill categories. As illustrated in Fig. 5, given an original image  $I$ , an edited image  $\bar{I}$ , and an instruction from an instructional image editing dataset [64], our approach involves three key steps. First, we use an offline MLLM [61, 72] to generate an image caption and an animation prompt that describes plausible motion within the image. Next, an image-to-video (I2V) model animates the image into a video  $V$ . Finally,
+
+![](images/519aa3cbb726c3f46600125ec3400403ec4db3182b9494a57667d8cbec97125a.jpg)  
+Figure 5. Our data generation pipeline for synthetic instructional video editing data. It injects dynamics into well-constructed instructional image editing datasets via the Image-to-Video (I2V) Model, and generates paired video data for instruction editing.
+
+we generate the corresponding edited video  $\bar{V}$  using a first-frame-conditioned video editing model [35], leveraging  $\bar{I}$  as a strong prior to ensure consistent edits across frames. Finally, to ensure data quality, we evaluate each original-edited video pair with automatic video quality evaluation metrics [30], which assess the generated videos from diverse dimensions, e.g., motion smoothness, image quality, and background consistency. This pipeline transforms carefully curated image-based datasets into instructional video-editing resources while preserving the precision of the original edits. As a result, our data method expands the availability of high-quality synthetic video-editing data, supporting a wider range of editing tasks in our end-to-end unified framework. More details on data generation, prompting, examples, and pre/post-processing are in the Appendix.
+
+# 4. Experiments
+
+We first introduce the VEG-Bench Benchmark and then demonstrate the superiority of VEGGIE across diverse video instructional editing skills. More experiments, visualization, and implementation details are in the Appendix.
+
+# 4.1. VEG-Bench and Metrics
+
+As no existing benchmark is designed for fine-grained instructional video editing skills, we manually collect and annotate VEG-Bench, containing 132 video-instruction pairs that balanced cover 8 different video generative skills (15-20 for each). Beyond standard metrics, including text-to-video alignment (CLIP-Text [55]), video smoothness (CLIP-F [55]), and image quality (MUSIQ [32]), we also
+
+first introduce MLLM-as-a-Judge to give a holistic evaluation score according to the given original video, edited video, and user instruction. It is achieved by prompting GPT4o [21] to evaluate whether the requested semantic change has been fulfilled, using a scale from 1 to 10. For addition and removal, we also introduce an object detector (GroundingDiNo [45]) to detect if the object is added/removed faithfully. For grounding and reasoning segmentation, we following video grounding tasks [12, 14, 33, 58] and adopt the Jaccard index  $(\mathcal{I})$  [31], F-measure  $(\mathcal{F})$  [13], and their mean  $(\mathcal{J} \& \mathcal{F})$ . We also compute SSIM between the generated video and the original video masked with GT masks. More evaluation/metrics details are included in Appendix.
+
+# 4.2. Experimental Results
+
+Instructional Video Editing over Diverse Skills. As shown in Tab. 2, we evaluate 7 different models on VEG-Bench across 8 distinct editing skills. Overall, VEGGIE demonstrates the best performance among instructional video editing models. Compared to VEGGIE, non-instructional models often struggle with concept removal and addition. This limitation arises because these models rely on attention control or additional conditions (e.g., depth maps) that impose strong priors, constraining the model and making object addition or removal challenging. We also observe that InsV2V achieves high scores in quality and smoothness metrics, but underperforms in alignment and MLLM judgment, which demand faithful semantic changes. Qualitative examples in Fig. 6 illustrate that InsV2V often makes minimal changes to the input video, resulting in high video quality but unfaithful outputs. In contrast, VEGGIE strikes a better balance, delivering both high-quality visuals and accurate semantic alignment with the intended edits.
+
+Can Multi-Task Help Each Other? To test the previous hypothesis, we train our model on the VPLM [75] dataset, which includes paired grounding and removal tasks (approximately 5.5K samples for each task). We focus on these tasks as representative examples due to their straightforward evaluation against ground truth. As shown in Table 3, multitask training yields a lower FVD score and a higher SSIM score, demonstrating that learning to locate and remove a video concept can mutually reinforce performance. We show an example in Fig. 7. However, this conclusion only holds with a balanced data combination. We also observe that an excessive amount of grounding data can introduce more artifacts and negatively impact visual editing skills.
+
+Emergent Zero-shot Multimodal Instruction Following. We also highlight the emergent behavior of VEGGIE on multi-modal instruction following, even without dedicated training data for this specific editing instruction. Notably, VEGGIE demonstrates the ability to perform zero-shot multimodal instructional video editing. As illustrated in Fig. 2, VEGGIE can transfer styles or add objects from a reference
+
+![](images/448541813648206c696b8060be6836952b32cd5e03312d900e585ca11af2f430.jpg)  
+[Addition] Please add a ball in the given video frames.
+
+![](images/67afd19caa3fbcb6b392c5ad751d2b0f2e3161441709a94940152b6d86723f7e.jpg)  
+[Removal] Please remove the man in black in given video frames.
+
+![](images/373ec45a7f7d566ba4d3a78d8e06b09d7bad2b84179f4cdf62f91a27bbbf530a.jpg)  
+[Swap] Replace golden building with a white mountain.
+
+![](images/41198c72d6faf22421eb58cefcc01dfd03a7dbb2d35994a01bd56dc226535d45.jpg)  
+[Environment] Make it on the beach.
+
+![](images/2aaed0e07fec7c8843e49de92130946b5b89bbd966dcb8e209bd72feffe8789e.jpg)  
+[Color] Make the swan white.
+
+![](images/a238051c85810f38dffc43dcef81eb1d2021aded79c1a77a3508a3bafc592ab6.jpg)  
+[Texture] Make the rhinocero furry.
+
+![](images/a41b7f9d6e96f7f28878ef89d70835e2d12c0c4a44228a64db4c7f7f34d6c0d5.jpg)  
+[Style] Make it chinese ink style.  
+Figure 6. Qualitative comparison of editing results across 8 different abilities (splitting visual features into color and texture). We provide zoom-in details for a more detailed comparison. Best viewed in color. More in Appendix.
+
+![](images/6fecf891dfe80f439d7ac4732ec66bb78527fab1d6f6d18a6235978f3c780c18.jpg)  
+[Grounding] Could you locate the knife
+
+![](images/b7c113e68f1b140f6b2fd89134382a5e2e6aa73949f6da185815fa51ef01fdde.jpg)  
+[Reasoning] What can be used for heating food?
+
+<table><tr><td rowspan="2">Methods
+Generator</td><td colspan="3">Non-Instructional Editing Model</td><td colspan="4">Instructional Editing Model</td></tr><tr><td>VidToMe [41] (SD1.5)</td><td>TokenFlow [20] (SD2.1)</td><td>Flatten [10] (SD2.1)</td><td>InstructDiff [19] (SD1.5)</td><td>LGVI [66] (SD1.5)</td><td>InsV2V [9] (SD1.5)</td><td>VEGGIE (Ours) (SD1.5)</td></tr><tr><td colspan="8">Concept Addition</td></tr><tr><td>MLLM-Judge (↑)</td><td>5.00</td><td>5.80</td><td>6.62</td><td>7.26</td><td>2.73</td><td>5.69</td><td>7.44</td></tr><tr><td>Alignment (↑)</td><td>27.80</td><td>29.30</td><td>28.22</td><td>28.10</td><td>25.06</td><td>28.27</td><td>29.27</td></tr><tr><td>Smoothness (↑)</td><td>96.79</td><td>97.26</td><td>95.74</td><td>93.66</td><td>96.09</td><td>96.94</td><td>94.93</td></tr><tr><td>Quality (↑)</td><td>62.81</td><td>65.62</td><td>55.45</td><td>54.08</td><td>41.59</td><td>56.24</td><td>61.31</td></tr><tr><td>Detection (↑)</td><td>47.98</td><td>49.53</td><td>49.74</td><td>55.36</td><td>14.42</td><td>48.01</td><td>57.96</td></tr><tr><td colspan="8">Concept Removal</td></tr><tr><td>MLLM-Judge (↑)</td><td>2.60</td><td>3.73</td><td>4.46</td><td>6.12</td><td>6.59</td><td>2.78</td><td>5.07</td></tr><tr><td>Alignment (↑)</td><td>75.01</td><td>75.99</td><td>78.40</td><td>75.51</td><td>75.67</td><td>74.41</td><td>75.63</td></tr><tr><td>Smoothness (↑)</td><td>96.13</td><td>96.47</td><td>95.82</td><td>91.83</td><td>97.03</td><td>96.99</td><td>95.04</td></tr><tr><td>Quality (↑)</td><td>66.32</td><td>71.52</td><td>50.77</td><td>55.08</td><td>42.31</td><td>58.79</td><td>50.99</td></tr><tr><td>Detection (↑)</td><td>34.31</td><td>55.16</td><td>70.91</td><td>64.81</td><td>78.40</td><td>25.64</td><td>70.22</td></tr><tr><td colspan="8">Object Changing</td></tr><tr><td>MLLM-Judge (↑)</td><td>5.00</td><td>6.53</td><td>7.37</td><td>7.00</td><td>2.06</td><td>6.60</td><td>6.63</td></tr><tr><td>Alignment (↑)</td><td>25.69</td><td>28.76</td><td>27.06</td><td>27.36</td><td>22.17</td><td>26.60</td><td>27.77</td></tr><tr><td>Smoothness (↑)</td><td>96.23</td><td>97.21</td><td>96.13</td><td>92.07</td><td>95.66</td><td>96.74</td><td>95.44</td></tr><tr><td>Quality (↑)</td><td>64.06</td><td>69.97</td><td>59.37</td><td>55.01</td><td>38.20</td><td>60.90</td><td>58.15</td></tr><tr><td colspan="8">Environment &amp; Background Changing</td></tr><tr><td>MLLM-Judge (↑)</td><td>5.81</td><td>7.35</td><td>7.37</td><td>6.05</td><td>2.37</td><td>6.60</td><td>7.18</td></tr><tr><td>Alignment (↑)</td><td>28.17</td><td>30.00</td><td>30.04</td><td>28.03</td><td>21.94</td><td>28.27</td><td>29.15</td></tr><tr><td>Smoothness (↑)</td><td>95.76</td><td>96.96</td><td>95.90</td><td>89.85</td><td>95.66</td><td>96.03</td><td>94.58</td></tr><tr><td>Quality (↑)</td><td>61.95</td><td>67.06</td><td>54.58</td><td>53.06</td><td>38.97</td><td>54.94</td><td>54.25</td></tr><tr><td colspan="8">Visual Feature Changing (Color &amp; Texture)</td></tr><tr><td>MLLM-Judge (↑)</td><td>5.86</td><td>6.85</td><td>6.60</td><td>6.43</td><td>2.14</td><td>7.53</td><td>7.33</td></tr><tr><td>Alignment (↑)</td><td>27.99</td><td>29.25</td><td>29.46</td><td>27.54</td><td>23.18</td><td>28.88</td><td>28.69</td></tr><tr><td>Smoothness (↑)</td><td>95.93</td><td>97.10</td><td>95.83</td><td>91.71</td><td>94.75</td><td>96.66</td><td>94.52</td></tr><tr><td>Quality (↑)</td><td>65.80</td><td>69.31</td><td>53.32</td><td>58.29</td><td>36.27</td><td>59.36</td><td>57.91</td></tr><tr><td colspan="8">Stylization</td></tr><tr><td>MLLM-Judge (↑)</td><td>7.23</td><td>7.62</td><td>8.31</td><td>7.41</td><td>3.71</td><td>8.07</td><td>8.26</td></tr><tr><td>Alignment (↑)</td><td>29.84</td><td>30.25</td><td>29.00</td><td>27.74</td><td>22.80</td><td>29.14</td><td>29.38</td></tr><tr><td>Smoothness (↑)</td><td>96.31</td><td>97.23</td><td>96.71</td><td>88.97</td><td>95.62</td><td>96.50</td><td>95.69</td></tr><tr><td>Quality (↑)</td><td>64.05</td><td>68.22</td><td>53.18</td><td>54.15</td><td>35.76</td><td>62.59</td><td>57.00</td></tr><tr><td colspan="8">Object Grounding</td></tr><tr><td>SSIM (↑)</td><td>40.47</td><td>50.46</td><td>47.21</td><td>37.98</td><td>66.84</td><td>49.65</td><td>70.90</td></tr><tr><td>Jaccard Index J (↑)</td><td>13.85</td><td>19.29</td><td>25.62</td><td>19.88</td><td>1.52</td><td>13.89</td><td>37.74</td></tr><tr><td>F-measure F (↑)</td><td>15.50</td><td>16.86</td><td>17.60</td><td>12.81</td><td>3.07</td><td>17.37</td><td>21.83</td></tr><tr><td colspan="8">Reasoning Segmentation</td></tr><tr><td>SSIM (↑)</td><td>-</td><td>-</td><td>-</td><td>32.39</td><td>44.47</td><td>59.86</td><td>68.41</td></tr><tr><td>Jaccard Index J (↑)</td><td>-</td><td>-</td><td>-</td><td>14.02</td><td>10.12</td><td>16.89</td><td>22.53</td></tr><tr><td>F-measure F (↑)</td><td>-</td><td>-</td><td>-</td><td>8.07</td><td>9.06</td><td>10.45</td><td>15.97</td></tr><tr><td>Avg. Ranking</td><td>2.61</td><td>1.41</td><td>1.96</td><td>3.00</td><td>3.21</td><td>2.00</td><td>1.78</td></tr></table>
+
+Table 2. Comparison of video editing task with instructional / non-instructional models on VEG-Bench. -: the task is not capable of non-instructional models. We gray out numbers of non-instructional models that are in different categories.  
+
+<table><tr><td>Settings</td><td>Removal (FVD ↓)</td><td>Grounding (SSIM ↑)</td></tr><tr><td>Grd.-only</td><td>-</td><td>52.34</td></tr><tr><td>Rmv.-only</td><td>1098.52</td><td>-</td></tr><tr><td>Mixed</td><td>987.80</td><td>55.21</td></tr></table>
+
+Table 3. An ablation study on whether multi-task learning provides transferable benefits that enhance performance across tasks. We focus on removal and grounding tasks as representative examples.
+
+![](images/16dd7320589fd8681a64e1346642e6554bace0aed5cea65f14131a25d7187071.jpg)  
+Figure 7. Comparison between single- and multi-skill models with different data training. We find tasks can help each other.
+
+image into the input video based on instructions.
+
+Emergent Few-shot In-Context Editing. As shown in Fig. 3, VEGGIE can effectively utilize a few example image pairs to transfer the intended editing changes seamlessly to the input video. We observe that VEGGIE exhibits in-context learning for image editing without the need for language instructions. Instead, it uses image pairs as examples to infer and apply the desired editing intention directly.
+
+# 5. Conclusion
+
+We present VEGGIE, a unified end-to-end model for instructional video editing that handles diverse pixel-level tasks. VEGGIE leverages MLLM for robust instruction understanding and employs a video diffusion model to execute pixel-level edits. Our framework uses a single diffusion loss for end-to-end optimization across varied tasks/skills. We also introduce a novel synthetic data generation pipeline and VEG-Bench, a benchmark that assesses a broad range of editing skills. Our VEGGIE outperforms previous methods as a versatile, all-in-one solution. We hope our model, data, and benchmark to advance research on instructional generative video models.
+
+# References
+
+[1] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 2  
+[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3  
+[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 5  
+[4] 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. OpenAI, https://openai.com/research/video-generation-models-as-world-simulators, 2024.2.3  
+[5] Alper Canberk, Maksym Bondarenko, Ege Ozguroglu, Ruoshi Liu, and Carl Vondrick. Erasedraw: Learning to insert objects by erasing them from images. arXiv preprint arXiv:2409.00522, 2024. 5  
+[6] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23206-23217, 2023. 3  
+[7] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stablevideo: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 3  
+[8] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3  
+[9] Jiaxin Cheng, Tianjun Xiao, and Tong He. Consistent video-to-video transfer using synthetic dataset. In The Twelfth International Conference on Learning Representations, 2024. 5, 8, 1, 2  
+[10] Yuren Cong, Mengmeng Xu, christian simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: optical FLOW-guided ATTENtion for consistent text-to-video editing. In The Twelfth International Conference on Learning Representations, 2024. 8, 1  
+[11] Andong Deng, Tongjia Chen, Shoubin Yu, Taojiannan Yang, Lincoln Spencer, Yapeng Tian, Ajmal Saeed Mian, Mohit Bansal, and Chen Chen. Motion-grounded video reasoning: Understanding and perceiving motion at pixel level. arXiv preprint arXiv:2411.09921, 2024. 3  
+[12] Andong Deng, Tongjia Chen, Shoubin Yu, Taojiannan Yang, Lincoln Spencer, Yapeng Tian, Ajmal Saeed Mian, Bansal Mohit, and Chen. Chen. Motion-grounded video reasoning: Understanding and perceiving motion at pixel level. 2024. 5, 6, 2
+
+[13] Lee R Dice. Measures of the amount of ecologic association between species. Ecology, 26(3):297-302, 1945. 6  
+[14] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 5, 6  
+[15] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In CVPR, pages 7346-7356, 2023. 3  
+[16] Hao Fei, Shengqiong Wu, Hanwang Zhang, Tat-Seng Chua, and Shuicheng Yan. Vitron: A unified pixel-level vision llm for understanding, generating, segmenting, editing, 2024. 2  
+[17] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 2, 3, 5  
+[18] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. Seed-data-edit technical report: A hybrid dataset for instructional image editing. arXiv preprint arXiv:2405.04007, 2024. 5  
+[19] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024. 8, 1, 2  
+[20] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 2, 3, 8, 1  
+[21] gpt 4o. https://openai.com/index/hello-gpt-4o/.2024.6  
+[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 5, 1  
+[23] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022.5  
+[24] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2  
+[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 2022. 2  
+[26] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In The Eleventh International Conference on Learning Representations, 2023. 3  
+[27] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 1
+
+[28] Jiahao Hu, Tianxiong Zhong, Xuebo Wang, Boyuan Jiang, Xingye Tian, Fei Yang, Pengfei Wan, and Di Zhang. Vivid-10m: A dataset and baseline for versatile and interactive video local editing. arXiv preprint arXiv:2411.15260, 2024. 2  
+[29] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024. 2, 5  
+[30] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6, 2  
+[31] Paul Jaccard. The distribution of the flora in the alpine zone. 1. New phytologist, 11(2):37-50, 1912. 6  
+[32] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6  
+[33] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 6  
+[34] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 2  
+[35] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv preprint arXiv:2403.14468, 2024. 2, 6  
+[36] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 3  
+[37] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 5  
+[38] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1  
+[39] Jialu Li, Shoubin Yu, Han Lin, Jaemin Cho, Jaehong Yoon, and Mohit Bansal. Training-free guidance in text-to-video generation via multimodal planning and structured noise initialization. arXiv preprint arXiv:2504.08641, 2025. 2  
+[40] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language
+
+image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 3  
+[41] Xirui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidthome: Video token merging for zero-shot video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7486-7495, 2024. 2, 8, 1  
+[42] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 2  
+[43] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 2  
+[44] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23592-23601, 2023. 3, 5  
+[45] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6  
+[46] Shaoteng Liu, Tianyu Wang, Jui-Hsien Wang, Qing Liu, Zhifei Zhang, Joon-Young Lee, Yijun Li, Bei Yu, Zhe Lin, Soo Ye Kim, et al. Generative video propagation. arXiv preprint arXiv:2412.19761, 2024. 2  
+[47] Ziqiao Ma, Jiayi Pan, and Joyce Chai. World-to-words: Grounded open vocabulary acquisition through fast mapping in vision-language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 524–544, 2023. 3  
+[48] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 3  
+[49] Kangfu Mei and Vishal Patel. Vidm: Video implicit diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 9117-9125, 2023. 3  
+[50] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, Mubarak Shah, and Ajmal Mian. Towards temporally consistent referring video object segmentation. https://arxiv.org/abs/2403.19407, 2024. 2  
+[51] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. Grounding multimodal large language models to the world. In The Twelfth International Conference on Learning Representations, 2024. 3  
+[52] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, et al. Detgpt: Detect what you need via reasoning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14172-14189, 2023. 3  
+[53] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing
+
+attentions for zero-shot text-based video editing. In ICCV, pages 15932-15942, 2023. 2, 3  
+[54] Bosheng Qin, Juncheng Li, Siliang Tang, Tat-Seng Chua, and Yueting Zhuang. Instructvid2vid: Controllable video editing with natural language instructions. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2, 3, 5  
+[55] 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. 6  
+[56] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3  
+[57] 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  
+[58] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 6  
+[59] Shangkun Sun, Xiaoyu Liang, Songlin Fan, Wenxu Gao, and Wei Gao. Ve-bench: Subjective-aligned benchmark suite for text-driven video editing quality assessment. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2  
+[60] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019. 5  
+[61] 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. arXiv preprint arXiv:2409.12191, 2024. 5, 2  
+[62] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 2, 3  
+[63] Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal IIm as an agent for unified image generation and editing. arXiv preprint arXiv:2407.05600, 2024. 2  
+[64] Cong Wei, Zheyang Xiong, Weiming Ren, Xinrun Du, Ge Zhang, and Wenhu Chen. Omniedit: Building image edit
+
+ing generalist models through specialist supervision. arXiv preprint arXiv:2411.07199, 2024. 5, 2  
+[65] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 5  
+[66] Jianzong Wu, Xiangtai Li, Chenyang Si, Shangchen Zhou, Jingkang Yang, Jiangning Zhang, Yining Li, Kai Chen, Yunhai Tong, Ziwei Liu, et al. Towards language-driven video inpainting via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12501-12511, 2024. 2, 3, 5, 8, 1  
+[67] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7623-7633, 2023. 3, 5  
+[68] Jay Zhangjie Wu, Xiuyu Li, Difei Gao, Zhen Dong, Jinbin Bai, Aishani Singh, Xiaoyu Xiang, Youzeng Li, Zuwei Huang, Yuanxi Sun, et al. Cvpr 2023 text guided video editing competition. arXiv preprint arXiv:2310.16003, 2023. 2  
+[69] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3  
+[70] Zhen Xing, Qi Dai, Zihao Zhang, Hui Zhang, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Vidiff: Translating videos via multi-modal instructions with diffusion models. arXiv preprint arXiv:2311.18837, 2023. 3  
+[71] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3  
+[72] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 5  
+[73] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1  
+[74] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2
+
+[75] Jaehong Yoon, Shoubin Yu, and Mohit Bansal. Raccoon: Remove, add, and change video content with auto-generated narratives. arXiv preprint arXiv:2405.18406, 2024. 2, 5, 6  
+[76] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. In The Twelfth International Conference on Learning Representations, 2023. 3  
+[77] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 3  
+[78] Shoubin Yu, Jacob Zhiyuan Fang, Jian Zheng, Gunnar Sigurdsson, Vicente Ordonez, Robinson Piramuthu, and Mohit Bansal. Zero-shot controllable image-to-video animation via motion decomposition. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3332-3341, 2024. 2  
+[79] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2, 3  
+[80] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975-11986, 2023. 1  
+[81] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Lei Zhang, Chunyuan Li, et al. Llava-grounding: Grounded visual chat with large multimodal models. arXiv preprint arXiv:2312.02949, 2023. 3  
+[82] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferret-v2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 3  
+[83] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 4, 5, 1  
+[84] Yichi Zhang, Ziqiao Ma, Xiaofeng Gao, Suhaila Shakiah, Qiaozi Gao, and Joyce Chai. Groundhog: Grounding large language models to holistic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3  
+[85] 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. arXiv preprint arXiv:2312.03816, 2023. 2, 3  
+[86] Zhenghao Zhang, Zuozhuo Dai, Long Qin, and Weizhi Wang. Effived: Efficient video editing via text-instruction diffusion models. arXiv preprint arXiv:2403.11568, 2024. 2, 3  
+[87] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. 3

2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78c227e9126fa5888a5e552a39c6477995db227dc770bbca52b8786616e40c28
+size 1338865

2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/b08d8e36-4b30-4260-aad4-6cb0fd4d67ca_content_list.json

@@ -0,0 +1,1631 @@
+[
+    {
+        "type": "text",
+        "text": "VFlowOpt: A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization",
+        "text_level": 1,
+        "bbox": [
+            109,
+            128,
+            888,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sihan Yang<sup>1</sup> Runsen Xu<sup>1,2</sup> Chenhang Cui<sup>3</sup> Tai Wang<sup>1†</sup> Dahua Lin<sup>1,2</sup> Jiangmiao Pang<sup>1†</sup>",
+        "bbox": [
+            120,
+            200,
+            875,
+            222
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$  Shanghai AI Laboratory  $^{2}$  The Chinese University of Hong Kong  $^{3}$  National University of Singapore",
+        "bbox": [
+            184,
+            223,
+            813,
+            239
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "† Corresponding Author taiwang.me@gmail.com pangjiangmiao@gmail.com",
+        "bbox": [
+            217,
+            239,
+            772,
+            256
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            291,
+            326,
+            306
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large Multimodal Models (LMMs) excel in visual-language tasks by leveraging numerous visual tokens for fine-grained visual information, but this token redundancy results in significant computational costs. Previous research aimed at reducing visual tokens during inference typically leverages importance maps derived from attention scores among vision-only tokens or vision-language tokens to prune tokens across one or multiple pruning stages. Despite this progress, pruning frameworks and strategies remain simplistic and insufficiently explored, often resulting in substantial performance degradation. In this paper, we propose VFlowOpt, a token pruning framework that introduces an importance map derivation process and a progressive pruning module with a recycling mechanism. The hyperparameters of its pruning strategy are further optimized by a visual information flow-guided method. Specifically, we compute an importance map for image tokens based on their attention-derived context relevance and patch-level information entropy. We then decide which tokens to retain or prune and aggregate the pruned ones as recycled tokens to avoid potential information loss. Finally, we apply a visual information flow-guided method that regards the last token in the LMM as the most representative signal of text-visual interactions. This method minimizes the discrepancy between token representations in LMMs with and without pruning, thereby enabling superior pruning strategies tailored to different LMMs. Experiments demonstrate that VFlowOpt can prune  $90\\%$  of visual tokens while maintaining comparable performance, leading to an  $89\\%$  reduction in KV-Cache memory and  $3.8 \\times$  faster inference.",
+        "bbox": [
+            88,
+            324,
+            485,
+            779
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            91,
+            811,
+            220,
+            828
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large Multimodal Models (LMMs) [3, 8, 21, 22] have attained remarkable performance in tasks such as visual question answering [2, 12] and multimodal reasoning [44, 45], making them indispensable for applications like au",
+        "bbox": [
+            89,
+            839,
+            483,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/4bd6a78824a2c043ef30859d3263d0397b752420a0d40c549f030e7784e45242.jpg",
+        "image_caption": [
+            "Figure 1. VFlowOpt Performance and Efficiency. Our VFlowOpt significantly outperforms other LMM token reduction methods across three general benchmarks (MME, MMStar, and MMBench) on LLaVA-OneVision-7B. It achieves zero performance degradation when reducing  $50\\%$  of visual tokens. Furthermore, when retaining only  $10\\%$  of the tokens, it achieves  $90\\%$  of the original performance while reducing KV-Cache memory usage by  $89\\%$  and accelerating inference by  $3.8\\times$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            280,
+            903,
+            462
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tonomous driving [9, 31, 35] and robotics [18, 20, 24]. To capture fine-grained visual details, modern LMMs treat images as token sequences, with models such as LLaVA-1.5 [21] processing hundreds of tokens and LLaVA-OneVision [17] handling up to several thousand tokens. Nonetheless, the increasing quantity of visual tokens significantly magnifies computational cost, memory usage, and inference time, creating a critical bottleneck for LMM deployment, particularly in resource-constrained environments or latency-sensitive scenarios.",
+        "bbox": [
+            509,
+            594,
+            906,
+            744
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To alleviate these constraints, recent work explores various methods for reducing the number of visual tokens during inference. Most existing approaches evaluate the importance of visual tokens in LMMs using attention-based importance maps derived either from vision-only tokens [40, 49] or from text-visual interactions [36, 52]. Following a predefined pruning ratio, visual tokens are pruned in a single stage [33] or progressively across multiple stages [42], often with heuristic strategies applied uniformly across different LMMs. Although promising, these methods are sim",
+        "bbox": [
+            509,
+            750,
+            908,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV 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": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "23924",
+        "bbox": [
+            478,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "plastic and underexplored, with coarse-grained pruning often causing significant performance drops due to lacking fitness for different LMMs models' characteristics.",
+        "bbox": [
+            89,
+            90,
+            480,
+            136
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In response to these limitations, we propose a novel token pruning framework, VFlowOpt, that more accurately estimates token importance by integrating attention calibration and entropy, and reduces information loss through token recycling and progressive pruning strategy with fine-grained pruning ratios. Furthermore, by customizing pruning strategies for different models, our framework better preserves model performance. Specifically, the pruning framework consists of importance map computation and a progressive pruning module with a recycling mechanism. We decompose the importance map of image tokens into two aspects: the importance of tokens in the visual context, reflected by their attention maps, and the richness of visual information in each image patch, captured by its information entropy. Based on the weighted summation of these two factors, we determine which tokens to retain or prune across multiple pruning stages and introduce a recycling mechanism to aggregate pruned tokens to avoid potential information loss. Built with this pruning strategy with several hyperparameters, a key ingredient is the visual information flow-guided method to optimize the pruning strategy, which treats the last token in each pruning stage as the most representative signal of text-visual interactions during inference. By minimizing the difference between token representations in LMMs with and without pruning, VFlowOpt ultimately delivers superior pruning strategies specifically tailored to different LMMs.",
+        "bbox": [
+            91,
+            138,
+            486,
+            544
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Comprehensive experiments on multiple vision-language benchmarks validate the efficacy of VFlowOpt. In particular, it substantially lowers computational cost while preserving competitive performance. As shown in Fig. 1, VFlowOpt can prune  $50\\%$  of visual tokens with negligible performance loss. Moreover, it can prune  $90\\%$  of visual tokens while maintaining  $90\\%$  of the model's performance, resulting in an  $89\\%$  reduction in KV-Cache memory usage and a  $3.8 \\times$  speedup in inference. These experimental results highlight its effectiveness as a practical and efficient solution for deploying LMMs in resource-constrained environments.",
+        "bbox": [
+            89,
+            547,
+            482,
+            728
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            89,
+            744,
+            232,
+            760
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Large Multimodal Models",
+        "text_level": 1,
+        "bbox": [
+            89,
+            771,
+            328,
+            786
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Large Multimodal Models (LMMs) [3, 8, 21, 22] extend the reasoning capabilities of Large Language Models (LLMs) [4, 32, 38, 39] to vision-language tasks by integrating a pre-trained vision encoder [29, 46] with a language model, linked by an alignment module such as an MLP, or a query-based network. This design transforms visual inputs into token sequences that the LLM can pro",
+        "bbox": [
+            89,
+            794,
+            482,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "cess, facilitating multimodal prompts for tasks like visual question answering [2, 6, 12, 44]. To enhance performance, advanced LMMs, such as LLaVA-OneVision [17] and Qwen2-VL [34], can encode higher-resolution images into more image tokens, thereby capturing more granular visual details. However, as image resolution grows, the number of visual tokens rises exponentially, leading to substantially higher computational costs. For example, LLaVA-1.5 [21] processes  $336 \\times 336$  images into 576 tokens, whereas LLaVA-OneVision can handle  $1152 \\times 1152$  images, producing 7,290 tokens. This challenge becomes even more pronounced in video-based models like LongVA [48], which must process tokens across numerous frames. Fine-grained visual tokenization boosts LMM performance but poses an inference bottleneck, driving efforts to balance performance and cost through token reduction.",
+        "bbox": [
+            511,
+            90,
+            906,
+            333
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Token Reduction for LMMs",
+        "text_level": 1,
+        "bbox": [
+            511,
+            340,
+            764,
+            356
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Token reduction has become a key strategy for improving the efficiency of LMMs by mitigating the computational cost associated with extensive visual token sequences. Existing methods can be broadly classified into training-based and training-free approaches. Training-based methods, such as LLaVA-Mini [50] and LLaVolta [5], introduce additional modules during model training to compress visual tokens and preserve critical information, while approaches like ATP-LLaVA [43] and p-MoD [47] train pruning modules to dynamically retain important tokens across LLM layers. However, these methods demand substantial computational resources to retrain the models, limiting their real-world applicability. In contrast, training-free methods prune tokens without additional training, often leveraging attention mechanisms to identify and discard redundant tokens. For instance, FastV[7] and SparseVLM exploit text-visual attention to rank token importance, whereas FasterVLM [49], VisionZip [40], and VTC-CLS [33] rely on [CLS] token attention in the vision encoder to evaluate token importance and prune redundant tokens. FitPrune [42] and PDrop [36] propose a progressive multistage pruning strategy in LMMs to fully utilize visual information. However, the pruning strategies in these methods often lack adaptability to different LMMs, which frequently leads to significant performance drops.",
+        "bbox": [
+            511,
+            364,
+            908,
+            741
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            511,
+            753,
+            604,
+            768
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce VFlowOpt, a framework that employs a pruning strategy to reduce redundant visual tokens during LMM inference while preserving essential visual information. This strategy involves multiple hyperparameters that our VFlowOpt framework automatically optimizes for proper configuration—an essential factor for maintaining model performance. Specifically, Section 3.1 explains how our method evaluates the importance of visual",
+        "bbox": [
+            511,
+            779,
+            908,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "23925",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/11f3eec939c6bc6f8cac28c20264c1deb05bddf83b8a2d827d4fc8af5103a5c8.jpg",
+        "image_caption": [
+            "(a) Token Importance Estimation"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            99,
+            109,
+            429,
+            262
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/bd996f6af1ded17ff73c3562764b3550a45b47a57cd004df5ca48fe7e66614ac.jpg",
+        "image_caption": [
+            "(b) Token Recycling"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            447,
+            113,
+            604,
+            256
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/ac77e9c733e7b164f80cd00a8207b0782e0c95b091775067933bce49f861a589.jpg",
+        "image_caption": [
+            "(1) Inference Stage"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            98,
+            290,
+            609,
+            467
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/3c405a4b6ddb48846a13a61c0ebe6a3c45543abb3506bdd88281f3b31d79216c.jpg",
+        "image_caption": [
+            "(2) Optimization Stage"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            617,
+            88,
+            893,
+            406
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/08f38546afb6798412810532e22ad42dbf3845ad1e93985beeff65b10ec97b23.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            450,
+            784,
+            464
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/9c4ac0f406f77133d4fb72a2bffd19daf953d57a3f3ef903f97d9543a1ade086.jpg",
+        "image_caption": [
+            "Figure 2. Overview of VFlowOpt. (1) During inference, VFlowOpt first assesses the importance of visual tokens, based on which progressive token pruning is performed. After the initial pruning stage, the pruned tokens are merged and recycled. The pruning strategy used in this process is defined by the (2) Optimization Stage. (a) The importance map is computed by combining the attention of relatively important tokens with the entropy of image patches. (b) The pruned tokens are grouped into grid cells, where each cell has a side length of a. Within each grid cell, the pruned tokens are fused using a weighted average, with their importance values as weights, and then recycled. (2) VFlowOpt optimizes the pruning strategy by minimizing the discrepancy of the last token in the final layer of the LMM with and without applying the pruning strategy.",
+            "Figure 3. The attention of redundant tokens (marked in red) fails to reflect the importance of other tokens and instead focuses on similar tokens, such as background elements."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            107,
+            575,
+            464,
+            713
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "tokens, Section 3.2 presents progressive pruning and token recycling strategies, and finally, Section 3.3 details how VFlowOpt customizes the superior pruning strategy (i.e., selects the pruning hyperparameters) for different LMMs.",
+        "bbox": [
+            89,
+            779,
+            482,
+            839
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Visual Token Importance Estimation",
+        "text_level": 1,
+        "bbox": [
+            89,
+            848,
+            410,
+            864
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For a general approach to evaluate visual token importance, the previous works [14, 41] propose that the importance of",
+        "bbox": [
+            89,
+            869,
+            483,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "visual tokens can be estimated using the average attention from all other tokens in ViTs. Although it has achieved some success, we find that redundant tokens (e.g., tokens corresponding to background regions) often assign disproportionately high attention to other similar redundant tokens (shown in Fig. 3), reducing the reliability of the importance estimation. To address this, we first identify relatively important tokens based on the attention they receive from all tokens (tokens receiving higher attention are considered relatively important), as shown in Fig. 2 (a). We then exclude redundant tokens and use the attention from these relatively important tokens as a more robust metric. Additionally, we incorporate the information entropy of image patches into the importance score to prioritize tokens corresponding to visually informative regions.",
+        "bbox": [
+            511,
+            577,
+            906,
+            803
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To identify relatively important visual tokens, we define a threshold:",
+        "bbox": [
+            511,
+            805,
+            906,
+            833
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\tau = t \\cdot \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} A _ {i j}, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            630,
+            834,
+            906,
+            877
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\tau$  is the threshold,  $A_{ij}$  is the attention weight from",
+        "bbox": [
+            513,
+            885,
+            906,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "23926",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "token  $i$  to token  $j$  within a ViT layer,  $N$  is the total number of tokens, and  $t$  is a sensitivity hyperparameter. We then treat tokens with a total attention exceeding  $\\tau$  as relatively important:",
+        "bbox": [
+            89,
+            90,
+            480,
+            150
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {K} = \\{j \\mid \\sum_ {i = 1} ^ {N} A _ {i j} > \\tau \\}. \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            204,
+            151,
+            480,
+            191
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Here,  $\\mathcal{K}$  denotes the set of indices of these relatively important tokens.",
+        "bbox": [
+            89,
+            196,
+            480,
+            224
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The information entropy of an image patch corresponding to token  $i$  is defined as:",
+        "bbox": [
+            89,
+            226,
+            480,
+            255
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nH \\left(V _ {i}\\right) = - \\sum_ {k = 0} ^ {L - 1} p _ {k} \\log p _ {k}, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            197,
+            266,
+            480,
+            306
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $V_{i}$  is the image patch corresponding to token  $i$ ,  $p_{k}$  is the proportion of pixels in the patch with gray level  $k$ , and  $L$  is the total number of possible gray levels (256 for 8-bit images). Here, the gray level is defined as the average of the RGB channel values, providing a scalar representation of pixel brightness. Higher entropy reflects greater diversity in pixel intensities and richer visual information.",
+        "bbox": [
+            89,
+            316,
+            482,
+            420
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The importance score of token  $i$  is computed by combining the attention it receives from tokens in  $\\mathcal{K}$  with the entropy of its corresponding image patch, normalized via the softmax function:",
+        "bbox": [
+            89,
+            422,
+            483,
+            481
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nI _ {i} = \\sum_ {k \\in \\mathcal {K}} A _ {k i} + \\alpha \\cdot \\frac {\\exp (H (V _ {i}))}{\\sum_ {j = 1} ^ {N} \\exp (H (V _ {j}))}, \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            145,
+            489,
+            480,
+            529
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $I_{i}$  denotes the importance score of token  $i$ ,  $A_{ki}$  is the attention weight from token  $k$  to token  $i$ ,  $H(V_{i})$  is the entropy of the image patch corresponding to token  $i$ ,  $\\alpha$  is a hyperparameter that determines the entropy term's contribution, and  $N$  is the total number of tokens. This approach mitigates biases from redundant tokens and leverages the visual information of image patches, thereby facilitating effective estimation of visual token importance.",
+        "bbox": [
+            89,
+            537,
+            482,
+            657
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Progressive Pruning and Token Recycling",
+        "text_level": 1,
+        "bbox": [
+            89,
+            667,
+            446,
+            683
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Previous studies [37, 51] indicate that visual tokens play a more critical role in the shallower layers of LMMs, whereas redundancy tends to increase in deeper layers. Based on this, we adopt a progressive pruning strategy. To balance the pruning process's simplicity with its fine-grained configuration, we evenly divide the LMM into three stages. At the beginning of each stage, a predetermined fraction of visual tokens with higher importance scores (as defined in Sec. 3.1) is retained according to the stage-specific retention ratios  $R = [R_{1}, R_{2}, R_{3}]$ , while the rest are pruned. Notably, the position IDs of visual tokens remain unchanged following pruning, preserving the original spatial structure of the visual input. Following the previous work [37], we compute the average visual token retention rate across the entire",
+        "bbox": [
+            89,
+            688,
+            482,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "LMM as follows:",
+        "bbox": [
+            513,
+            90,
+            633,
+            104
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\bar {R} = \\frac {R _ {1} \\cdot L _ {1} + R _ {1} \\cdot R _ {2} \\cdot L _ {2} + R _ {1} \\cdot R _ {2} \\cdot R _ {3} \\cdot L _ {3}}{L}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            532,
+            114,
+            903,
+            143
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $L_{1}, L_{2}$ , and  $L_{3}$  represent the number of layers in the three stages, respectively, and  $L$  denotes the total number of layers in the LMM.",
+        "bbox": [
+            511,
+            156,
+            903,
+            200
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To prevent the loss of any small but potentially significant information during the initial pruning process (before visual tokens are fed into the LLM), we propose a token merging and recycling strategy to compactly represent redundant visual information. Specifically, We use a square grid (with each cell having side length  $a$ ) to group pruned tokens that fall into the same grid cell, as shown in Fig. 2 (b). Within each cell, the pruned tokens are fused into a single token by computing a weighted average of their representations, using their importance scores as weights. This fused token then replaces the pruned token with the highest importance score in that cell and is incorporated into the set of retained tokens.",
+        "bbox": [
+            511,
+            202,
+            903,
+            397
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Formally, a token  $t_i$  with spatial coordinates  $(x_i, y_i)$  belongs to a grid cell  $\\mathcal{G}_{p,q}$  if",
+        "bbox": [
+            511,
+            398,
+            903,
+            431
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\lfloor x _ {i} / a \\rfloor = p \\quad \\text {a n d} \\quad \\lfloor y _ {i} / a \\rfloor = q, \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            601,
+            446,
+            903,
+            464
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $p$  and  $q$  are the row and column indices of the grid cell, respectively. Suppose there are  $k$  pruned tokens in  $\\mathcal{G}_{p,q}$  with corresponding importance scores  $I_{1}, I_{2}, \\ldots, I_{k}$  and token representations  $\\mathbf{t}_1, \\mathbf{t}_2, \\ldots, \\mathbf{t}_k$ . The fused token  $\\mathbf{t}_{\\mathrm{merged}}^{p,q}$  is computed as:",
+        "bbox": [
+            511,
+            472,
+            903,
+            547
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {t} _ {\\text {m e r g e d}} ^ {p, q} = \\frac {\\sum_ {i = 1} ^ {k} I _ {i} \\cdot \\mathbf {t} _ {i}}{\\sum_ {i = 1} ^ {k} I _ {i}}. \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            630,
+            561,
+            903,
+            601
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We then assign  $\\mathbf{t}_{\\mathrm{merged}}^{p,q}$  to the position of the pruned token with the highest importance score  $I_{\\mathrm{max}}$  in  $\\mathcal{G}_{p,q}$ , counting it among the retained tokens. By combining token pruning with this recycling strategy, we reduce token redundancy while avoiding potential information loss.",
+        "bbox": [
+            511,
+            608,
+            903,
+            685
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Pruning Strategy Optimization",
+        "text_level": 1,
+        "bbox": [
+            511,
+            696,
+            787,
+            713
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The pruning strategies described in Sec. 3.1 and Sec. 3.2 involve several key hyperparameters that directly affect the performance of LMMs after token pruning. Properly defining these hyperparameters is crucial for preserving model performance. However, existing approaches often rely on manually designed pruning strategies and apply the same strategy across different LMMs, without considering the unique characteristics of each model. This coarse-grained approach can lead to significant performance degradation due to its lack of fitness for different LMMs, leaving the task of designing pruning strategies tailored to specific LMMs as a formidable challenge.",
+        "bbox": [
+            511,
+            719,
+            906,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "23927",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Previous interpretability studies of LMMs [15, 53] offer crucial insights into their internal mechanisms. Specifically, these works reveal that in the lower and middle layers of LMMs, visual information from visual tokens is aggregated into the corresponding query text tokens. In the higher layers, the multimodal representation encoded in the query text tokens is further progressively propagated to the final position of the input sequence, ultimately influencing the subsequent inference process.",
+        "bbox": [
+            89,
+            90,
+            483,
+            226
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Inspired by this insight, we propose a framework that requires only a small amount of unlabeled data and leverages the internal flow of visual information within LMMs to search for the superior pruning strategy tailored to the characteristics of different LMMs, as shown in Fig. 2 (2). Specifically, we recast the task of designing the superior pruning strategy as an optimization problem, aiming to minimize the discrepancy in visual information flow with and without applying a pruning strategy to perform visual token pruning. In this framework, we treat the final token in the last layer as the representative outcome of the visual information flow. We use cosine similarity to measure the similarity of the final token representation with and without visual token pruning. A higher similarity indicates that the discrepancy between the visual information flow with and without pruning is smaller. Accordingly, we define the following optimization objective:",
+        "bbox": [
+            91,
+            227,
+            485,
+            484
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\max  _ {s \\in \\mathcal {S}} f (s) = \\operatorname {C o s i n e S i m} \\left(h _ {f}, g _ {s} \\left(h _ {f}\\right)\\right). \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            158,
+            496,
+            482,
+            517
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Here,  $h_f$  represents the representation of the final token in the last layer before pruning, and  $g_s(h_f)$  represents the final token representation after applying the pruning strategy  $s$  (with  $g_s$  modeling its effect on the token feature), and  $S$  is the solution space.",
+        "bbox": [
+            89,
+            529,
+            483,
+            604
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To efficiently search for the superior pruning strategy  $s$ , we employ Bayesian optimization, which systematically explores the hyperparameter space—including the threshold sensitivity  $t$ , the entropy weight  $\\alpha$ , the grid size  $a$ , and the pruning ratios  $R_{1}, R_{2}$ , and  $R_{3}$ —to maximize the target function  $f(s)$ . In short, Bayesian optimization constructs a surrogate model to approximate the objective function and employs an acquisition function to balance exploration and exploitation, thereby efficiently guiding the search for promising hyperparameter settings. The details of this optimization process are presented in Algorithm 1.",
+        "bbox": [
+            89,
+            604,
+            483,
+            771
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            89,
+            784,
+            223,
+            801
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we evaluate our approach on various LMMs across diverse image and video benchmarks, followed by an efficiency analysis and an ablation study of each component. Finally, we illustrate how our method affects discrepancies in the visual information flow with and without pruning, offering deeper insights into our approach.",
+        "bbox": [
+            89,
+            809,
+            483,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Algorithm 1 VFlowOpt with Bayesian Optimization",
+        "bbox": [
+            514,
+            90,
+            864,
+            106
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Input: LMM  $\\theta$ , data samples  $\\mathcal{D}$ , number of Bayesian optimization iterations  $T$ , computation budget  $\\overline{R}$ , target function  $f(\\cdot)$ : the sum of cosine similarities between the last token in the last layer of the LMM  $\\theta$  with and without visual token pruning, computed over all data samples  $\\mathcal{D}$ .  \nOutput: retention rates  $R_{1}, R_{2}, R_{3}$ ; threshold sensitivity  $t$ ; entropy weight in importance score  $\\alpha$ ; grid size  $a$ .",
+        "bbox": [
+            514,
+            109,
+            906,
+            229
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1: Initialize a Gaussian Process model  $\\mathcal{GP}$",
+            "2: Define the acquisition function  $A(\\cdot)$  (Expected Improvement is adopted)",
+            "3: Uniformly sample initial points:"
+        ],
+        "bbox": [
+            524,
+            231,
+            903,
+            291
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nX _ {0} = \\left\\{\\left(R _ {1}, R _ {2}, t, \\alpha , a\\right) \\mid R _ {v}, R _ {1}, t, \\alpha , a \\in \\text {v a l i d r a n g e s} \\right\\}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            540,
+            303,
+            915,
+            320
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4: for all  $x \\in X_0$  do",
+        "bbox": [
+            524,
+            332,
+            663,
+            345
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "5: Calculate  $R_{3}$  using the constraint:  $\\overline{R} = (R_{1} \\cdot L_{1} + R_{1} \\cdot R_{2} \\cdot L_{2} + R_{1} \\cdot R_{2} \\cdot R_{3} \\cdot L_{3}) / L$",
+            "6: Form pruning strategy:  $s = (R_1, R_2, R_3, t, \\alpha, a)$",
+            "7: Evaluate the target function  $f(s \\mid \\theta, \\mathcal{D})$"
+        ],
+        "bbox": [
+            524,
+            347,
+            903,
+            407
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "8: end for",
+        "bbox": [
+            524,
+            407,
+            596,
+            420
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "9: for  $n = 0$  to  $T - 1$  do",
+        "bbox": [
+            524,
+            422,
+            692,
+            436
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "10: Fit  $\\mathcal{GP}$  to the observed data  $(S_{n},f(S_{n}\\mid \\theta ,\\mathcal{D}))$",
+            "11: Select the next point:  $x_{n + 1}\\gets \\arg \\max_xA(x;\\mathcal{GP})$",
+            "12: Calculate  $R_{3}$  using the constraint:  $\\overline{R} = (R_{1} \\cdot L_{1} + R_{1} \\cdot R_{2} \\cdot L_{2} + R_{1} \\cdot R_{2} \\cdot R_{3} \\cdot L_{3}) / L$",
+            "13: Form pruning strategy:  $s_{n+1} = (R_1, R_2, R_3, t, \\alpha, a)_{n+1}$",
+            "14: Update  $S_{n + 1}\\gets S_n\\cup \\{s_{n + 1}\\}$",
+            "15: end for",
+            "16: return  $(R_{1}, R_{2}, R_{3}, t, \\alpha, a)^{*}$  that maximize  $f$"
+        ],
+        "bbox": [
+            519,
+            436,
+            903,
+            558
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Experiment Setting",
+        "text_level": 1,
+        "bbox": [
+            513,
+            590,
+            699,
+            606
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Datasets. We evaluate our method on ten image-based multimodal benchmarks: GQA [1], VizWiz [13], ScienceQA-IMG [26], TextVQA [30], ChartQA [27], POPE [19], MME [10], MMBench [25], MMStar [6], and DocVQA [28]. For video understanding, we adopt two datasets—SeedBench (video) [16] and VideoMME [11], where VideoMME is partitioned by video length into short, medium, and long subsets. Further details are presented in the Appendix.",
+        "bbox": [
+            511,
+            613,
+            906,
+            750
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Model Architectures. We integrate VFlowOpt into multiple LMMs, including LLaVA-OneVision-7B, LLaVA-NeXT-7B [23], and Qwen2-VL-7B [34]. These models employ various vision encoders (SigLIP, CLIP, and a ViT designed for Qwen2-VL) and LLM backbones (Qwen2-7B and Vicuna-7B). To prevent out-of-memory issues in Qwen2-VL, we set maxPixels = 3000000.",
+        "bbox": [
+            511,
+            750,
+            905,
+            854
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Comparison Methods. We compare VFlowOpt against five baseline methods: FastV [7], FitPrune [42], PDrop [36], SparseVLM [52], and VisionZip [40]. FastV,",
+        "bbox": [
+            511,
+            854,
+            903,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "23928",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/c08e4252dfd62ec3e56072a83d39363f69a64d4c10b6a680c0e600cc68ee8b4b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>MMStar</td><td>MME</td><td>MMB</td><td>SQA</td><td>POPE</td><td>GQA</td><td>VizWiz</td><td>\\(VQA^{Text}\\)</td><td>ChartQA</td><td>DocVQA</td><td>Avg.</td></tr><tr><td colspan=\"12\">Upper Bound, 100% Tokens</td></tr><tr><td>Vanilla</td><td>61.7</td><td>1581</td><td>80.8</td><td>95.8</td><td>89.1</td><td>62.2</td><td>60.4</td><td>76.0</td><td>80.0</td><td>87.5</td><td>100%</td></tr><tr><td colspan=\"12\">Retain 50% Tokens (↓ 50%)</td></tr><tr><td>FastV (ECCV24)</td><td>58.9</td><td>1549</td><td>79.4</td><td>92.8</td><td>87.9</td><td>61.5</td><td>61.1</td><td>72.5</td><td>68.6</td><td>84.0</td><td>96.5%</td></tr><tr><td>FitPrune (AAAI25)</td><td>4.5</td><td>96.6</td><td>1.0</td><td>8.6</td><td>18.8</td><td>16.2</td><td>6.8</td><td>37.6</td><td>37.7</td><td>30.7</td><td>21.4%</td></tr><tr><td>PDrop (CVPR25)</td><td>37.1</td><td>628.6</td><td>32.1</td><td>64.8</td><td>81.3</td><td>44.2</td><td>48.9</td><td>63.7</td><td>62.7</td><td>62.2</td><td>71.3%</td></tr><tr><td>SparseVLM (2024.10)</td><td>59.8</td><td>1577</td><td>80.5</td><td>94.1</td><td>88.1</td><td>61.9</td><td>60.4</td><td>73.9</td><td>70.5</td><td>80.8</td><td>97.1%</td></tr><tr><td>VisionZip (CVPR25)</td><td>60.4</td><td>1587</td><td>80.3</td><td>94.6</td><td>89.3</td><td>62.7</td><td>59.8</td><td>74.2</td><td>75.4</td><td>88.4</td><td>98.9%</td></tr><tr><td>VFlowOpt</td><td>61.3</td><td>1591</td><td>81.1</td><td>95.4</td><td>89.4</td><td>62.4</td><td>60.0</td><td>75.1</td><td>77.8</td><td>90.0</td><td>99.9%</td></tr><tr><td colspan=\"12\">Retain 25% Tokens (↓ 75%)</td></tr><tr><td>FastV (ECCV24)</td><td>54.0</td><td>1539</td><td>77.0</td><td>88.6</td><td>83.8</td><td>58.2</td><td>61.0</td><td>58.3</td><td>42.7</td><td>62.9</td><td>86.4%</td></tr><tr><td>FitPrune (AAAI25)</td><td>5.2</td><td>92.6</td><td>1.1</td><td>5.9</td><td>20.2</td><td>14.5</td><td>5.1</td><td>29.7</td><td>27.3</td><td>23.2</td><td>17.6%</td></tr><tr><td>PDrop (CVPR25)</td><td>36.8</td><td>574.4</td><td>31.3</td><td>62.8</td><td>78.0</td><td>42.2</td><td>47.1</td><td>56.3</td><td>47.5</td><td>47.0</td><td>65.2%</td></tr><tr><td>SparseVLM (2024.10)</td><td>56.6</td><td>1520</td><td>78.7</td><td>90.3</td><td>87.2</td><td>59.7</td><td>60.8</td><td>66.3</td><td>54.0</td><td>66.6</td><td>90.5%</td></tr><tr><td>VisionZip (CVPR25)</td><td>54.6</td><td>1562</td><td>78.9</td><td>90.4</td><td>88.8</td><td>61.0</td><td>60.4</td><td>70.0</td><td>66.3</td><td>79.6</td><td>94.3%</td></tr><tr><td>VFlowOpt</td><td>57.8</td><td>1570</td><td>79.9</td><td>92.3</td><td>89.1</td><td>61.2</td><td>60.4</td><td>72.5</td><td>69.1</td><td>82.3</td><td>96.3%</td></tr><tr><td colspan=\"12\">Retain 10% Tokens (↓ 90%)</td></tr><tr><td>FastV (ECCV24)</td><td>46.0</td><td>1209</td><td>70.1</td><td>81.7</td><td>77.0</td><td>51.5</td><td>56.4</td><td>35.6</td><td>21.3</td><td>33.2</td><td>69.7%</td></tr><tr><td>FitPrune (AAAI25)</td><td>5.4</td><td>27.0</td><td>0.9</td><td>3.5</td><td>2.3</td><td>4.6</td><td>0.7</td><td>6.9</td><td>5.9</td><td>6.9</td><td>5.1%</td></tr><tr><td>PDrop (CVPR25)</td><td>31.8</td><td>456.3</td><td>22.5</td><td>59.0</td><td>65.9</td><td>35.7</td><td>43.3</td><td>33.7</td><td>19.6</td><td>18.5</td><td>49.9%</td></tr><tr><td>SparseVLM (2024.10)</td><td>45.1</td><td>1191</td><td>71.8</td><td>83.7</td><td>80.0</td><td>54.6</td><td>56.4</td><td>39.8</td><td>37.6</td><td>39.6</td><td>74.0%</td></tr><tr><td>VisionZip (CVPR25)</td><td>49.5</td><td>1389</td><td>74.8</td><td>86.2</td><td>86.1</td><td>57.2</td><td>56.8</td><td>56.4</td><td>46.1</td><td>49.0</td><td>82.1%</td></tr><tr><td>VFlowOpt</td><td>52.0</td><td>1464</td><td>75.1</td><td>88.4</td><td>85.2</td><td>57.3</td><td>57.3</td><td>60.2</td><td>53.6</td><td>56.1</td><td>85.5%</td></tr></table>",
+        "bbox": [
+            99,
+            88,
+            898,
+            417
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Performance comparison on LLaVA-OneVision-7B under different token retention conditions. \"Avg.\" refers to average accuracy on 10 benchmarks. For each reduction ratio, the best average performance is shown in bold.",
+        "bbox": [
+            89,
+            424,
+            906,
+            454
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "FitPrune, PDrop, and SparseVLM rely on text-visual attention in the LLM to prune visual tokens but differ as follows: FastV performs a single pruning step after the second LLM layer; FitPrune prunes tokens in each layer according to predefined, layer-specific ratios; PDrop conducts four pruning steps within the LLM, using the same retention ratio each time; and SparseVLM uses the attention weights of preselected text tokens to evaluate the importance of visual tokens. VisionZip, in contrast, determines token importance based on the [cls] token's attention; for models lacking a [cls] token, we follow VisionZip's original procedure by computing the average attention each token receives from every other token in the ViTs.",
+        "bbox": [
+            88,
+            474,
+            482,
+            671
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. For LLaVA-OneVision-7B and Qwen2-VL-7B, we perform token pruning at three distinct points: before the LLM, and after the 9th and 18th layers. For LLaVA-NeXT-7B, pruning is conducted before the LLM, and again after the 10th and 20th layers. During optimization, we sample 30 unlabeled instances from each model's training datasets; for models without publicly available training dataset, we instead use random samples from the LLaVA-OneVision training set. The optimization is performed for a total of 50 iterations. For LMMs that modify visual tokens output by the vision encoder (e.g., unpadding and interpolation in LLaVA-OneVision, and unpadding in LLaVA-NeXT), we apply corresponding transformations to the importance maps so they remain fully aligned with the final visual tokens.",
+        "bbox": [
+            89,
+            671,
+            482,
+            898
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Image Understanding Tasks",
+        "text_level": 1,
+        "bbox": [
+            511,
+            474,
+            764,
+            491
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We evaluate the proposed VFlowOpt on image understanding benchmarks with LLaVA-OneVision-7B using various pruning ratios, and present the results in Tab. 1. Compared to other baselines, VFlowOpt consistently maintains superior accuracy across different levels of token pruning. With  $50\\%$  token retention, VFlowOpt achieves  $99.9\\%$  of the original performance, exceeding the second-best approach by  $1.0\\%$ . This negligible performance drop underscores the method's strong potential for practical deployment. Under more extreme pruning conditions with very few retained tokens, VFlowOpt's advantage becomes more pronounced. When only  $10\\%$  of the visual tokens are retained, VFlowOpt preserves  $85.5\\%$  of the original performance, surpassing the second-best approach by  $3.3\\%$ . This finding suggests that VFlowOpt can effectively leverage limited visual information to maintain high performance under strict computational budgets. When applied to LLaVA-OneVision, both FitPrune and PDrop disrupt the model's normal output, indicating their inability to devise effective pruning strategies for different architectures. In contrast, VFlowOpt demonstrates a clear advantage by tailoring its pruning strategy to each model's unique characteristics.",
+        "bbox": [
+            509,
+            502,
+            906,
+            835
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To further validate VFlowOpt's generalization capability, we evaluate it on LLaVA-NeXT-7B and Qwen2-VL-7B. As shown in Tab. 2 and Tab. 3, VFlowOpt once again delivers the top results on these LMMs. When retaining  $25\\%$",
+        "bbox": [
+            511,
+            839,
+            908,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "23929",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/89770b1edab563629f5768444b850141eae77d84f4414fa572579c778afb191b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>MMStar</td><td>MME</td><td>MMB</td><td>SQA</td><td>POPE</td><td>GQA</td><td>Avg.</td></tr><tr><td colspan=\"8\">Upper Bound, 100% Tokens</td></tr><tr><td>Vanilla</td><td>37.6</td><td>1519</td><td>67.4</td><td>70.1</td><td>86.5</td><td>64.2</td><td>100%</td></tr><tr><td colspan=\"8\">Retain 25% Tokens (↓ 75%)</td></tr><tr><td>FastV</td><td>35.1</td><td>1477</td><td>65.6</td><td>67.4</td><td>83.1</td><td>60.4</td><td>95.7%</td></tr><tr><td>VisionZip</td><td>35.8</td><td>1501</td><td>65.4</td><td>67.9</td><td>86.7</td><td>61.5</td><td>97.3%</td></tr><tr><td>VFlowOpt</td><td>37.0</td><td>1514</td><td>67.0</td><td>67.7</td><td>87.6</td><td>62.6</td><td>98.8%</td></tr><tr><td colspan=\"8\">Retain 10% Tokens (↓ 90%)</td></tr><tr><td>FastV</td><td>29.2</td><td>1282</td><td>61.6</td><td>63.8</td><td>71.7</td><td>55.9</td><td>85.7%</td></tr><tr><td>VisionZip</td><td>32.6</td><td>1378</td><td>61.5</td><td>67.1</td><td>83.5</td><td>57.0</td><td>91.6%</td></tr><tr><td>VFlowOpt</td><td>35.1</td><td>1393</td><td>62.9</td><td>67.4</td><td>83.6</td><td>57.3</td><td>93.4%</td></tr></table>",
+        "bbox": [
+            96,
+            89,
+            486,
+            243
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/36abdb835487ebf46093e346e5ae6ca0c9428afb043cb14d96319d2e489f57f2.jpg",
+        "table_caption": [
+            "Table 2. Comparative experiments on LLaVA-NeXT-7B."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>MMStar</td><td>MME</td><td>MMB</td><td>SQA</td><td>POPE</td><td>GQA</td><td>Avg.</td></tr><tr><td colspan=\"8\">Upper Bound, 100% Tokens</td></tr><tr><td>Vanilla</td><td>57.5</td><td>1680</td><td>80.3</td><td>84.7</td><td>88.4</td><td>62.2</td><td>100%</td></tr><tr><td colspan=\"8\">Retain 25% Tokens (↓ 75%)</td></tr><tr><td>FastV</td><td>54.5</td><td>1597</td><td>76.3</td><td>79.0</td><td>81.8</td><td>57.2</td><td>93.8</td></tr><tr><td>VisionZip</td><td>55.2</td><td>1618</td><td>78.9</td><td>81.3</td><td>86.9</td><td>60.0</td><td>96.8</td></tr><tr><td>VFlowOpt</td><td>55.9</td><td>1659</td><td>79.8</td><td>81.5</td><td>87.1</td><td>60.4</td><td>97.8</td></tr><tr><td colspan=\"8\">Retain 10% Tokens (↓ 90%)</td></tr><tr><td>FastV</td><td>44.9</td><td>1405</td><td>70.0</td><td>75.5</td><td>75.8</td><td>51.6</td><td>84.4</td></tr><tr><td>VisionZip</td><td>49.3</td><td>1518</td><td>76.5</td><td>78.2</td><td>84.3</td><td>54.1</td><td>90.9</td></tr><tr><td>VFlowOpt</td><td>51.0</td><td>1591</td><td>78.0</td><td>78.6</td><td>84.5</td><td>54.8</td><td>92.8</td></tr></table>",
+        "bbox": [
+            96,
+            271,
+            482,
+            426
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/40d07dfea5d2b525bc1b5d6e804cc5d1b171db9c191f8b647fb6c7255033c3b0.jpg",
+        "table_caption": [
+            "Table 3. Comparative experiments on Qwen2-VL-7B."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">SeedBench (video)</td><td colspan=\"3\">VideoMME</td><td rowspan=\"2\">Avg.</td></tr><tr><td>S</td><td>M</td><td>L</td></tr><tr><td colspan=\"6\">Upper Bound, 100% Tokens</td></tr><tr><td>Vanilla</td><td>56.9</td><td>70.5</td><td>54.6</td><td>49.5</td><td>100%</td></tr><tr><td colspan=\"6\">Retain 25% Tokens (↓75%)</td></tr><tr><td>FastV</td><td>54.7</td><td>66.9</td><td>53.2</td><td>47.7</td><td>96.2%</td></tr><tr><td>VisionZip</td><td>56.4</td><td>68.3</td><td>55.3</td><td>49.0</td><td>99.0%</td></tr><tr><td>VFlowOpt</td><td>56.8</td><td>68.9</td><td>55.8</td><td>49.2</td><td>100%</td></tr><tr><td colspan=\"6\">Retain 10% Tokens (↓90%)</td></tr><tr><td>FastV</td><td>48.7</td><td>53.7</td><td>47.6</td><td>42.3</td><td>83.6%</td></tr><tr><td>VisionZip</td><td>55.0</td><td>59.7</td><td>51.8</td><td>46.3</td><td>92.4%</td></tr><tr><td>VFlowOpt</td><td>55.5</td><td>63.3</td><td>52.7</td><td>48.6</td><td>95.5%</td></tr></table>",
+        "bbox": [
+            96,
+            452,
+            480,
+            625
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Comparative experiments on video understanding tasks.",
+        "bbox": [
+            94,
+            627,
+            477,
+            641
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "of the tokens, VFlowOpt preserves  $98.8\\%$  of the original performance on LLaVA-NeXT-7B and  $97.8\\%$  on Qwen2-VL-7B. Even under the stringent condition of retaining only  $10\\%$  of the tokens, VFlowOpt maintains  $93.4\\%$  of the performance on LLaVA-NeXT-7B and  $92.8\\%$  on Qwen2-VL-7B, demonstrating its strong generalizability.",
+        "bbox": [
+            89,
+            657,
+            482,
+            748
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3. Generalization to Video Tasks",
+        "text_level": 1,
+        "bbox": [
+            89,
+            757,
+            356,
+            771
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We further explore VFlowOpt's generalization across different modalities by evaluating it on video benchmarks using LLaVA-OneVision-7B. As shown in Tab. 4, VFlowOpt surpasses other baselines under various token retention ratios. Notably, when only  $25\\%$  of the tokens are preserved, VFlowOpt incurs virtually no performance loss on LLaVA-OneVision-7B, maintaining  $100\\%$  of its original performance. Even under the more stringent condition of retain",
+        "bbox": [
+            89,
+            779,
+            482,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "ing just  $10\\%$  of the tokens, VFlowOpt still retains  $95.5\\%$  of the original performance, underscoring its robust generalization in the video domain.",
+        "bbox": [
+            511,
+            90,
+            906,
+            136
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4. Efficiency Analysis",
+        "text_level": 1,
+        "bbox": [
+            511,
+            147,
+            694,
+            165
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We demonstrate the efficiency of VFlowOpt by conducting a comparative study on LLaVA-OneVision-7B running on a single NVIDIA A100-SXM4-80GB GPU, focusing on FLOPs, KV-Cache memory usage, and inference latency. The efficiency analysis of other baseline methods is provided in the Appendix.",
+        "bbox": [
+            511,
+            170,
+            903,
+            261
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We assess various token pruning ratios on the MME benchmark, measuring overall performance and average efficiency metrics across all samples. As shown in Tab. 5, VFlowOpt significantly enhances LLaVA-OneVision-7B's computational efficiency by reducing FLOPs, shrinking KV-Cache memory usage, and accelerating inference speed. Notably, at a  $50\\%$  token pruning ratio, VFlowOpt achieves a  $49.5\\%$  reduction in KV-Cache memory and a  $1.8\\times$  speedup in inference, while boosting performance by  $0.7\\%$ . At a more aggressive  $75\\%$  pruning ratio, we observe a  $74.2\\%$  reduction in KV-Cache memory and a  $3.1\\times$  speedup, with only a  $0.7\\%$  performance drop. Under extremely tight computational budgets, pruning  $90\\%$  of visual tokens reduces the KV-Cache memory footprint by  $89\\%$ , accelerates inference by  $3.8\\times$ , and degrades performance by only  $7.4\\%$ .",
+        "bbox": [
+            511,
+            261,
+            906,
+            487
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Such reductions in inference latency substantially benefit user-facing applications demanding real-time performance, such as autonomous driving and robotics. Moreover, by significantly shrinking the KV-Cache, large-batch inference on LMMs can accommodate more user requests simultaneously, thereby substantially reducing overall inference costs. In summary, VFlowOpt uses far less GPU memory and delivers faster inference while preserving model performance, offering a highly practical solution for efficiently deploying LMMs in real-world scenarios.",
+        "bbox": [
+            511,
+            488,
+            908,
+            640
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.5. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            511,
+            650,
+            666,
+            666
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To confirm the contributions of each component in our token pruning strategy, we conduct ablation experiments on LLaVA-OneVision-7B under a computational budget of retaining  $25\\%$  of tokens. As shown in Tab. 6, removing importance calibration (i.e., directly using the mean attention over all tokens received by each visual token as its importance score), omitting token recycling, or discarding progressive pruning (i.e., maintaining the same number of tokens at each layer) leads to noticeable performance degradation across MMStar, MMBench, and SQA. By contrast, the complete VFlowOpt method consistently exhibits minimal performance loss, highlighting the effectiveness of all its components in pruning redundant tokens while preserving essential visual information.",
+        "bbox": [
+            511,
+            672,
+            906,
+            883
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We further examine how the number of samples and the",
+        "bbox": [
+            532,
+            885,
+            903,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "23930",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/3af5b2956741a6de8945bd4b3d799f7aa67ca963a5c5c8b6c2c586d111305f97.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>Token Reduction</td><td>FLOPs ↓ (T)</td><td>Δ</td><td>Latency ↓ (ms)</td><td>Δ</td><td>KV Cache ↓ (MB)</td><td>Δ</td><td>Performance ↑</td><td>Δ</td></tr><tr><td>LLaVA-OneVision-7B</td><td>-</td><td>71.4</td><td>-</td><td>1040.1</td><td>-</td><td>1786.4</td><td>-</td><td>1581</td><td>-</td></tr><tr><td rowspan=\"3\">+ VFlowOpt</td><td>50%</td><td>37.2</td><td>-48.0%</td><td>584.2</td><td>-43.8%</td><td>902.8</td><td>-49.5%</td><td>1591</td><td>+0.6%</td></tr><tr><td>75%</td><td>19.1</td><td>-73.2%</td><td>328.5</td><td>-68.4%</td><td>460.6</td><td>-74.2%</td><td>1570</td><td>-0.7%</td></tr><tr><td>90%</td><td>7.7</td><td>-89.2%</td><td>272.1</td><td>-73.8%</td><td>197.1</td><td>-89.0%</td><td>1464</td><td>-7.4%</td></tr></table>",
+        "bbox": [
+            143,
+            88,
+            854,
+            180
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/7df6f58322eeff0d18022495fcc373c29148d80151a63306e98a2dbebc3eff21.jpg",
+        "table_caption": [
+            "Table 5. Efficiency analysis of LLaVA-OneVision-7B with VFlowOpt. The detailed metric includes computation (FLOPs), latency, and KV-Cache memory.  $(\\Delta)$  denotes the reduction ratio."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Pruning Strategy</td><td>MMStar</td><td>MMBench</td><td>SQA</td></tr><tr><td>VFlowOpt</td><td>57.8</td><td>79.9</td><td>92.3</td></tr><tr><td>w/o Importance Calibration</td><td>56.2</td><td>79.4</td><td>91.8</td></tr><tr><td>w/o Token Merging</td><td>57.6</td><td>79.8</td><td>91.9</td></tr><tr><td>w/o Progressive Pruning</td><td>56.0</td><td>79.1</td><td>91.0</td></tr></table>",
+        "bbox": [
+            117,
+            232,
+            455,
+            313
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6. Ablation studies of pruning strategy components.",
+        "bbox": [
+            114,
+            319,
+            457,
+            335
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/2747e4b8f4cda23aeadfcdea5c5eba1512cdc543528df07152841897b6e5eb9c.jpg",
+        "image_caption": [
+            "Figure 4. Relationships of sample size (left) and optimization steps (right) with optimization time and final performance."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            343,
+            475,
+            452
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "number of optimization iterations in the Bayesian optimization procedure influence both optimization time and final performance. As illustrated in Fig. 4, the optimization time increases linearly with the number of samples and steps. Using 30 samples and 50 steps strikes an effective balance between time efficiency and performance, requiring only about 30 minutes to reach an exceptional result. More ablation results are provided in the supplementary material.",
+        "bbox": [
+            89,
+            506,
+            483,
+            628
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.6. Visual Information Flow Analysis",
+        "text_level": 1,
+        "bbox": [
+            89,
+            636,
+            385,
+            652
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We further investigate how Bayesian optimization formulates the superior pruning strategy by examining its impact on visual information flow. Specifically, we compare the randomly initialized pruning strategy derived from the Bayesian optimization process with its final optimized strategy, measuring their respective effects on the text tokens that follow the visual tokens at each LLM layer. Because these text tokens can receive information from the visual tokens via the LLM's unidirectional attention mechanism, examining differences in these tokens with and without pruning effectively reveals how visual information flow is altered. As shown in Fig. 5, under the same computational budget, the initial strategy causes a pronounced discrepancy between text token representations with and without pruning, indicating that it significantly disrupts the visual information flow. In contrast, when using the optimized strategy,",
+        "bbox": [
+            89,
+            659,
+            482,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/d2a427acb1d1889fcb27ddbe93caecc35d46160cebc60341d0b621e7434e3262.jpg",
+        "image_caption": [
+            "Figure 5. Visualization of the discrepancies in visual information flow corresponding to the pruning strategies before and after optimization, along with an example of visual question answering."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            232,
+            898,
+            484
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "the text tokens remain much more similar to those observed without visual token pruning. This important finding suggests that the optimization process yields a pruning strategy that preserves crucial visual information by minimizing the difference in visual information flow with and without pruning, thereby retaining LMM performance to the greatest extent possible.",
+        "bbox": [
+            511,
+            542,
+            906,
+            647
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            511,
+            661,
+            633,
+            678
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we address the pressing challenge of computational inefficiency in LMMs by introducing a general framework for visual token pruning. We reformulate pruning as an optimization problem focused on minimizing the divergence of visual information flow with and without pruning, thereby facilitating a tailored pruning strategy for diverse LMMs. By leveraging calibrated attention in ViTs to evaluate token significance and merging pruned tokens using importance-based weighting, we preserve critical information during inference without requiring retraining or manual tuning. Extensive experiments underscore the efficiency and generalizability of our approach, making it a practical solution for deploying LMMs in real-world settings, particularly in resource-constrained environments.",
+        "bbox": [
+            511,
+            686,
+            906,
+            898
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "23931",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            91,
+            90,
+            256,
+            107
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This research was supported by Shanghai Artificial Intelligence Laboratory.",
+        "bbox": [
+            89,
+            116,
+            480,
+            148
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            91,
+            165,
+            187,
+            181
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Joshua Ainslie, James Lee-Thorp, Michiel De Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245, 2023. 5",
+            "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 1, 2",
+            "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1, 2",
+            "[4] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, Xiaoyi Dong, Haodong Duan, Qi Fan, Zhaoye Fei, Yang Gao, Jiaye Ge, Chenya Gu, Yuzhe Gu, Tao Gui, Aijia Guo, Qipeng Guo, Conghui He, Yingfan Hu, Ting Huang, Tao Jiang, Penglong Jiao, Zhenjiang Jin, Zhikai Lei, Jiaxing Li, Jingwen Li, Linyang Li, Shuaibin Li, Wei Li, Yinling Li, Hongwei Liu, Jiangning Liu, Jiawei Hong, Kaiwen Liu, Kuikun Liu, Xiaoran Liu, Chengqi Lv, Haijun Lv, Kai Lv, Li Ma, Runyuan Ma, Zerun Ma, Wenchang Ning, Linke Ouyang, Jiantao Qiu, Yuan Qu, Fukai Shang, Yunfan Shao, Demin Song, Zifan Song, Zhihao Sui, Peng Sun, Yu Sun, Huanze Tang, Bin Wang, Guoteng Wang, Jiaqi Wang, Jiayu Wang, Rui Wang, Yudong Wang, Ziyi Wang, Xingjian Wei, Qizhen Weng, Fan Wu, Yingtong Xiong, Chao Xu, Ruiliang Xu, Hang Yan, Yirong Yan, Xiaogui Yang, Haochen Ye, Huaiyuan Ying, Jia Yu, Jing Yu, Yuhang Zang, Chuyu Zhang, Li Zhang, Pan Zhang, Peng Zhang, Ruijie Zhang, Shuo Zhang, Songyang Zhang, Wenjian Zhang, Wenwei Zhang, Xingcheng Zhang, Xinyue Zhang, Hui Zhao, Qian Zhao, Xiaomeng Zhao, Fengzhe Zhou, Zaida Zhou, Jingming Zhuo, Yicheng Zou, Xipeng Qiu, Yu Qiao and Dahua Lin. Internl m2 technical report, 2024. 2",
+            "[5] Jieneng Chen, Luoxin Ye, Ju He, Zhao-Yang Wang, Daniel Khashabi, and Alan Yuille. Llavolta: Efficient multi-modal models via stage-wise visual context compression. arXiv preprint arXiv:2406.20092, 2024. 2",
+            "[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 2, 5",
+            "[7] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for"
+        ],
+        "bbox": [
+            99,
+            190,
+            480,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 2, 5",
+            "[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 1, 2",
+            "[9] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, Tianren Gao, Erlong Li, Kun Tang, Zhipeng Cao, Tong Zhou, Ao Liu, Xinrui Yan, Shuqi Mei, Jianguo Cao, Ziran Wang, and Chao Zheng. A survey on multimodal large language models for autonomous driving, 2023. 1",
+            "[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 5",
+            "[11] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5",
+            "[12] 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, 2017. 1, 2",
+            "[13] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 5",
+            "[14] Lianyu Hu, Fanhua Shang, Liang Wan, and Wei Feng. illava: An image is worth fewer than 1/3 input tokens in large multimodal models. arXiv preprint arXiv:2412.06263, 2024. 3",
+            "[15] Omri Kaduri, Shai Bagon, and Tali Dekel. What's in the image? a deep-dive into the vision of vision language models. arXiv preprint arXiv:2411.17491, 2024. 5",
+            "[16] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5",
+            "[17] 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, 2",
+            "[18] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation, 2023. 1",
+            "[19] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 5",
+            "[20] Yuelei Li, Ge Yan, Annabella Macaluso, Mazeyu Ji, Xueyan Zou, and Xiaolong Wang. Integrating lmm planners and 3d"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "23932",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "skill policies for generalizable manipulation. arXiv preprint arXiv:2501.18733, 2025. 1",
+            "[21] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2",
+            "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2",
+            "[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5",
+            "[24] Jiaming Liu, Chenxuan Li, Guanqun Wang, Lily Lee, Kaichen Zhou, Sixiang Chen, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, and Shanghang Zhang. Self-corrected multimodal large language model for end-to-end robot manipulation, 2024. 1",
+            "[25] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 5",
+            "[26] 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. 5",
+            "[27] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5",
+            "[28] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5",
+            "[29] 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",
+            "[30] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5",
+            "[31] Haoxiang Tian, Xingshuo Han, Guoquan Wu, Yuan Zhou, Shuo Li, Jun Wei, Dan Ye, Wei Wang, and Tianwei Zhang. An llm-enhanced multi-objective evolutionary search for autonomous driving test scenario generation. arXiv preprint arXiv:2406.10857, 2024. 1",
+            "[32] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn,"
+        ],
+        "bbox": [
+            91,
+            90,
+            483,
+            901
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 2",
+            "[33] Ao Wang, Fengyuan Sun, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. [cls] token tells everything needed for training-free efficient mllms. arXiv preprint arXiv:2412.05819, 2024. 1, 2",
+            "[34] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 5",
+            "[35] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, Hao Tian, Lewei Lu, Xizhou Zhu, Xiaogang Wang, Yu Qiao, and Jifeng Dai. Drivemlm: Aligning multimodal large language models with behavioral planning states for autonomous driving, 2023. 1",
+            "[36] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 1, 2, 5",
+            "[37] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 4",
+            "[38] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2",
+            "[39] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin,"
+        ],
+        "bbox": [
+            516,
+            90,
+            903,
+            901
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "23933",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 2",
+            "[40] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 1, 2, 5",
+            "[41] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 3",
+            "[42] Weihao Ye, Qiong Wu, Wenhao Lin, and Yiyi Zhou. Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. arXiv preprint arXiv:2409.10197, 2024. 1, 2, 5",
+            "[43] Xubing Ye, Yukang Gan, Yixiao Ge, Xiao-Ping Zhang, and Yansong Tang. Atp-llava: Adaptive token pruning for large vision language models. arXiv preprint arXiv:2412.00447, 2024. 2",
+            "[44] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 1, 2",
+            "[45] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark, 2024. 1",
+            "[46] 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. 2",
+            "[47] Jun Zhang, Desen Meng, Ji Qi, Zhenpeng Huang, Tao Wu, and Limin Wang. p-mod: Building mixture-of-depths mllms via progressive ratio decay. arXiv preprint arXiv:2412.04449, 2024. 2",
+            "[48] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 2",
+            "[49] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster, 2024. 1, 2",
+            "[50] Shaolei Zhang, Qingkai Fang, Zhe Yang, and Yang Feng. Llava-mini: Efficient image and video large multimodal models with one vision token. In The Thirteenth International Conference on Learning Representations. 2",
+            "[51] Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu,"
+        ],
+        "bbox": [
+            91,
+            90,
+            480,
+            900
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Hao Tang, and Jieping Ye. From redundancy to relevance: Information flow in lvlms across reasoning tasks. arXiv preprint arXiv:2406.06579, 2024. 4",
+            "[52] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 1, 5",
+            "[53] Zhi Zhang, Srishti Yadav, Fengze Han, and Ekaterina Shutova. Cross-modal information flow in multimodal large language models. arXiv preprint arXiv:2411.18620, 2024. 5"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            247
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "23934",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/b08d8e36-4b30-4260-aad4-6cb0fd4d67ca_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1db02048554fc50466f7266ceb76a5da8a8a203bc05d5c03d8be2d14f3c16f6c
+size 4892235

2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/full.md

@@ -0,0 +1,333 @@
+# VFlowOpt: A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization
+
+Sihan Yang<sup>1</sup> Runsen Xu<sup>1,2</sup> Chenhang Cui<sup>3</sup> Tai Wang<sup>1†</sup> Dahua Lin<sup>1,2</sup> Jiangmiao Pang<sup>1†</sup>
+
+$^{1}$  Shanghai AI Laboratory  $^{2}$  The Chinese University of Hong Kong  $^{3}$  National University of Singapore
+
+† Corresponding Author taiwang.me@gmail.com pangjiangmiao@gmail.com
+
+# Abstract
+
+Large Multimodal Models (LMMs) excel in visual-language tasks by leveraging numerous visual tokens for fine-grained visual information, but this token redundancy results in significant computational costs. Previous research aimed at reducing visual tokens during inference typically leverages importance maps derived from attention scores among vision-only tokens or vision-language tokens to prune tokens across one or multiple pruning stages. Despite this progress, pruning frameworks and strategies remain simplistic and insufficiently explored, often resulting in substantial performance degradation. In this paper, we propose VFlowOpt, a token pruning framework that introduces an importance map derivation process and a progressive pruning module with a recycling mechanism. The hyperparameters of its pruning strategy are further optimized by a visual information flow-guided method. Specifically, we compute an importance map for image tokens based on their attention-derived context relevance and patch-level information entropy. We then decide which tokens to retain or prune and aggregate the pruned ones as recycled tokens to avoid potential information loss. Finally, we apply a visual information flow-guided method that regards the last token in the LMM as the most representative signal of text-visual interactions. This method minimizes the discrepancy between token representations in LMMs with and without pruning, thereby enabling superior pruning strategies tailored to different LMMs. Experiments demonstrate that VFlowOpt can prune  $90\%$  of visual tokens while maintaining comparable performance, leading to an  $89\%$  reduction in KV-Cache memory and  $3.8 \times$  faster inference.
+
+# 1. Introduction
+
+Large Multimodal Models (LMMs) [3, 8, 21, 22] have attained remarkable performance in tasks such as visual question answering [2, 12] and multimodal reasoning [44, 45], making them indispensable for applications like au
+
+![](images/4bd6a78824a2c043ef30859d3263d0397b752420a0d40c549f030e7784e45242.jpg)  
+Figure 1. VFlowOpt Performance and Efficiency. Our VFlowOpt significantly outperforms other LMM token reduction methods across three general benchmarks (MME, MMStar, and MMBench) on LLaVA-OneVision-7B. It achieves zero performance degradation when reducing  $50\%$  of visual tokens. Furthermore, when retaining only  $10\%$  of the tokens, it achieves  $90\%$  of the original performance while reducing KV-Cache memory usage by  $89\%$  and accelerating inference by  $3.8\times$ .
+
+tonomous driving [9, 31, 35] and robotics [18, 20, 24]. To capture fine-grained visual details, modern LMMs treat images as token sequences, with models such as LLaVA-1.5 [21] processing hundreds of tokens and LLaVA-OneVision [17] handling up to several thousand tokens. Nonetheless, the increasing quantity of visual tokens significantly magnifies computational cost, memory usage, and inference time, creating a critical bottleneck for LMM deployment, particularly in resource-constrained environments or latency-sensitive scenarios.
+
+To alleviate these constraints, recent work explores various methods for reducing the number of visual tokens during inference. Most existing approaches evaluate the importance of visual tokens in LMMs using attention-based importance maps derived either from vision-only tokens [40, 49] or from text-visual interactions [36, 52]. Following a predefined pruning ratio, visual tokens are pruned in a single stage [33] or progressively across multiple stages [42], often with heuristic strategies applied uniformly across different LMMs. Although promising, these methods are sim
+
+plastic and underexplored, with coarse-grained pruning often causing significant performance drops due to lacking fitness for different LMMs models' characteristics.
+
+In response to these limitations, we propose a novel token pruning framework, VFlowOpt, that more accurately estimates token importance by integrating attention calibration and entropy, and reduces information loss through token recycling and progressive pruning strategy with fine-grained pruning ratios. Furthermore, by customizing pruning strategies for different models, our framework better preserves model performance. Specifically, the pruning framework consists of importance map computation and a progressive pruning module with a recycling mechanism. We decompose the importance map of image tokens into two aspects: the importance of tokens in the visual context, reflected by their attention maps, and the richness of visual information in each image patch, captured by its information entropy. Based on the weighted summation of these two factors, we determine which tokens to retain or prune across multiple pruning stages and introduce a recycling mechanism to aggregate pruned tokens to avoid potential information loss. Built with this pruning strategy with several hyperparameters, a key ingredient is the visual information flow-guided method to optimize the pruning strategy, which treats the last token in each pruning stage as the most representative signal of text-visual interactions during inference. By minimizing the difference between token representations in LMMs with and without pruning, VFlowOpt ultimately delivers superior pruning strategies specifically tailored to different LMMs.
+
+Comprehensive experiments on multiple vision-language benchmarks validate the efficacy of VFlowOpt. In particular, it substantially lowers computational cost while preserving competitive performance. As shown in Fig. 1, VFlowOpt can prune  $50\%$  of visual tokens with negligible performance loss. Moreover, it can prune  $90\%$  of visual tokens while maintaining  $90\%$  of the model's performance, resulting in an  $89\%$  reduction in KV-Cache memory usage and a  $3.8 \times$  speedup in inference. These experimental results highlight its effectiveness as a practical and efficient solution for deploying LMMs in resource-constrained environments.
+
+# 2. Related Work
+
+# 2.1. Large Multimodal Models
+
+Large Multimodal Models (LMMs) [3, 8, 21, 22] extend the reasoning capabilities of Large Language Models (LLMs) [4, 32, 38, 39] to vision-language tasks by integrating a pre-trained vision encoder [29, 46] with a language model, linked by an alignment module such as an MLP, or a query-based network. This design transforms visual inputs into token sequences that the LLM can pro
+
+cess, facilitating multimodal prompts for tasks like visual question answering [2, 6, 12, 44]. To enhance performance, advanced LMMs, such as LLaVA-OneVision [17] and Qwen2-VL [34], can encode higher-resolution images into more image tokens, thereby capturing more granular visual details. However, as image resolution grows, the number of visual tokens rises exponentially, leading to substantially higher computational costs. For example, LLaVA-1.5 [21] processes  $336 \times 336$  images into 576 tokens, whereas LLaVA-OneVision can handle  $1152 \times 1152$  images, producing 7,290 tokens. This challenge becomes even more pronounced in video-based models like LongVA [48], which must process tokens across numerous frames. Fine-grained visual tokenization boosts LMM performance but poses an inference bottleneck, driving efforts to balance performance and cost through token reduction.
+
+# 2.2. Token Reduction for LMMs
+
+Token reduction has become a key strategy for improving the efficiency of LMMs by mitigating the computational cost associated with extensive visual token sequences. Existing methods can be broadly classified into training-based and training-free approaches. Training-based methods, such as LLaVA-Mini [50] and LLaVolta [5], introduce additional modules during model training to compress visual tokens and preserve critical information, while approaches like ATP-LLaVA [43] and p-MoD [47] train pruning modules to dynamically retain important tokens across LLM layers. However, these methods demand substantial computational resources to retrain the models, limiting their real-world applicability. In contrast, training-free methods prune tokens without additional training, often leveraging attention mechanisms to identify and discard redundant tokens. For instance, FastV[7] and SparseVLM exploit text-visual attention to rank token importance, whereas FasterVLM [49], VisionZip [40], and VTC-CLS [33] rely on [CLS] token attention in the vision encoder to evaluate token importance and prune redundant tokens. FitPrune [42] and PDrop [36] propose a progressive multistage pruning strategy in LMMs to fully utilize visual information. However, the pruning strategies in these methods often lack adaptability to different LMMs, which frequently leads to significant performance drops.
+
+# 3. Method
+
+In this section, we introduce VFlowOpt, a framework that employs a pruning strategy to reduce redundant visual tokens during LMM inference while preserving essential visual information. This strategy involves multiple hyperparameters that our VFlowOpt framework automatically optimizes for proper configuration—an essential factor for maintaining model performance. Specifically, Section 3.1 explains how our method evaluates the importance of visual
+
+![](images/11f3eec939c6bc6f8cac28c20264c1deb05bddf83b8a2d827d4fc8af5103a5c8.jpg)  
+(a) Token Importance Estimation
+
+![](images/bd996f6af1ded17ff73c3562764b3550a45b47a57cd004df5ca48fe7e66614ac.jpg)  
+(b) Token Recycling
+
+![](images/ac77e9c733e7b164f80cd00a8207b0782e0c95b091775067933bce49f861a589.jpg)  
+(1) Inference Stage
+
+![](images/3c405a4b6ddb48846a13a61c0ebe6a3c45543abb3506bdd88281f3b31d79216c.jpg)  
+(2) Optimization Stage
+
+![](images/08f38546afb6798412810532e22ad42dbf3845ad1e93985beeff65b10ec97b23.jpg)
+
+![](images/9c4ac0f406f77133d4fb72a2bffd19daf953d57a3f3ef903f97d9543a1ade086.jpg)  
+Figure 2. Overview of VFlowOpt. (1) During inference, VFlowOpt first assesses the importance of visual tokens, based on which progressive token pruning is performed. After the initial pruning stage, the pruned tokens are merged and recycled. The pruning strategy used in this process is defined by the (2) Optimization Stage. (a) The importance map is computed by combining the attention of relatively important tokens with the entropy of image patches. (b) The pruned tokens are grouped into grid cells, where each cell has a side length of a. Within each grid cell, the pruned tokens are fused using a weighted average, with their importance values as weights, and then recycled. (2) VFlowOpt optimizes the pruning strategy by minimizing the discrepancy of the last token in the final layer of the LMM with and without applying the pruning strategy.  
+Figure 3. The attention of redundant tokens (marked in red) fails to reflect the importance of other tokens and instead focuses on similar tokens, such as background elements.
+
+tokens, Section 3.2 presents progressive pruning and token recycling strategies, and finally, Section 3.3 details how VFlowOpt customizes the superior pruning strategy (i.e., selects the pruning hyperparameters) for different LMMs.
+
+# 3.1. Visual Token Importance Estimation
+
+For a general approach to evaluate visual token importance, the previous works [14, 41] propose that the importance of
+
+visual tokens can be estimated using the average attention from all other tokens in ViTs. Although it has achieved some success, we find that redundant tokens (e.g., tokens corresponding to background regions) often assign disproportionately high attention to other similar redundant tokens (shown in Fig. 3), reducing the reliability of the importance estimation. To address this, we first identify relatively important tokens based on the attention they receive from all tokens (tokens receiving higher attention are considered relatively important), as shown in Fig. 2 (a). We then exclude redundant tokens and use the attention from these relatively important tokens as a more robust metric. Additionally, we incorporate the information entropy of image patches into the importance score to prioritize tokens corresponding to visually informative regions.
+
+To identify relatively important visual tokens, we define a threshold:
+
+$$
+\tau = t \cdot \frac {1}{N} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {N} A _ {i j}, \tag {1}
+$$
+
+where  $\tau$  is the threshold,  $A_{ij}$  is the attention weight from
+
+token  $i$  to token  $j$  within a ViT layer,  $N$  is the total number of tokens, and  $t$  is a sensitivity hyperparameter. We then treat tokens with a total attention exceeding  $\tau$  as relatively important:
+
+$$
+\mathcal {K} = \{j \mid \sum_ {i = 1} ^ {N} A _ {i j} > \tau \}. \tag {2}
+$$
+
+Here,  $\mathcal{K}$  denotes the set of indices of these relatively important tokens.
+
+The information entropy of an image patch corresponding to token  $i$  is defined as:
+
+$$
+H \left(V _ {i}\right) = - \sum_ {k = 0} ^ {L - 1} p _ {k} \log p _ {k}, \tag {3}
+$$
+
+where  $V_{i}$  is the image patch corresponding to token  $i$ ,  $p_{k}$  is the proportion of pixels in the patch with gray level  $k$ , and  $L$  is the total number of possible gray levels (256 for 8-bit images). Here, the gray level is defined as the average of the RGB channel values, providing a scalar representation of pixel brightness. Higher entropy reflects greater diversity in pixel intensities and richer visual information.
+
+The importance score of token  $i$  is computed by combining the attention it receives from tokens in  $\mathcal{K}$  with the entropy of its corresponding image patch, normalized via the softmax function:
+
+$$
+I _ {i} = \sum_ {k \in \mathcal {K}} A _ {k i} + \alpha \cdot \frac {\exp (H (V _ {i}))}{\sum_ {j = 1} ^ {N} \exp (H (V _ {j}))}, \tag {4}
+$$
+
+where  $I_{i}$  denotes the importance score of token  $i$ ,  $A_{ki}$  is the attention weight from token  $k$  to token  $i$ ,  $H(V_{i})$  is the entropy of the image patch corresponding to token  $i$ ,  $\alpha$  is a hyperparameter that determines the entropy term's contribution, and  $N$  is the total number of tokens. This approach mitigates biases from redundant tokens and leverages the visual information of image patches, thereby facilitating effective estimation of visual token importance.
+
+# 3.2. Progressive Pruning and Token Recycling
+
+Previous studies [37, 51] indicate that visual tokens play a more critical role in the shallower layers of LMMs, whereas redundancy tends to increase in deeper layers. Based on this, we adopt a progressive pruning strategy. To balance the pruning process's simplicity with its fine-grained configuration, we evenly divide the LMM into three stages. At the beginning of each stage, a predetermined fraction of visual tokens with higher importance scores (as defined in Sec. 3.1) is retained according to the stage-specific retention ratios  $R = [R_{1}, R_{2}, R_{3}]$ , while the rest are pruned. Notably, the position IDs of visual tokens remain unchanged following pruning, preserving the original spatial structure of the visual input. Following the previous work [37], we compute the average visual token retention rate across the entire
+
+LMM as follows:
+
+$$
+\bar {R} = \frac {R _ {1} \cdot L _ {1} + R _ {1} \cdot R _ {2} \cdot L _ {2} + R _ {1} \cdot R _ {2} \cdot R _ {3} \cdot L _ {3}}{L}, \tag {5}
+$$
+
+where  $L_{1}, L_{2}$ , and  $L_{3}$  represent the number of layers in the three stages, respectively, and  $L$  denotes the total number of layers in the LMM.
+
+To prevent the loss of any small but potentially significant information during the initial pruning process (before visual tokens are fed into the LLM), we propose a token merging and recycling strategy to compactly represent redundant visual information. Specifically, We use a square grid (with each cell having side length  $a$ ) to group pruned tokens that fall into the same grid cell, as shown in Fig. 2 (b). Within each cell, the pruned tokens are fused into a single token by computing a weighted average of their representations, using their importance scores as weights. This fused token then replaces the pruned token with the highest importance score in that cell and is incorporated into the set of retained tokens.
+
+Formally, a token  $t_i$  with spatial coordinates  $(x_i, y_i)$  belongs to a grid cell  $\mathcal{G}_{p,q}$  if
+
+$$
+\lfloor x _ {i} / a \rfloor = p \quad \text {a n d} \quad \lfloor y _ {i} / a \rfloor = q, \tag {6}
+$$
+
+where  $p$  and  $q$  are the row and column indices of the grid cell, respectively. Suppose there are  $k$  pruned tokens in  $\mathcal{G}_{p,q}$  with corresponding importance scores  $I_{1}, I_{2}, \ldots, I_{k}$  and token representations  $\mathbf{t}_1, \mathbf{t}_2, \ldots, \mathbf{t}_k$ . The fused token  $\mathbf{t}_{\mathrm{merged}}^{p,q}$  is computed as:
+
+$$
+\mathbf {t} _ {\text {m e r g e d}} ^ {p, q} = \frac {\sum_ {i = 1} ^ {k} I _ {i} \cdot \mathbf {t} _ {i}}{\sum_ {i = 1} ^ {k} I _ {i}}. \tag {7}
+$$
+
+We then assign  $\mathbf{t}_{\mathrm{merged}}^{p,q}$  to the position of the pruned token with the highest importance score  $I_{\mathrm{max}}$  in  $\mathcal{G}_{p,q}$ , counting it among the retained tokens. By combining token pruning with this recycling strategy, we reduce token redundancy while avoiding potential information loss.
+
+# 3.3. Pruning Strategy Optimization
+
+The pruning strategies described in Sec. 3.1 and Sec. 3.2 involve several key hyperparameters that directly affect the performance of LMMs after token pruning. Properly defining these hyperparameters is crucial for preserving model performance. However, existing approaches often rely on manually designed pruning strategies and apply the same strategy across different LMMs, without considering the unique characteristics of each model. This coarse-grained approach can lead to significant performance degradation due to its lack of fitness for different LMMs, leaving the task of designing pruning strategies tailored to specific LMMs as a formidable challenge.
+
+Previous interpretability studies of LMMs [15, 53] offer crucial insights into their internal mechanisms. Specifically, these works reveal that in the lower and middle layers of LMMs, visual information from visual tokens is aggregated into the corresponding query text tokens. In the higher layers, the multimodal representation encoded in the query text tokens is further progressively propagated to the final position of the input sequence, ultimately influencing the subsequent inference process.
+
+Inspired by this insight, we propose a framework that requires only a small amount of unlabeled data and leverages the internal flow of visual information within LMMs to search for the superior pruning strategy tailored to the characteristics of different LMMs, as shown in Fig. 2 (2). Specifically, we recast the task of designing the superior pruning strategy as an optimization problem, aiming to minimize the discrepancy in visual information flow with and without applying a pruning strategy to perform visual token pruning. In this framework, we treat the final token in the last layer as the representative outcome of the visual information flow. We use cosine similarity to measure the similarity of the final token representation with and without visual token pruning. A higher similarity indicates that the discrepancy between the visual information flow with and without pruning is smaller. Accordingly, we define the following optimization objective:
+
+$$
+\max  _ {s \in \mathcal {S}} f (s) = \operatorname {C o s i n e S i m} \left(h _ {f}, g _ {s} \left(h _ {f}\right)\right). \tag {8}
+$$
+
+Here,  $h_f$  represents the representation of the final token in the last layer before pruning, and  $g_s(h_f)$  represents the final token representation after applying the pruning strategy  $s$  (with  $g_s$  modeling its effect on the token feature), and  $S$  is the solution space.
+
+To efficiently search for the superior pruning strategy  $s$ , we employ Bayesian optimization, which systematically explores the hyperparameter space—including the threshold sensitivity  $t$ , the entropy weight  $\alpha$ , the grid size  $a$ , and the pruning ratios  $R_{1}, R_{2}$ , and  $R_{3}$ —to maximize the target function  $f(s)$ . In short, Bayesian optimization constructs a surrogate model to approximate the objective function and employs an acquisition function to balance exploration and exploitation, thereby efficiently guiding the search for promising hyperparameter settings. The details of this optimization process are presented in Algorithm 1.
+
+# 4. Experiments
+
+In this section, we evaluate our approach on various LMMs across diverse image and video benchmarks, followed by an efficiency analysis and an ablation study of each component. Finally, we illustrate how our method affects discrepancies in the visual information flow with and without pruning, offering deeper insights into our approach.
+
+Algorithm 1 VFlowOpt with Bayesian Optimization
+
+Input: LMM  $\theta$ , data samples  $\mathcal{D}$ , number of Bayesian optimization iterations  $T$ , computation budget  $\overline{R}$ , target function  $f(\cdot)$ : the sum of cosine similarities between the last token in the last layer of the LMM  $\theta$  with and without visual token pruning, computed over all data samples  $\mathcal{D}$ .  
+Output: retention rates  $R_{1}, R_{2}, R_{3}$ ; threshold sensitivity  $t$ ; entropy weight in importance score  $\alpha$ ; grid size  $a$ .
+
+1: Initialize a Gaussian Process model  $\mathcal{GP}$  
+2: Define the acquisition function  $A(\cdot)$  (Expected Improvement is adopted)  
+3: Uniformly sample initial points:
+
+$$
+X _ {0} = \left\{\left(R _ {1}, R _ {2}, t, \alpha , a\right) \mid R _ {v}, R _ {1}, t, \alpha , a \in \text {v a l i d r a n g e s} \right\}
+$$
+
+4: for all  $x \in X_0$  do
+
+5: Calculate  $R_{3}$  using the constraint:  $\overline{R} = (R_{1} \cdot L_{1} + R_{1} \cdot R_{2} \cdot L_{2} + R_{1} \cdot R_{2} \cdot R_{3} \cdot L_{3}) / L$  
+6: Form pruning strategy:  $s = (R_1, R_2, R_3, t, \alpha, a)$  
+7: Evaluate the target function  $f(s \mid \theta, \mathcal{D})$
+
+8: end for
+
+9: for  $n = 0$  to  $T - 1$  do
+
+10: Fit  $\mathcal{GP}$  to the observed data  $(S_{n},f(S_{n}\mid \theta ,\mathcal{D}))$  
+11: Select the next point:  $x_{n + 1}\gets \arg \max_xA(x;\mathcal{GP})$  
+12: Calculate  $R_{3}$  using the constraint:  $\overline{R} = (R_{1} \cdot L_{1} + R_{1} \cdot R_{2} \cdot L_{2} + R_{1} \cdot R_{2} \cdot R_{3} \cdot L_{3}) / L$  
+13: Form pruning strategy:  $s_{n+1} = (R_1, R_2, R_3, t, \alpha, a)_{n+1}$  
+14: Update  $S_{n + 1}\gets S_n\cup \{s_{n + 1}\}$  
+15: end for  
+16: return  $(R_{1}, R_{2}, R_{3}, t, \alpha, a)^{*}$  that maximize  $f$
+
+# 4.1. Experiment Setting
+
+Datasets. We evaluate our method on ten image-based multimodal benchmarks: GQA [1], VizWiz [13], ScienceQA-IMG [26], TextVQA [30], ChartQA [27], POPE [19], MME [10], MMBench [25], MMStar [6], and DocVQA [28]. For video understanding, we adopt two datasets—SeedBench (video) [16] and VideoMME [11], where VideoMME is partitioned by video length into short, medium, and long subsets. Further details are presented in the Appendix.
+
+Model Architectures. We integrate VFlowOpt into multiple LMMs, including LLaVA-OneVision-7B, LLaVA-NeXT-7B [23], and Qwen2-VL-7B [34]. These models employ various vision encoders (SigLIP, CLIP, and a ViT designed for Qwen2-VL) and LLM backbones (Qwen2-7B and Vicuna-7B). To prevent out-of-memory issues in Qwen2-VL, we set maxPixels = 3000000.
+
+Comparison Methods. We compare VFlowOpt against five baseline methods: FastV [7], FitPrune [42], PDrop [36], SparseVLM [52], and VisionZip [40]. FastV,
+
+<table><tr><td>Method</td><td>MMStar</td><td>MME</td><td>MMB</td><td>SQA</td><td>POPE</td><td>GQA</td><td>VizWiz</td><td>\(VQA^{Text}\)</td><td>ChartQA</td><td>DocVQA</td><td>Avg.</td></tr><tr><td colspan="12">Upper Bound, 100% Tokens</td></tr><tr><td>Vanilla</td><td>61.7</td><td>1581</td><td>80.8</td><td>95.8</td><td>89.1</td><td>62.2</td><td>60.4</td><td>76.0</td><td>80.0</td><td>87.5</td><td>100%</td></tr><tr><td colspan="12">Retain 50% Tokens (↓ 50%)</td></tr><tr><td>FastV (ECCV24)</td><td>58.9</td><td>1549</td><td>79.4</td><td>92.8</td><td>87.9</td><td>61.5</td><td>61.1</td><td>72.5</td><td>68.6</td><td>84.0</td><td>96.5%</td></tr><tr><td>FitPrune (AAAI25)</td><td>4.5</td><td>96.6</td><td>1.0</td><td>8.6</td><td>18.8</td><td>16.2</td><td>6.8</td><td>37.6</td><td>37.7</td><td>30.7</td><td>21.4%</td></tr><tr><td>PDrop (CVPR25)</td><td>37.1</td><td>628.6</td><td>32.1</td><td>64.8</td><td>81.3</td><td>44.2</td><td>48.9</td><td>63.7</td><td>62.7</td><td>62.2</td><td>71.3%</td></tr><tr><td>SparseVLM (2024.10)</td><td>59.8</td><td>1577</td><td>80.5</td><td>94.1</td><td>88.1</td><td>61.9</td><td>60.4</td><td>73.9</td><td>70.5</td><td>80.8</td><td>97.1%</td></tr><tr><td>VisionZip (CVPR25)</td><td>60.4</td><td>1587</td><td>80.3</td><td>94.6</td><td>89.3</td><td>62.7</td><td>59.8</td><td>74.2</td><td>75.4</td><td>88.4</td><td>98.9%</td></tr><tr><td>VFlowOpt</td><td>61.3</td><td>1591</td><td>81.1</td><td>95.4</td><td>89.4</td><td>62.4</td><td>60.0</td><td>75.1</td><td>77.8</td><td>90.0</td><td>99.9%</td></tr><tr><td colspan="12">Retain 25% Tokens (↓ 75%)</td></tr><tr><td>FastV (ECCV24)</td><td>54.0</td><td>1539</td><td>77.0</td><td>88.6</td><td>83.8</td><td>58.2</td><td>61.0</td><td>58.3</td><td>42.7</td><td>62.9</td><td>86.4%</td></tr><tr><td>FitPrune (AAAI25)</td><td>5.2</td><td>92.6</td><td>1.1</td><td>5.9</td><td>20.2</td><td>14.5</td><td>5.1</td><td>29.7</td><td>27.3</td><td>23.2</td><td>17.6%</td></tr><tr><td>PDrop (CVPR25)</td><td>36.8</td><td>574.4</td><td>31.3</td><td>62.8</td><td>78.0</td><td>42.2</td><td>47.1</td><td>56.3</td><td>47.5</td><td>47.0</td><td>65.2%</td></tr><tr><td>SparseVLM (2024.10)</td><td>56.6</td><td>1520</td><td>78.7</td><td>90.3</td><td>87.2</td><td>59.7</td><td>60.8</td><td>66.3</td><td>54.0</td><td>66.6</td><td>90.5%</td></tr><tr><td>VisionZip (CVPR25)</td><td>54.6</td><td>1562</td><td>78.9</td><td>90.4</td><td>88.8</td><td>61.0</td><td>60.4</td><td>70.0</td><td>66.3</td><td>79.6</td><td>94.3%</td></tr><tr><td>VFlowOpt</td><td>57.8</td><td>1570</td><td>79.9</td><td>92.3</td><td>89.1</td><td>61.2</td><td>60.4</td><td>72.5</td><td>69.1</td><td>82.3</td><td>96.3%</td></tr><tr><td colspan="12">Retain 10% Tokens (↓ 90%)</td></tr><tr><td>FastV (ECCV24)</td><td>46.0</td><td>1209</td><td>70.1</td><td>81.7</td><td>77.0</td><td>51.5</td><td>56.4</td><td>35.6</td><td>21.3</td><td>33.2</td><td>69.7%</td></tr><tr><td>FitPrune (AAAI25)</td><td>5.4</td><td>27.0</td><td>0.9</td><td>3.5</td><td>2.3</td><td>4.6</td><td>0.7</td><td>6.9</td><td>5.9</td><td>6.9</td><td>5.1%</td></tr><tr><td>PDrop (CVPR25)</td><td>31.8</td><td>456.3</td><td>22.5</td><td>59.0</td><td>65.9</td><td>35.7</td><td>43.3</td><td>33.7</td><td>19.6</td><td>18.5</td><td>49.9%</td></tr><tr><td>SparseVLM (2024.10)</td><td>45.1</td><td>1191</td><td>71.8</td><td>83.7</td><td>80.0</td><td>54.6</td><td>56.4</td><td>39.8</td><td>37.6</td><td>39.6</td><td>74.0%</td></tr><tr><td>VisionZip (CVPR25)</td><td>49.5</td><td>1389</td><td>74.8</td><td>86.2</td><td>86.1</td><td>57.2</td><td>56.8</td><td>56.4</td><td>46.1</td><td>49.0</td><td>82.1%</td></tr><tr><td>VFlowOpt</td><td>52.0</td><td>1464</td><td>75.1</td><td>88.4</td><td>85.2</td><td>57.3</td><td>57.3</td><td>60.2</td><td>53.6</td><td>56.1</td><td>85.5%</td></tr></table>
+
+Table 1. Performance comparison on LLaVA-OneVision-7B under different token retention conditions. "Avg." refers to average accuracy on 10 benchmarks. For each reduction ratio, the best average performance is shown in bold.
+
+FitPrune, PDrop, and SparseVLM rely on text-visual attention in the LLM to prune visual tokens but differ as follows: FastV performs a single pruning step after the second LLM layer; FitPrune prunes tokens in each layer according to predefined, layer-specific ratios; PDrop conducts four pruning steps within the LLM, using the same retention ratio each time; and SparseVLM uses the attention weights of preselected text tokens to evaluate the importance of visual tokens. VisionZip, in contrast, determines token importance based on the [cls] token's attention; for models lacking a [cls] token, we follow VisionZip's original procedure by computing the average attention each token receives from every other token in the ViTs.
+
+Implementation Details. For LLaVA-OneVision-7B and Qwen2-VL-7B, we perform token pruning at three distinct points: before the LLM, and after the 9th and 18th layers. For LLaVA-NeXT-7B, pruning is conducted before the LLM, and again after the 10th and 20th layers. During optimization, we sample 30 unlabeled instances from each model's training datasets; for models without publicly available training dataset, we instead use random samples from the LLaVA-OneVision training set. The optimization is performed for a total of 50 iterations. For LMMs that modify visual tokens output by the vision encoder (e.g., unpadding and interpolation in LLaVA-OneVision, and unpadding in LLaVA-NeXT), we apply corresponding transformations to the importance maps so they remain fully aligned with the final visual tokens.
+
+# 4.2. Image Understanding Tasks
+
+We evaluate the proposed VFlowOpt on image understanding benchmarks with LLaVA-OneVision-7B using various pruning ratios, and present the results in Tab. 1. Compared to other baselines, VFlowOpt consistently maintains superior accuracy across different levels of token pruning. With  $50\%$  token retention, VFlowOpt achieves  $99.9\%$  of the original performance, exceeding the second-best approach by  $1.0\%$ . This negligible performance drop underscores the method's strong potential for practical deployment. Under more extreme pruning conditions with very few retained tokens, VFlowOpt's advantage becomes more pronounced. When only  $10\%$  of the visual tokens are retained, VFlowOpt preserves  $85.5\%$  of the original performance, surpassing the second-best approach by  $3.3\%$ . This finding suggests that VFlowOpt can effectively leverage limited visual information to maintain high performance under strict computational budgets. When applied to LLaVA-OneVision, both FitPrune and PDrop disrupt the model's normal output, indicating their inability to devise effective pruning strategies for different architectures. In contrast, VFlowOpt demonstrates a clear advantage by tailoring its pruning strategy to each model's unique characteristics.
+
+To further validate VFlowOpt's generalization capability, we evaluate it on LLaVA-NeXT-7B and Qwen2-VL-7B. As shown in Tab. 2 and Tab. 3, VFlowOpt once again delivers the top results on these LMMs. When retaining  $25\%$
+
+<table><tr><td>Method</td><td>MMStar</td><td>MME</td><td>MMB</td><td>SQA</td><td>POPE</td><td>GQA</td><td>Avg.</td></tr><tr><td colspan="8">Upper Bound, 100% Tokens</td></tr><tr><td>Vanilla</td><td>37.6</td><td>1519</td><td>67.4</td><td>70.1</td><td>86.5</td><td>64.2</td><td>100%</td></tr><tr><td colspan="8">Retain 25% Tokens (↓ 75%)</td></tr><tr><td>FastV</td><td>35.1</td><td>1477</td><td>65.6</td><td>67.4</td><td>83.1</td><td>60.4</td><td>95.7%</td></tr><tr><td>VisionZip</td><td>35.8</td><td>1501</td><td>65.4</td><td>67.9</td><td>86.7</td><td>61.5</td><td>97.3%</td></tr><tr><td>VFlowOpt</td><td>37.0</td><td>1514</td><td>67.0</td><td>67.7</td><td>87.6</td><td>62.6</td><td>98.8%</td></tr><tr><td colspan="8">Retain 10% Tokens (↓ 90%)</td></tr><tr><td>FastV</td><td>29.2</td><td>1282</td><td>61.6</td><td>63.8</td><td>71.7</td><td>55.9</td><td>85.7%</td></tr><tr><td>VisionZip</td><td>32.6</td><td>1378</td><td>61.5</td><td>67.1</td><td>83.5</td><td>57.0</td><td>91.6%</td></tr><tr><td>VFlowOpt</td><td>35.1</td><td>1393</td><td>62.9</td><td>67.4</td><td>83.6</td><td>57.3</td><td>93.4%</td></tr></table>
+
+Table 2. Comparative experiments on LLaVA-NeXT-7B.  
+
+<table><tr><td>Method</td><td>MMStar</td><td>MME</td><td>MMB</td><td>SQA</td><td>POPE</td><td>GQA</td><td>Avg.</td></tr><tr><td colspan="8">Upper Bound, 100% Tokens</td></tr><tr><td>Vanilla</td><td>57.5</td><td>1680</td><td>80.3</td><td>84.7</td><td>88.4</td><td>62.2</td><td>100%</td></tr><tr><td colspan="8">Retain 25% Tokens (↓ 75%)</td></tr><tr><td>FastV</td><td>54.5</td><td>1597</td><td>76.3</td><td>79.0</td><td>81.8</td><td>57.2</td><td>93.8</td></tr><tr><td>VisionZip</td><td>55.2</td><td>1618</td><td>78.9</td><td>81.3</td><td>86.9</td><td>60.0</td><td>96.8</td></tr><tr><td>VFlowOpt</td><td>55.9</td><td>1659</td><td>79.8</td><td>81.5</td><td>87.1</td><td>60.4</td><td>97.8</td></tr><tr><td colspan="8">Retain 10% Tokens (↓ 90%)</td></tr><tr><td>FastV</td><td>44.9</td><td>1405</td><td>70.0</td><td>75.5</td><td>75.8</td><td>51.6</td><td>84.4</td></tr><tr><td>VisionZip</td><td>49.3</td><td>1518</td><td>76.5</td><td>78.2</td><td>84.3</td><td>54.1</td><td>90.9</td></tr><tr><td>VFlowOpt</td><td>51.0</td><td>1591</td><td>78.0</td><td>78.6</td><td>84.5</td><td>54.8</td><td>92.8</td></tr></table>
+
+Table 3. Comparative experiments on Qwen2-VL-7B.  
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">SeedBench (video)</td><td colspan="3">VideoMME</td><td rowspan="2">Avg.</td></tr><tr><td>S</td><td>M</td><td>L</td></tr><tr><td colspan="6">Upper Bound, 100% Tokens</td></tr><tr><td>Vanilla</td><td>56.9</td><td>70.5</td><td>54.6</td><td>49.5</td><td>100%</td></tr><tr><td colspan="6">Retain 25% Tokens (↓75%)</td></tr><tr><td>FastV</td><td>54.7</td><td>66.9</td><td>53.2</td><td>47.7</td><td>96.2%</td></tr><tr><td>VisionZip</td><td>56.4</td><td>68.3</td><td>55.3</td><td>49.0</td><td>99.0%</td></tr><tr><td>VFlowOpt</td><td>56.8</td><td>68.9</td><td>55.8</td><td>49.2</td><td>100%</td></tr><tr><td colspan="6">Retain 10% Tokens (↓90%)</td></tr><tr><td>FastV</td><td>48.7</td><td>53.7</td><td>47.6</td><td>42.3</td><td>83.6%</td></tr><tr><td>VisionZip</td><td>55.0</td><td>59.7</td><td>51.8</td><td>46.3</td><td>92.4%</td></tr><tr><td>VFlowOpt</td><td>55.5</td><td>63.3</td><td>52.7</td><td>48.6</td><td>95.5%</td></tr></table>
+
+Table 4. Comparative experiments on video understanding tasks.
+
+of the tokens, VFlowOpt preserves  $98.8\%$  of the original performance on LLaVA-NeXT-7B and  $97.8\%$  on Qwen2-VL-7B. Even under the stringent condition of retaining only  $10\%$  of the tokens, VFlowOpt maintains  $93.4\%$  of the performance on LLaVA-NeXT-7B and  $92.8\%$  on Qwen2-VL-7B, demonstrating its strong generalizability.
+
+# 4.3. Generalization to Video Tasks
+
+We further explore VFlowOpt's generalization across different modalities by evaluating it on video benchmarks using LLaVA-OneVision-7B. As shown in Tab. 4, VFlowOpt surpasses other baselines under various token retention ratios. Notably, when only  $25\%$  of the tokens are preserved, VFlowOpt incurs virtually no performance loss on LLaVA-OneVision-7B, maintaining  $100\%$  of its original performance. Even under the more stringent condition of retain
+
+ing just  $10\%$  of the tokens, VFlowOpt still retains  $95.5\%$  of the original performance, underscoring its robust generalization in the video domain.
+
+# 4.4. Efficiency Analysis
+
+We demonstrate the efficiency of VFlowOpt by conducting a comparative study on LLaVA-OneVision-7B running on a single NVIDIA A100-SXM4-80GB GPU, focusing on FLOPs, KV-Cache memory usage, and inference latency. The efficiency analysis of other baseline methods is provided in the Appendix.
+
+We assess various token pruning ratios on the MME benchmark, measuring overall performance and average efficiency metrics across all samples. As shown in Tab. 5, VFlowOpt significantly enhances LLaVA-OneVision-7B's computational efficiency by reducing FLOPs, shrinking KV-Cache memory usage, and accelerating inference speed. Notably, at a  $50\%$  token pruning ratio, VFlowOpt achieves a  $49.5\%$  reduction in KV-Cache memory and a  $1.8\times$  speedup in inference, while boosting performance by  $0.7\%$ . At a more aggressive  $75\%$  pruning ratio, we observe a  $74.2\%$  reduction in KV-Cache memory and a  $3.1\times$  speedup, with only a  $0.7\%$  performance drop. Under extremely tight computational budgets, pruning  $90\%$  of visual tokens reduces the KV-Cache memory footprint by  $89\%$ , accelerates inference by  $3.8\times$ , and degrades performance by only  $7.4\%$ .
+
+Such reductions in inference latency substantially benefit user-facing applications demanding real-time performance, such as autonomous driving and robotics. Moreover, by significantly shrinking the KV-Cache, large-batch inference on LMMs can accommodate more user requests simultaneously, thereby substantially reducing overall inference costs. In summary, VFlowOpt uses far less GPU memory and delivers faster inference while preserving model performance, offering a highly practical solution for efficiently deploying LMMs in real-world scenarios.
+
+# 4.5. Ablation Study
+
+To confirm the contributions of each component in our token pruning strategy, we conduct ablation experiments on LLaVA-OneVision-7B under a computational budget of retaining  $25\%$  of tokens. As shown in Tab. 6, removing importance calibration (i.e., directly using the mean attention over all tokens received by each visual token as its importance score), omitting token recycling, or discarding progressive pruning (i.e., maintaining the same number of tokens at each layer) leads to noticeable performance degradation across MMStar, MMBench, and SQA. By contrast, the complete VFlowOpt method consistently exhibits minimal performance loss, highlighting the effectiveness of all its components in pruning redundant tokens while preserving essential visual information.
+
+We further examine how the number of samples and the
+
+<table><tr><td>Methods</td><td>Token Reduction</td><td>FLOPs ↓ (T)</td><td>Δ</td><td>Latency ↓ (ms)</td><td>Δ</td><td>KV Cache ↓ (MB)</td><td>Δ</td><td>Performance ↑</td><td>Δ</td></tr><tr><td>LLaVA-OneVision-7B</td><td>-</td><td>71.4</td><td>-</td><td>1040.1</td><td>-</td><td>1786.4</td><td>-</td><td>1581</td><td>-</td></tr><tr><td rowspan="3">+ VFlowOpt</td><td>50%</td><td>37.2</td><td>-48.0%</td><td>584.2</td><td>-43.8%</td><td>902.8</td><td>-49.5%</td><td>1591</td><td>+0.6%</td></tr><tr><td>75%</td><td>19.1</td><td>-73.2%</td><td>328.5</td><td>-68.4%</td><td>460.6</td><td>-74.2%</td><td>1570</td><td>-0.7%</td></tr><tr><td>90%</td><td>7.7</td><td>-89.2%</td><td>272.1</td><td>-73.8%</td><td>197.1</td><td>-89.0%</td><td>1464</td><td>-7.4%</td></tr></table>
+
+Table 5. Efficiency analysis of LLaVA-OneVision-7B with VFlowOpt. The detailed metric includes computation (FLOPs), latency, and KV-Cache memory.  $(\Delta)$  denotes the reduction ratio.  
+
+<table><tr><td>Pruning Strategy</td><td>MMStar</td><td>MMBench</td><td>SQA</td></tr><tr><td>VFlowOpt</td><td>57.8</td><td>79.9</td><td>92.3</td></tr><tr><td>w/o Importance Calibration</td><td>56.2</td><td>79.4</td><td>91.8</td></tr><tr><td>w/o Token Merging</td><td>57.6</td><td>79.8</td><td>91.9</td></tr><tr><td>w/o Progressive Pruning</td><td>56.0</td><td>79.1</td><td>91.0</td></tr></table>
+
+Table 6. Ablation studies of pruning strategy components.
+
+![](images/2747e4b8f4cda23aeadfcdea5c5eba1512cdc543528df07152841897b6e5eb9c.jpg)  
+Figure 4. Relationships of sample size (left) and optimization steps (right) with optimization time and final performance.
+
+number of optimization iterations in the Bayesian optimization procedure influence both optimization time and final performance. As illustrated in Fig. 4, the optimization time increases linearly with the number of samples and steps. Using 30 samples and 50 steps strikes an effective balance between time efficiency and performance, requiring only about 30 minutes to reach an exceptional result. More ablation results are provided in the supplementary material.
+
+# 4.6. Visual Information Flow Analysis
+
+We further investigate how Bayesian optimization formulates the superior pruning strategy by examining its impact on visual information flow. Specifically, we compare the randomly initialized pruning strategy derived from the Bayesian optimization process with its final optimized strategy, measuring their respective effects on the text tokens that follow the visual tokens at each LLM layer. Because these text tokens can receive information from the visual tokens via the LLM's unidirectional attention mechanism, examining differences in these tokens with and without pruning effectively reveals how visual information flow is altered. As shown in Fig. 5, under the same computational budget, the initial strategy causes a pronounced discrepancy between text token representations with and without pruning, indicating that it significantly disrupts the visual information flow. In contrast, when using the optimized strategy,
+
+![](images/d2a427acb1d1889fcb27ddbe93caecc35d46160cebc60341d0b621e7434e3262.jpg)  
+Figure 5. Visualization of the discrepancies in visual information flow corresponding to the pruning strategies before and after optimization, along with an example of visual question answering.
+
+the text tokens remain much more similar to those observed without visual token pruning. This important finding suggests that the optimization process yields a pruning strategy that preserves crucial visual information by minimizing the difference in visual information flow with and without pruning, thereby retaining LMM performance to the greatest extent possible.
+
+# 5. Conclusion
+
+In this paper, we address the pressing challenge of computational inefficiency in LMMs by introducing a general framework for visual token pruning. We reformulate pruning as an optimization problem focused on minimizing the divergence of visual information flow with and without pruning, thereby facilitating a tailored pruning strategy for diverse LMMs. By leveraging calibrated attention in ViTs to evaluate token significance and merging pruned tokens using importance-based weighting, we preserve critical information during inference without requiring retraining or manual tuning. Extensive experiments underscore the efficiency and generalizability of our approach, making it a practical solution for deploying LMMs in real-world settings, particularly in resource-constrained environments.
+
+# Acknowledgements
+
+This research was supported by Shanghai Artificial Intelligence Laboratory.
+
+# References
+
+[1] Joshua Ainslie, James Lee-Thorp, Michiel De Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245, 2023. 5  
+[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 1, 2  
+[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1, 2  
+[4] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, Xiaoyi Dong, Haodong Duan, Qi Fan, Zhaoye Fei, Yang Gao, Jiaye Ge, Chenya Gu, Yuzhe Gu, Tao Gui, Aijia Guo, Qipeng Guo, Conghui He, Yingfan Hu, Ting Huang, Tao Jiang, Penglong Jiao, Zhenjiang Jin, Zhikai Lei, Jiaxing Li, Jingwen Li, Linyang Li, Shuaibin Li, Wei Li, Yinling Li, Hongwei Liu, Jiangning Liu, Jiawei Hong, Kaiwen Liu, Kuikun Liu, Xiaoran Liu, Chengqi Lv, Haijun Lv, Kai Lv, Li Ma, Runyuan Ma, Zerun Ma, Wenchang Ning, Linke Ouyang, Jiantao Qiu, Yuan Qu, Fukai Shang, Yunfan Shao, Demin Song, Zifan Song, Zhihao Sui, Peng Sun, Yu Sun, Huanze Tang, Bin Wang, Guoteng Wang, Jiaqi Wang, Jiayu Wang, Rui Wang, Yudong Wang, Ziyi Wang, Xingjian Wei, Qizhen Weng, Fan Wu, Yingtong Xiong, Chao Xu, Ruiliang Xu, Hang Yan, Yirong Yan, Xiaogui Yang, Haochen Ye, Huaiyuan Ying, Jia Yu, Jing Yu, Yuhang Zang, Chuyu Zhang, Li Zhang, Pan Zhang, Peng Zhang, Ruijie Zhang, Shuo Zhang, Songyang Zhang, Wenjian Zhang, Wenwei Zhang, Xingcheng Zhang, Xinyue Zhang, Hui Zhao, Qian Zhao, Xiaomeng Zhao, Fengzhe Zhou, Zaida Zhou, Jingming Zhuo, Yicheng Zou, Xipeng Qiu, Yu Qiao and Dahua Lin. Internl m2 technical report, 2024. 2  
+[5] Jieneng Chen, Luoxin Ye, Ju He, Zhao-Yang Wang, Daniel Khashabi, and Alan Yuille. Llavolta: Efficient multi-modal models via stage-wise visual context compression. arXiv preprint arXiv:2406.20092, 2024. 2  
+[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 2, 5  
+[7] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for
+
+large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 2, 5  
+[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 1, 2  
+[9] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, Tianren Gao, Erlong Li, Kun Tang, Zhipeng Cao, Tong Zhou, Ao Liu, Xinrui Yan, Shuqi Mei, Jianguo Cao, Ziran Wang, and Chao Zheng. A survey on multimodal large language models for autonomous driving, 2023. 1  
+[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 5  
+[11] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5  
+[12] 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, 2017. 1, 2  
+[13] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 5  
+[14] Lianyu Hu, Fanhua Shang, Liang Wan, and Wei Feng. illava: An image is worth fewer than 1/3 input tokens in large multimodal models. arXiv preprint arXiv:2412.06263, 2024. 3  
+[15] Omri Kaduri, Shai Bagon, and Tali Dekel. What's in the image? a deep-dive into the vision of vision language models. arXiv preprint arXiv:2411.17491, 2024. 5  
+[16] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5  
+[17] 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, 2  
+[18] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation, 2023. 1  
+[19] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 5  
+[20] Yuelei Li, Ge Yan, Annabella Macaluso, Mazeyu Ji, Xueyan Zou, and Xiaolong Wang. Integrating lmm planners and 3d
+
+skill policies for generalizable manipulation. arXiv preprint arXiv:2501.18733, 2025. 1  
+[21] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2  
+[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2  
+[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5  
+[24] Jiaming Liu, Chenxuan Li, Guanqun Wang, Lily Lee, Kaichen Zhou, Sixiang Chen, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, and Shanghang Zhang. Self-corrected multimodal large language model for end-to-end robot manipulation, 2024. 1  
+[25] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 5  
+[26] 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. 5  
+[27] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5  
+[28] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5  
+[29] 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  
+[30] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5  
+[31] Haoxiang Tian, Xingshuo Han, Guoquan Wu, Yuan Zhou, Shuo Li, Jun Wei, Dan Ye, Wei Wang, and Tianwei Zhang. An llm-enhanced multi-objective evolutionary search for autonomous driving test scenario generation. arXiv preprint arXiv:2406.10857, 2024. 1  
+[32] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn,
+
+Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 2  
+[33] Ao Wang, Fengyuan Sun, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. [cls] token tells everything needed for training-free efficient mllms. arXiv preprint arXiv:2412.05819, 2024. 1, 2  
+[34] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 5  
+[35] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, Hao Tian, Lewei Lu, Xizhou Zhu, Xiaogang Wang, Yu Qiao, and Jifeng Dai. Drivemlm: Aligning multimodal large language models with behavioral planning states for autonomous driving, 2023. 1  
+[36] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 1, 2, 5  
+[37] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 4  
+[38] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2  
+[39] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin,
+
+Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 2  
+[40] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 1, 2, 5  
+[41] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 3  
+[42] Weihao Ye, Qiong Wu, Wenhao Lin, and Yiyi Zhou. Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. arXiv preprint arXiv:2409.10197, 2024. 1, 2, 5  
+[43] Xubing Ye, Yukang Gan, Yixiao Ge, Xiao-Ping Zhang, and Yansong Tang. Atp-llava: Adaptive token pruning for large vision language models. arXiv preprint arXiv:2412.00447, 2024. 2  
+[44] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 1, 2  
+[45] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark, 2024. 1  
+[46] 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. 2  
+[47] Jun Zhang, Desen Meng, Ji Qi, Zhenpeng Huang, Tao Wu, and Limin Wang. p-mod: Building mixture-of-depths mllms via progressive ratio decay. arXiv preprint arXiv:2412.04449, 2024. 2  
+[48] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 2  
+[49] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster, 2024. 1, 2  
+[50] Shaolei Zhang, Qingkai Fang, Zhe Yang, and Yang Feng. Llava-mini: Efficient image and video large multimodal models with one vision token. In The Thirteenth International Conference on Learning Representations. 2  
+[51] Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu,
+
+Hao Tang, and Jieping Ye. From redundancy to relevance: Information flow in lvlms across reasoning tasks. arXiv preprint arXiv:2406.06579, 2024. 4  
+[52] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 1, 5  
+[53] Zhi Zhang, Srishti Yadav, Fengze Han, and Ekaterina Shutova. Cross-modal information flow in multimodal large language models. arXiv preprint arXiv:2411.18620, 2024. 5

2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4fb6be8870c5452c86ce6bf00456960ec966e6551d96f2a4d445ea182d9284b
+size 649996

2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/a6042e74-fb7f-4488-a331-f3b644cd7772_content_list.json

@@ -0,0 +1,1617 @@
+[
+    {
+        "type": "image",
+        "img_path": "images/68614e85b835a66db0667fcd63c2b5cc77b7b39f4a49010171318974b8b9b39b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            156,
+            126,
+            192,
+            151
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "VGGSounder: Audio-Visual Evaluations for Foundation Models",
+        "text_level": 1,
+        "bbox": [
+            194,
+            130,
+            841,
+            150
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Daniil Zverev<sup>1,*</sup> Thaddäus Wiedemer<sup>2,3,4,*</sup> Ameya Prabhu<sup>2,3</sup>",
+        "bbox": [
+            251,
+            179,
+            743,
+            199
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Matthias Bethge $^{2,3}$  Wieland Brendel $^{3,4}$  A. Sophia Koepke $^{1,2,3}$",
+        "bbox": [
+            243,
+            202,
+            751,
+            222
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Technical University of Munich, MCML  $^{2}$ University of Tübingen  $^{3}$ Tübingen AI Center  $^{4}$ MPI for Intelligent Systems, ELLIS Institute Tübingen",
+        "bbox": [
+            129,
+            229,
+            869,
+            273
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{daniil.zverev,a-sophia.koepke}@tum.de,{thaddaeus.wiedemer,ameya.prabhu}@uni-tuebingen.de",
+        "bbox": [
+            102,
+            279,
+            887,
+            295
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            246,
+            329,
+            326,
+            345
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The emergence of audio-visual foundation models underscores the importance of reliably assessing their multimodal understanding. The VGGSound dataset is commonly used as a benchmark for evaluation audio-visual classification. However, our analysis identifies several limitations of VGGSound, including incomplete labelling, partially overlapping classes, and misaligned modalities. These lead to distorted evaluations of auditory and visual capabilities. To address these limitations, we introduce VGGSounder, a comprehensively re-annotated, multi-label test set that extends VGGSound and is specifically designed to evaluate audio-visual foundation models. VGGSounder features detailed modality annotations, enabling precise analyses of modality-specific performance. Furthermore, we reveal model limitations by analysing performance degradation when adding another input modality with our new modality confusion metric. Our dataset and project page are available at https://vqgsounder.github.io/",
+        "bbox": [
+            89,
+            362,
+            483,
+            633
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            89,
+            666,
+            220,
+            681
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Rigorous evaluation benchmarks have been instrumental in assessing the effectiveness of audio-visual models [33, 43, 49, 57]. Specifically, multi-modal foundation models integrating visual and auditory data aim to achieve a holistic understanding of audio-visual content. However, the field lacks large-scale modality-aware classification benchmarks with ground-truth annotations indicating whether each label is visible, audible, or both. Such annotations would allow detailed evaluations of multi-modal model capabilities. To address this gap, we introduce VGGSounder, an enhanced version of the widely-used audio-visual classification dataset VGGSound [13], which facilitates modality-",
+        "bbox": [
+            88,
+            691,
+            482,
+            875
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/2b489317aa6ca2aff9ddd7b4f939ceccbb63d7e5eb5c60b2e953cee0770ab8f8.jpg",
+        "image_caption": [
+            "Figure 1. We introduce VGGSounder, a multi-label audio-visual classification benchmark with modality annotations. We extend the original VGGSound test set with human-annotated audible, visible, and visible+audible labels. We add meta labels for common confounders, such as background music. We benchmark eleven recent audio-visual models on VGGSounder. It enables selective analysis of a model's auditory and visual capabilities on classes relevant for the queried modality."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            329,
+            903,
+            570
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "aware evaluation of audio-visual foundation models.",
+        "bbox": [
+            511,
+            715,
+            861,
+            731
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "VGGSound has several notable limitations. First, its data is inherently multi-label; for instance, a single sample may simultaneously include labels such as playing drum kit and playing acoustic guitar when multiple instruments are present. Additionally, evaluating how different modalities contribute to model performance becomes difficult without explicit modality annotations, as some labels are either not visually present or not audible (e.g., certain instruments might only be audible but not visible in advertisements). Moreover, overlapping label classes present another challenge; for example,",
+        "bbox": [
+            509,
+            734,
+            906,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV 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": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*equal contribution",
+        "bbox": [
+            109,
+            887,
+            215,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "1027",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "the orchestra label often coincides with labels for individual instruments. These issues result in systematic under-evaluation of multi-modal audio-visual foundation models.",
+        "bbox": [
+            89,
+            90,
+            480,
+            135
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To overcome these limitations, we present VGGSounder, an improved benchmark inspired by similar advancements in other domains [12, 28]. We re-announce the dataset to create a comprehensive multi-label classification setting by collecting detailed annotations for each sample, including (1) additional classes present, (2) explicit modality annotations to label modality misalignment, (3) metadata indicating the presence of background music, voice-over, or static images, and (4) merging of classes to address overlapping classes. Consequently, VGGSounder provides a robust, foundation-model-ready benchmark enabling structured analysis of whether models rely on audio or visual cues. Furthermore, we include meta-labels (e.g., background music, voice-over, or static images) to easily filter out unreliable labels during evaluation. Utilising VGGSounder, we evaluate audio-visual foundation models, demonstrating their poor performance on our benchmark. We find that the state-of-the-art, closed-source Gemini models consistently rely exclusively on the visual modality. In addition to that, we measure the modality confusion, i.e. when models get distracted by an additional input modality, which exposes the unsuccessful merging of modalities. These findings highlight the importance of the audio-focused VGGSounder benchmark as a critical tool for accurately assessing audio-visual foundation models.",
+        "bbox": [
+            91,
+            138,
+            483,
+            513
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We make the following contributions:",
+        "bbox": [
+            102,
+            532,
+            352,
+            547
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. We illustrate limitations of VGGSound in Sec. 3",
+            "2. We curate VGGSounder with multi-modal human annotations for multi-label classification in Sec. 4",
+            "3. We evaluate state-of-the-art audio-visual models, observing differences between embedding models and autoregressive foundation models in Sec. 5.",
+            "4. We propose new metrics to quantify the negative impact of using multiple input modalities in Sec. 5."
+        ],
+        "bbox": [
+            101,
+            547,
+            470,
+            670
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related work",
+        "text_level": 1,
+        "bbox": [
+            89,
+            689,
+            227,
+            705
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Audio-visual learning Many prior works consider audio-visual tasks that include sound source localisation and separation [3, 5, 9, 15, 27, 56, 62, 67, 75, 80, 85, 86, 90], event localisation [50, 51, 74, 78], audio-visual question answering [48, 54, 83, 84], audio-visual synchronisation [14, 23, 25, 38, 39, 42], audio synthesis using visual information [19, 26, 31, 44, 45, 61, 69-71, 87], or audio-driven face image synthesis [7, 40, 77]. Audio-visual data has also been leveraged for speech-related tasks, including speech and speaker recognition [2, 4, 59], or the spotting of spoken keywords [58, 66].",
+        "bbox": [
+            89,
+            719,
+            482,
+            883
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, the natural alignment between audio and",
+        "bbox": [
+            89,
+            885,
+            482,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/3086e0733e283efca0637873a1cab8fd54187d2014223a4b780e6becda7f20a0.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td># Clips</td><td># Classes</td><td>Multi- label</td><td>Modality labels</td><td>Annotation pipeline</td></tr><tr><td>Flickr-SoundNet [10]</td><td>2M</td><td>-</td><td>✘</td><td>✘</td><td>-</td></tr><tr><td>Kinetics-Sound [7]</td><td>18.8k</td><td>34</td><td>✘</td><td>✘</td><td>MTurk</td></tr><tr><td>AudioSet [26]</td><td>2.1M</td><td>537</td><td>✓</td><td>✘</td><td>Manual</td></tr><tr><td>AVE [62]</td><td>4k</td><td>28</td><td>✓</td><td>✘</td><td>Manual</td></tr><tr><td>VEGSAS [76]</td><td>132k</td><td>10</td><td>✘</td><td>✘</td><td>MTurk</td></tr><tr><td>Visually Aligned Sounds [15]</td><td>13k</td><td>8</td><td>✘</td><td>✘</td><td>MTurk</td></tr><tr><td>VGGSound [12]</td><td>200k</td><td>309</td><td>✘</td><td>✘</td><td>Classifiers+Manual</td></tr><tr><td>VGGSound-Sparse [32]</td><td>7.1k</td><td>12</td><td>✘</td><td>✘</td><td>Manual</td></tr><tr><td>Visual Sound [66]</td><td>91k</td><td>309</td><td>✘</td><td>✘</td><td>ImageBind [30]</td></tr><tr><td>VGGSounder</td><td>15.4k</td><td>309</td><td>✓</td><td>✓</td><td>MTurk</td></tr></table>",
+        "bbox": [
+            517,
+            88,
+            903,
+            198
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Comparison of audio-visual classification benchmarks.",
+        "bbox": [
+            519,
+            208,
+            897,
+            220
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "video has been exploited to learn improved audio-visual embeddings for downstream tasks [6, 10, 11, 18, 20, 21, 46, 60, 63, 65, 79]. Using both modalities jointly generally leads to performance boosts over using one modality in isolation. We examine this observation closely and aim to evaluate the effective use of multiple input modalities for the video classification task. To enable this, we propose — to the best of our knowledge, the first multi-label video classification benchmark that includes per-modality annotations for every sample (see Tab. 1).",
+        "bbox": [
+            511,
+            241,
+            903,
+            393
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Audio-visual foundation models Recently, multi-modal general-purpose models have emerged that can handle diverse downstream tasks without task-specific finetuning - also referred to as multi-modal foundation models. For instance, images or language were used as the bridge between modalities including audio, image, and text [30, 89]. Building on this, PandaGPT [72] leverages Vicuna [22] and ImageBind's embedding space to train a general multi-modal model exclusively on image-text pairs. Unified-IO 2 [53] employs universal tokenisation to process audio, video, and text. VideoLLaMA2 [21] uses a Spatial-Temporal Convolution connector in the visual branch before projecting audio and visual information into the LLM input space. The recently introduced Ola model [52] advances omnimodal processing through progressive modality alignment, using video to bridge audio and visual information. The Gemini models [73] are closed-source multi-modal models that achieve impressive performance on diverse downstream tasks. We use VGGSounder to benchmark the audio-visual capabilities of the aforementioned models.",
+        "bbox": [
+            511,
+            397,
+            906,
+            698
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Audio-visual classification benchmarks Audio-visual classification is distinct from general video classification (e.g. on YouTube-8M [1]), as classes typically cover both audible and visible actions or events. Commonly used datasets for audio-visual classification include KineticsSound [8] sourced from the Kinetics dataset [41], Flickr-SoundNet [11] scraped from Flickr, and AudioSet [29] and VGGSound [13], both sourced from YouTube.",
+        "bbox": [
+            511,
+            703,
+            903,
+            823
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Kinetics-Sound features manual labels of human actions, but covers only 34 classes. Flickr-SoundNet is much larger, but only a small subset is labelled. Similarly, only a small fraction of the roughly 2M AudioSet samples are annotated and have aligned audio and video.",
+        "bbox": [
+            511,
+            825,
+            903,
+            898
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "1028",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Figure 2. Limitations of VGGSound. We show video frames from videos in the VGGSound test set along with their annotated label (grey) to demonstrate various limitations. A. VGGSound samples are labelled with a single class, yet many videos contain multiple distinct classes. B. Additionally, many classes partially overlap or are ambiguous. C. Some samples are labelled with classes that are not present in one of the modalities (i.e., the labelled class is not visible or audible).",
+        "bbox": [
+            89,
+            292,
+            906,
+            349
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In contrast, VGGSound ensures audio-visual correspondence for around 200 000 samples and was curated with an automatic pipeline involving class-list generation, and auditory and visual content verification. The visual verification step ensures that a class is represented in the centre frame.",
+        "bbox": [
+            88,
+            369,
+            480,
+            444
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The VEGAS dataset [88] provides better quality assurances for a small subset of AudioSet with only 10 classes. Visually Aligned Sounds [16] subsamples VEGAS and AudioSet after human verification, and Visual Sound [76] subsamples VGGSound using a multi-modal embedding model, both aiming for high audio-visual correspondence. Similarly, VGGSound-Sparse [37] is a subset of VGGSound with a focus on temporally and spatially sparse synchronisation signals (e.g., short loud noises).",
+        "bbox": [
+            88,
+            444,
+            482,
+            580
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Overall, VGGSound strikes the best balance between size, generality, and annotations, making it a common benchmark for audio-visual classification. We update VGGSound to sustain its usability for the development of the next generation of multi-modal foundation models.",
+        "bbox": [
+            89,
+            580,
+            482,
+            655
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Limitations of VGGSound",
+        "text_level": 1,
+        "bbox": [
+            89,
+            667,
+            336,
+            684
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Since we are interested in the VGGSound dataset for benchmarking audio-visual multi-modal models, our analysis focuses on the VGGSound test set, which consists of 15446 video clips, each 10s long and labelled with one of 309 classes. We qualitatively identify several limitations of the VGGSound annotations outlined below and in Fig. 2.",
+        "bbox": [
+            89,
+            693,
+            482,
+            785
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Co-occurring classes While VGGSound's visual verification aimed to minimise multiple classes co-occurring in a clip, we find that most samples nevertheless clearly contain multiple classes, see Fig. 2A. In some cases, classes are temporally separated, e.g., showing male speech, man speaking and then cutting to",
+        "bbox": [
+            89,
+            787,
+            482,
+            878
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "footage of firing cannon. Most often, classes co-occur at the same time, sometimes for the entire duration of the video clip. Overlapping classes are often related, such as different instruments in a band or orchestra, but can also be entirely unrelated, e.g., donkey, ass braying co-occurring with playing violin. As additional empirical evidence, we provide a co-occurrence matrix computed using CAV-MAE [33], a state-of-the-art audio-visual model, in Appendix D.",
+        "bbox": [
+            511,
+            369,
+            906,
+            507
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Overlapping classes The issue of co-occurring classes is exacerbated by many of the 309 automatically generated classes partially overlapping in their definition, as illustrated in Fig. 2B. We found two pairs of synonymous classes: timpani and tympani and dog barking and dog bow-wow. Additionally, some classes are strict subclasses of others, such as the gender-specific versions of cattle moing: cow lowing and bull bellowing; or the more specific variants of people eating: people eating noodle and people eating apple. Finally, several classes commonly appear together, such as playing snare drum which is often played as part of a drum kit, or semantically similar concepts: running electric fan and air conditioning noise, and sloshing water and splashing water.",
+        "bbox": [
+            511,
+            513,
+            908,
+            755
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Modality misalignment Despite VGGSound's auditory and visual content verification, we find that many of the annotated classes are not visible or not audible, as shown in Fig. 2C. A large fraction of videos contains background music, voice-over and narration, or other background sounds like bird chirping, tweeting or cricket chirping without a visible source. Similarly, some videos contain visible but inaudible cues for classes like sea waves. Static images and slide shows accompa",
+        "bbox": [
+            511,
+            763,
+            908,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "1 Although these issues most likely also apply to the training set.",
+        "bbox": [
+            107,
+            886,
+            447,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "1029",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/7c9ba699608f529a550d998b95fefef92562d56bd4f6c77fefb12df0bb131ed6.jpg",
+        "image_caption": [
+            "Figure 3. Overview of VGGSounder. A. Most samples contain more than one label. B. More than a quarter of labels are audible but not visible. In contrast, only a tiny fraction is visible but not audible. C. Speech and bird sounds are the most common classes; more details can be found in Appendix B. D. Forty percent of the samples contain some combination of background music, voice over, and static image(s), making the classification task significantly harder."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            93,
+            90,
+            339,
+            270
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/94fbb8f345cb4997d6a8c853c6e957bfcd248d0cf8614eb593fba6ff5bc8f081.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            362,
+            123,
+            514,
+            239
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/d3c1a4923a33d3f9b62441bb900fbf905856b1e68189ceeca9616911ae16440e.jpg",
+        "image_caption": [
+            "B Label modality distribution",
+            "C Most common labels"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            531,
+            104,
+            707,
+            169
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/c63d15d87efbc62ba4c4c351ed68e88dbdc145a6c81095ccf49958c67c05d52b.jpg",
+        "image_caption": [
+            "D Meta labels"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            722,
+            104,
+            897,
+            260
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "nied by music or other sounds are other frequent sources of misaligned modalities. Finally, some classes are misaligned by definition: wind noise is only audible and not visible. Overall,  $48.43\\%$  of the original VGGSound test samples have misaligned modalities. This finding challenges the widely held assumption that VGGSound has strong modality alignment [32, 47].",
+        "bbox": [
+            88,
+            356,
+            482,
+            460
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Other datasets, such as Visually Aligned Sounds, Visual Sound, and VGGSound-Sparse (see Sec. 2) omitted samples with misaligned modalities. In contrast, we contend that inaudible or invisible cues are common in natural videos and should be considered when benchmarking multi-modal models. We, therefore, place particular emphasis on crafting reliable modality annotations for all samples, allowing users to evaluate models on samples that guarantee modality alignment, and on those where classes are only visible or only audible (see Tab. 1).",
+        "bbox": [
+            88,
+            462,
+            482,
+            613
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Takeaway 1 VGGSound suffers from several issues: class co-occurrence not captured by single labels, overlapping class definitions, and modality misalignment, see Fig. 2.",
+        "bbox": [
+            99,
+            628,
+            473,
+            691
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4. Building VGGSounder",
+        "text_level": 1,
+        "bbox": [
+            89,
+            709,
+            305,
+            724
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We propose a series of fixes for VGGSound's issues, ultimately resulting in the updated VGGSounder benchmark. We are not the first aiming to future-proof an existing benchmark: [12] analysed shortcomings of ImageNet [24], ultimately proposing switching to a multi-label classification task with additional manual labels. [28] similarly reannotated samples in MMLU [35, 36] to fix labelling errors. Both works inspire our approach to improve VGGSound. To deal with co-occurring classes, we switch to a multi-label classification setting. This effectively handles most overlapping class definitions: a strong model can assign a",
+        "bbox": [
+            89,
+            734,
+            482,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "high probability to multiple classes, even if they partially overlap. This also allows us to ensure that synonymous classes, as well as subclasses and their superclass, always appear together in the ground-truth labels.",
+        "bbox": [
+            511,
+            356,
+            903,
+            416
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To deal with modality misalignment, we add a modality annotation to each label. For example, we can label a video as containing people clapping in the audio and containing playing volleyball in the audio and video data. We also add meta-labels to indicate whether a sample contains background music, voice over, or static image(s) to optionally treat these cases separately during evaluation.",
+        "bbox": [
+            511,
+            417,
+            905,
+            539
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We employ a pipeline similar to [12] to annotate multiple labels per sample, which we outline below.",
+        "bbox": [
+            511,
+            541,
+            903,
+            570
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Collecting proposals We create a gold standard reference set by labelling a small, randomly selected subset of VGGSound test samples with four in-house computer vision experts. The interface used for this first labelling step is shown in Appendix A. We extend the subset until each class is covered at least once, leading to a final size of 417 samples. Labels from different annotators are merged via a simple majority vote.",
+        "bbox": [
+            511,
+            575,
+            905,
+            696
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Given the gold standard set, we want to find a solid strategy for automatically generating label proposals which are shown to the humans labelling the test set. This should have a recall greater than  $90\\%$  while maximising precision compared to the gold standard labels to produce a small set of proposals with good label coverage. Our final strategy combines predictions from several state-of-the-art models with a manual heuristic to obtain  $93\\%$  recall for an average of 30 proposals per sample, see Appendix A.",
+        "bbox": [
+            511,
+            699,
+            905,
+            835
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Human labelling We use Amazon Mechanical Turk to re-associate the entire VGGSound test set. For each sample, we first ask annotators to indicate whether the video contains background music, voice over, or",
+        "bbox": [
+            511,
+            839,
+            905,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "1030",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "static image(s). Then, annotators are asked to indicate for each label proposal whether the class is audible and/or visible. Finally, annotators can add a class if it is missing from the proposals. Annotators were paid the US minimum wage; the interface used is shown in Appendix A. We let annotators label the samples in batches of 20, each containing two gold standard samples as catch trials. We reject and re-associate all batches with a catch trial  $F_{1}$ -score below  $25\\%$ . In addition, we obtain modality annotations for the original VGGSound labels and meta-classes. Further details are provided in Appendix A.",
+        "bbox": [
+            89,
+            90,
+            480,
+            257
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Final labels We merge all obtained annotations using majority voting. Additionally, we automatically add synonymous classes and superclasses for a given subclass, e.g., we add cattle moing whenever cow lowing is in the set of labels. The full set of classes added this way is described in Appendix A.",
+        "bbox": [
+            89,
+            260,
+            480,
+            351
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Takeaway 2 We develop VGGSounder: A multi-label video classification benchmark extending VGGSound with human-annotated multi-labels, modality annotations, and meta-labels as summarised in Fig. 3",
+        "bbox": [
+            99,
+            366,
+            473,
+            429
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5. Benchmarking audio-visual models",
+        "text_level": 1,
+        "bbox": [
+            89,
+            445,
+            410,
+            462
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We use VGGSounder to benchmark four popular audiovisual embedding models and seven foundation models, and analyse their auditory and visual capabilities.",
+        "bbox": [
+            89,
+            470,
+            482,
+            516
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Models We evaluate the audio-visual embedding models CAV-MAE [33], DeepAVFusion [57], AV-Siam [49], and Equi-AV [43]. Those were finetuned on VGGSound.",
+        "bbox": [
+            89,
+            518,
+            482,
+            563
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We benchmark several models from the closed-source Gemini family [73] in a zero-shot evaluation protocol. Furthermore, we use LLM-assisted evaluation to evaluate the following four open-source autoregressive foundation models: VideoLLaMA-2 [21], Unified-IO-2 [53], Panda-GPT [72], and Ola [52]. All models are evaluated in three modes: using unimodal-audio, unimodal-visual, or multi-modal (audio and visual) inputs. Further details about models and their evaluation are provided in Appendix C.1",
+        "bbox": [
+            89,
+            564,
+            482,
+            700
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Metrics To benchmark the models on VGGSounder, we use multi-label classification metrics. For embedding models, all metrics are computed for the top- $k$  predictions, with  $k \\in \\{1, 3, 5, 10\\}$ . In contrast, prompting foundation models yields an unordered set of class predictions of varying size, and we compute only a single metric using the entire set. As a result, metrics are not directly comparable between embedding and foundation models. To get a sense of their relative performance, we report metrics for embedding models for  $k = 1$  in the main text, matching the median number of predictions per sample for the foundation models.",
+        "bbox": [
+            89,
+            704,
+            482,
+            869
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For open-source models such as VideoLLaMA-2, Ola, Unified-I0-2, and Panda-GPT, we employ LLM-assisted",
+        "bbox": [
+            89,
+            869,
+            482,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "evaluation [55, 81], in which the Qwen-3 model [82] is tasked to assess the correspondence between model outputs and target classes. Closed-source models from the Gemini family are evaluated by providing the full list of 309 classes as input. Further details on the evaluation procedures and exact prompts are provided in Appendix C",
+        "bbox": [
+            511,
+            90,
+            903,
+            181
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Subset accuracy compares the predicted label set to the ground-truth label set and reports the fraction of samples for which they match. This is our strictest metric.",
+        "bbox": [
+            511,
+            181,
+            903,
+            226
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "$F_{1}$ -score is the harmonic mean of precision and recall. It is strictly larger than the subset accuracy.",
+        "bbox": [
+            513,
+            227,
+            903,
+            257
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Hit reports the fraction of samples for which any of the predicted labels are part of the ground-truth label set. This is the most lenient metric which is strictly larger than the  $F_{1}$  score. We include this metric for ease of comparison to the \"ReaL-Accuracy\" used in [12].",
+        "bbox": [
+            511,
+            257,
+            903,
+            332
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "All metrics are computed separately for each input modality (audio, video, and audio-visual) and label modality. We use lowercase symbols a, v, and av to indicate the input modality: audio-only, visual-only, or audio-visual inputs, respectively. For label modality, we use uppercase symbols  $A$ ,  $V$ , and  $AV$ , referring to the subsets of the benchmark with audible, visible, and audio-visual labels. We further include  $A \\neg V$  (audible but not visible) and  $V \\neg A$  (visible but not audible) to analyze unaligned cues. For clarity, we define shorthand notations such as  $a = a(A)$  to denote the model's performance on the audible subset  $A$  using only audio input. Analogously,  $v = v(V)$  and  $av = av(AV)$  refer to video-only and audio-visual performance on their respective label subsets. Furthermore, we use micro-aggregation to balance the contribution from each class.",
+        "bbox": [
+            511,
+            333,
+            905,
+            558
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We additionally measure the negative impact of using multimodal inputs. In particular,  $\\mu$  is a new metric we propose to measure a model's modality confusion  $(\\mu)$ . We define it as",
+        "bbox": [
+            511,
+            559,
+            903,
+            604
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mu_ {\\mathrm {M}} = 1 0 0 \\cdot \\frac {\\sum_ {x \\in M} \\mathcal {I} [ m (x) - \\text {c o r r e c t} \\cap a v (x) - \\text {w r o n g} ]}{N _ {\\text {t o t a l}}}, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            519,
+            611,
+            903,
+            643
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $M \\in [A, V, A \\cap V]$  and their associated modality inputs are  $m \\in [a, v]$ , correct/wrong is determined as in the Hit score (with  $k = 1$  for embedding models).  $N_{total}$  refers to the total number of samples.  $\\mu$  measures the fraction of samples a model correctly classified given an input modality but got wrong when using both modalities simultaneously. We additionally report  $\\mu_{A \\cap V}$  as the percentage of samples a model could solve in either modality unimodally but could not solve multi-modally. In other words, the modality confusion  $\\mu$  captures how frequently a model is distracted by an additional input modality, which can indicate the unsuccessful merging of modalities.",
+        "bbox": [
+            511,
+            651,
+            906,
+            834
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Takeaway 3 We propose a new metric, modality confusion  $\\mu$ , that measures how frequently a model is distracted by an additional input modality; see Eq. (1).",
+        "bbox": [
+            521,
+            847,
+            897,
+            895
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "1031",
+        "bbox": [
+            483,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/7f3c2bebd34a8528a23fab9c42cd1cf3cda07f540249bab807c8564bc3c9e197.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"3\">Subset Accuracy ↑</td><td colspan=\"5\">F1↑</td><td colspan=\"3\">Hit ↑</td><td colspan=\"3\">μ↓</td></tr><tr><td>a</td><td>v</td><td>av</td><td>a</td><td>v</td><td>av</td><td>a(A-√V)</td><td>v(V-√A)</td><td>a</td><td>v</td><td>av</td><td>μA</td><td>μV</td><td>μA∩V</td></tr><tr><td colspan=\"15\">Embedding Models</td></tr><tr><td>CAV-MAE</td><td>13.19</td><td>19.23</td><td>24.49</td><td>34.46</td><td>34.91</td><td>42.62</td><td>13.94</td><td>19.00</td><td>62.29</td><td>53.44</td><td>64.17</td><td>3.58</td><td>6.43</td><td>0.77</td></tr><tr><td>DeepAVFusion</td><td>10.19</td><td>11.10</td><td>21.53</td><td>25.31</td><td>21.29</td><td>37.35</td><td>10.37</td><td>10.55</td><td>45.77</td><td>32.61</td><td>56.27</td><td>3.74</td><td>3.93</td><td>0.17</td></tr><tr><td>Equi-AV</td><td>11.60</td><td>10.52</td><td>20.00</td><td>29.39</td><td>20.42</td><td>34.69</td><td>12.55</td><td>10.65</td><td>53.12</td><td>31.26</td><td>52.24</td><td>6.97</td><td>7.13</td><td>1.38</td></tr><tr><td>AV-Siam</td><td>12.79</td><td>19.75</td><td>22.83</td><td>33.30</td><td>35.41</td><td>39.43</td><td>12.90</td><td>18.21</td><td>60.19</td><td>54.20</td><td>59.36</td><td>9.36</td><td>8.80</td><td>3.58</td></tr><tr><td colspan=\"15\">Closed-source Foundation Models</td></tr><tr><td>Gemini 1.5 Flash</td><td>1.78</td><td>14.44</td><td>16.44</td><td>14.49</td><td>36.98</td><td>42.52</td><td>15.61</td><td>21.61</td><td>32.73</td><td>47.36</td><td>59.10</td><td>10.22</td><td>4.25</td><td>0.77</td></tr><tr><td>Gemini 1.5 Pro</td><td>3.05</td><td>20.86</td><td>22.53</td><td>19.26</td><td>49.73</td><td>53.74</td><td>17.73</td><td>22.90</td><td>35.03</td><td>69.23</td><td>75.42</td><td>2.09</td><td>4.85</td><td>0.57</td></tr><tr><td>Gemini 2.0 Flash</td><td>1.85</td><td>12.54</td><td>12.69</td><td>11.80</td><td>34.08</td><td>36.45</td><td>6.19</td><td>18.90</td><td>18.51</td><td>43.83</td><td>47.72</td><td>2.39</td><td>5.43</td><td>1.00</td></tr><tr><td colspan=\"15\">Open-source Foundation Models</td></tr><tr><td>VideoLLaMA 2</td><td>12.86</td><td>19.85</td><td>24.47</td><td>38.87</td><td>47.82</td><td>52.35</td><td>20.34</td><td>28.08</td><td>58.91</td><td>52.02</td><td>59.80</td><td>12.72</td><td>5.46</td><td>2.95</td></tr><tr><td>Unified-IO 2</td><td>11.94</td><td>11.56</td><td>25.61</td><td>35.31</td><td>27.92</td><td>48.89</td><td>21.38</td><td>16.53</td><td>54.39</td><td>31.05</td><td>65.11</td><td>8.70</td><td>5.16</td><td>1.79</td></tr><tr><td>PandaGPT</td><td>3.19</td><td>4.19</td><td>5.46</td><td>18.73</td><td>18.56</td><td>20.85</td><td>16.82</td><td>14.40</td><td>21.08</td><td>17.01</td><td>18.82</td><td>7.59</td><td>5.90</td><td>2.47</td></tr><tr><td>OLA</td><td>14.11</td><td>8.69</td><td>18.19</td><td>47.70</td><td>24.85</td><td>46.48</td><td>40.44</td><td>13.45</td><td>59.05</td><td>24.57</td><td>51.51</td><td>15.47</td><td>6.80</td><td>2.49</td></tr></table>",
+        "bbox": [
+            94,
+            88,
+            903,
+            327
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Audio-visual video classification results on VGGSounder. We report multi-label classification metrics (subset accuracy,  $F_{1}$ -score, Hit accuracy, modality confusion  $(\\mu)$  for audio-  $a(A)$ , visual -  $v(V)$ , audio-visual -  $av(AV)$ , audio-only -  $a(A\\neg V)$  and video-only -  $v(V\\neg A)$  inputs. The embedding models CAV-MAE, DeepAVFusion, and Equi-AV were finetuned on the VGGSound training set. We report metrics for  $k = 1$  here and for other  $k$  in Appendix D The closed sourced multi-modal foundation models Gemini and open-sourced models use a zero-shot evaluation protocol and LLM-assisted protocol respectively.",
+        "bbox": [
+            89,
+            338,
+            906,
+            409
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. Re-evaluating the state of the art",
+        "text_level": 1,
+        "bbox": [
+            89,
+            424,
+            377,
+            441
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We present the benchmark performance of state-of-the-art audio-visual models in Tab. 2.",
+        "bbox": [
+            89,
+            446,
+            482,
+            477
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Overall performance Unsurprisingly, all models perform best with access to both input modalities (AV). Across all metrics, both open- and closed-source general-purpose models perform comparably to the purpose-built embedding models CAV-MAE and AV-Siam. This indicates that foundation models have reached — and for some modalities exceeded — the performance of specialised models. However, the embedding models finetuned on VGGSound generally have stronger unimodal performance with audio inputs  $(A)$  compared to visual inputs  $(V)$ , which is in line with their pretraining on AudioSet. Interestingly, this trend is reversed for most foundation models, which seem to be biased towards visual inputs, with unimodal video performance  $(V)$  being substantially higher than unimodal audio performance  $(A)$ .",
+        "bbox": [
+            88,
+            479,
+            482,
+            708
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Takeaway 4 Foundation models perform comparably to finetuned embedding models. Embedding models more heavily rely on audio cues than on visual ones, while foundation models exploit visible cues rather than audible ones, see Tab. 2",
+        "bbox": [
+            99,
+            724,
+            473,
+            804
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Modality confusion The modality confusion score  $\\mu$  shows that, for all models, a notable fraction of test samples  $(4 - 11\\%)$  were misclassified when an additional modality was included—despite being correctly classified with unimodal input. Furthermore, for all models, a small portion of",
+        "bbox": [
+            89,
+            824,
+            483,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "test samples is not solvable multi-modally even though they were solvable in both modalities alone  $(\\mu_{A\\cap V})$ . This insight is made possible by VGGSounder's per-label modality annotations and shows that all models are susceptible to being distracted given an additional modality. This is a concerning issue for multi-modal models since they should preserve unimodal capabilities when adding a second modality. Being able to evaluate this behaviour on the VGGSounder benchmark is a first step towards enabling the development of mitigation strategies, eventually resulting in stronger audio-visual models.",
+        "bbox": [
+            511,
+            425,
+            906,
+            592
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Takeaway 5 Our modality confusion score reveals that all models are negatively impacted by additional modalities for a substantial amount of samples (see Tab. ②).",
+        "bbox": [
+            521,
+            606,
+            898,
+            655
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Performance across modalities Fig. 4 shows the performance profiles across modalities. At first glance, we can see that VideoLLaMA-2's performance is well balanced for different input modalities, while models from the Gemini family distinctly underperform on audio inputs. In contrast, embedding models exhibit a moderate balance across modalities, with DeepAVFusion and EquiAV showing slight underperformance for visual input.",
+        "bbox": [
+            511,
+            672,
+            906,
+            794
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As Fig. 4 also illustrates, profiling of this kind is enabled through the modality annotations in VGGSounder. In contrast, VGGSound assumed that all classes are perceivable in both modalities, and did not account for background sounds. This resulted in consistent under-evaluation of foundation models (that were not finetuned on VGGSound) for audio inputs.",
+        "bbox": [
+            511,
+            795,
+            908,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "1032",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In addition to the radar plot in Fig. 4, we provide results on VGGSound in Appendix D showing that all models have substantially lower performance than their hit scores in Tab. 2. This confirms that many model predictions were incorrectly flagged as false positives in VGGSound due to the incomplete ground-truth labels, painting a distorted picture of models' limitations.",
+        "bbox": [
+            89,
+            90,
+            483,
+            196
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/cc5cecfd4791f27111ac0dae1e53c38773b7a815acb789f7c7016fb2e18df776.jpg",
+        "image_caption": [
+            "Figure 4. VGGSounder more accurately captures model performance across input modalities. We show the Hit score on VGGSounder and accuracy on VGGSound, normalised by the per-model maximum performance on each benchmark. Specifically for foundation models, we observe a significant difference in performance between VGGSound and VGGSounder."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            106,
+            210,
+            473,
+            465
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Takeaway 6 VGGSounder's more complete ground-truth labels allow for more accurate, modality-specific profiling of model performance (see Fig.4).",
+        "bbox": [
+            99,
+            584,
+            475,
+            633
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2. Performance analysis using meta-classes",
+        "text_level": 1,
+        "bbox": [
+            89,
+            651,
+            436,
+            667
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "VGGSounder includes annotations of three meta-classes for each sample: background music indicates whether samples contain music without a visible source, voice over similarly marks speech without a visible source, and static image(s) flags that the visual stream consists of one or only a few static visual frames.",
+        "bbox": [
+            89,
+            672,
+            482,
+            763
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "These new meta-classes allow us to evaluate the model behaviour in challenging scenarios where information from one modality dominates. We consider the performance difference between samples that do contain a meta-label and samples that do not contain it. Positive numbers indicate that the models perform better on the subset with metal-labels. In Tab.3, we summarise the main findings, focussing on  $F_{1}$ -score as the most balanced multi-label metric. Additional results are provided in Appendix D",
+        "bbox": [
+            89,
+            763,
+            482,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Background music All models perform worse on video samples containing background music. This indicates that it is challenging to decouple background audio from the rest of the video. The evaluated models are not good at differentiating between sound sources without visual cues, e.g. predicting different instruments in the background music.",
+        "bbox": [
+            511,
+            90,
+            903,
+            181
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Voice over In contrast to background music, we observe a clear difference between embedding models and foundation models for samples with voice-over. While the audio classification performance of embedding models drops substantially. This drop is only slight for VideoLLaMA-2 and Unified-IO 2, and the performance of other foundation models even improves. This indicates that the foundation models are less distracted by voice-over.",
+        "bbox": [
+            511,
+            185,
+            906,
+            306
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Static image(s) The impact of static images is more nuanced: First, audio classification performance improves for embedding models while it decreases for the foundation models. This shows that the purpose-built, VGGSoundfinetuned embedding models can more accurately predict specific sounds in the absence of other cues.",
+        "bbox": [
+            511,
+            309,
+            903,
+            398
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Second, visual classification performance on static images drops for all models, suggesting that models rely on rich temporal cues to make accurate predictions.",
+        "bbox": [
+            511,
+            400,
+            903,
+            445
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Third, the subset accuracy (i.e. exact label set matches) for uni-modal predictions shows that foundation models perform better in visual than audio classification with and without static images. In contrast, embedding models perform better in audio classification on static images but worse in visual classification for other samples. This suggests that non-static samples form a challenging subset for embedding models, where a model needs to favour one modality over another one to make a correct prediction.",
+        "bbox": [
+            511,
+            446,
+            905,
+            582
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Samples without any meta-label When comparing the model performance on samples without background music, static images, or voice-over annotations (column neither in Tab. 3) to the performance on the entire test set, we see a performance gain (here negative numbers indicate an increased  $\\mathbf{F}_1$  score). This finding concludes that these three categories form challenging subsets of the dataset.",
+        "bbox": [
+            511,
+            585,
+            905,
+            691
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Takeaway 7 Samples with background music, static images(s), and voice over provide distinct challenges for each model (see Tab.3). This highlights VGGSounder's value for comprehensive model evaluation.",
+        "bbox": [
+            521,
+            707,
+            898,
+            768
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3. Impact of VGGSounder labels",
+        "text_level": 1,
+        "bbox": [
+            511,
+            787,
+            781,
+            803
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Our relabelling pipeline adds two types of labels to those in VGGSound: (1) automatically generated labels based on synonymous classes and sub-/superclasses, and (2) human-curated labels. In Tab.4, we ablate the impact of each type of added labels in terms of the relative performance gains (Hit score). A complete breakdown of the effects across all",
+        "bbox": [
+            511,
+            810,
+            905,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "1033",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/6e96621888bbfcfb87649dd88334bc9a38904778b7eead2dce93dcf0fa39dfbb.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"3\">Model</td><td colspan=\"3\">background music</td><td>voice over ΔF1</td><td colspan=\"6\">static image(s)</td><td colspan=\"3\">neither ΔF1</td></tr><tr><td colspan=\"3\">ΔF1</td><td>ΔF1</td><td colspan=\"2\">ΔF1</td><td colspan=\"2\">Sub. Acc. w/</td><td colspan=\"2\">Sub. Acc. w/o</td><td>a</td><td>v</td><td>av</td></tr><tr><td>a</td><td>v</td><td>av</td><td>a</td><td>a</td><td>v</td><td>a</td><td>v</td><td>a</td><td>v</td><td>a</td><td>v</td><td>av</td></tr><tr><td colspan=\"14\">Embedding Models</td></tr><tr><td>CAV-MAE</td><td>-0.66</td><td>-0.59</td><td>-0.65</td><td>-1.13</td><td>0.31</td><td>-0.28</td><td>12.73</td><td>19.22</td><td>11.98</td><td>19.21</td><td>-1.00</td><td>-1.27</td><td>-1.24</td></tr><tr><td>DeepAVFusion</td><td>-0.71</td><td>-0.80</td><td>-0.70</td><td>-1.25</td><td>0.28</td><td>-0.18</td><td>9.79</td><td>10.81</td><td>9.30</td><td>10.81</td><td>-1.27</td><td>-1.01</td><td>-1.36</td></tr><tr><td>Equi-AV</td><td>-0.79</td><td>-0.42</td><td>-0.41</td><td>-0.97</td><td>0.28</td><td>-0.23</td><td>11.12</td><td>10.45</td><td>10.49</td><td>10.45</td><td>-0.97</td><td>-0.97</td><td>-0.89</td></tr><tr><td>AV-Siam</td><td>-0.73</td><td>-0.72</td><td>-0.82</td><td>-1.08</td><td>0.34</td><td>-0.26</td><td>12.34</td><td>19.53</td><td>11.57</td><td>19.52</td><td>-1.01</td><td>-1.34</td><td>-1.39</td></tr><tr><td colspan=\"14\">Closed-source Foundation Models</td></tr><tr><td>Gemini 1.5 Flash</td><td>-0.22</td><td>-0.41</td><td>-0.69</td><td>2.08</td><td>-0.36</td><td>-0.25</td><td>1.66</td><td>14.35</td><td>1.68</td><td>14.39</td><td>1.40</td><td>-1.10</td><td>-1.62</td></tr><tr><td>Gemini 1.5 Pro</td><td>-0.36</td><td>-0.61</td><td>-0.94</td><td>2.48</td><td>-0.31</td><td>-0.24</td><td>2.83</td><td>20.85</td><td>2.87</td><td>20.80</td><td>1.64</td><td>-0.89</td><td>-1.09</td></tr><tr><td>Gemini 2.0 Flash</td><td>-0.09</td><td>-0.32</td><td>-0.58</td><td>0.03</td><td>0.13</td><td>-0.26</td><td>1.70</td><td>12.33</td><td>1.53</td><td>12.40</td><td>-0.03</td><td>-1.07</td><td>-1.22</td></tr><tr><td colspan=\"14\">Open-source Foundation Models</td></tr><tr><td>VideoLLaMA 2</td><td>-0.47</td><td>-0.72</td><td>-0.84</td><td>-0.54</td><td>0.27</td><td>-0.31</td><td>12.55</td><td>19.64</td><td>12.04</td><td>19.70</td><td>-0.47</td><td>-1.26</td><td>-1.33</td></tr><tr><td>Unified-IO 2</td><td>-1.20</td><td>0.18</td><td>-0.65</td><td>-0.68</td><td>0.12</td><td>-0.21</td><td>11.39</td><td>11.89</td><td>10.88</td><td>12.00</td><td>-1.07</td><td>-0.15</td><td>-1.17</td></tr><tr><td>PandaGPT</td><td>-1.13</td><td>-0.15</td><td>-0.42</td><td>0.56</td><td>-0.22</td><td>-0.16</td><td>2.94</td><td>4.27</td><td>2.91</td><td>4.24</td><td>-0.60</td><td>-0.15</td><td>-0.65</td></tr><tr><td>OLA</td><td>-2.16</td><td>0.09</td><td>-0.44</td><td>1.50</td><td>-0.48</td><td>-0.18</td><td>13.03</td><td>8.88</td><td>12.88</td><td>8.89</td><td>-0.80</td><td>-0.13</td><td>-0.06</td></tr></table>",
+        "bbox": [
+            94,
+            88,
+            906,
+            305
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/61fa65244ee5c430de75afa7adb150130023403062ea9ab4ae8143a8f510ad6b.jpg",
+        "table_caption": [
+            "Table 3. Summary of performance differences in the presence/absence of meta-classes. Difference in  $\\mathbf{F}_1$  scores ( $\\Delta \\mathbf{F}_1$ ) for audio-visual video classification on VGGSounder between videos with a meta-class and those without it. Positive numbers ( $\\Delta$ ) indicate better performance when the meta-class is present. Additional results are provided in Appendix D"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"3\">Human labels ↑</td><td colspan=\"3\">Auto labels ↑</td></tr><tr><td>A</td><td>V</td><td>AV</td><td>A</td><td>V</td><td>AV</td></tr><tr><td>Gemini 1.5 Flash</td><td>29.28</td><td>14.59</td><td>16.36</td><td>0.48</td><td>0.93</td><td>1.51</td></tr><tr><td>Gemini 1.5 Pro</td><td>28.61</td><td>25.52</td><td>27.63</td><td>0.31</td><td>1.99</td><td>2.10</td></tr><tr><td>Gemini 2.0 Flash</td><td>8.80</td><td>12.16</td><td>11.13</td><td>0.22</td><td>1.12</td><td>1.28</td></tr></table>",
+        "bbox": [
+            93,
+            372,
+            480,
+            444
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Impact of added labels using different strategies in VGGSounder. We show the change in multi-label classification performance ( $\\Delta$  Hit) when adding automatically added (Auto) or human-annotated (Human) labels to VGGSound, and compare to the original VGGSound data.",
+        "bbox": [
+            89,
+            454,
+            482,
+            523
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "models and metrics is provided in Appendix D. While performance is consistently higher with automatically added labels (Auto), the increase is noticeably smaller than that for human-curated labels. Paired with the observation that models do frequently predict correct classes that were not part of the original VGGSound label set, this indicates that human-curated labels better cover the ground truth.",
+        "bbox": [
+            89,
+            542,
+            483,
+            650
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Takeaway 8 Automatically added labels are an important step, but human-curated labels have a bigger effect on eliminating incorrectly flagged false positives, underscoring the value of accurate human annotation.",
+        "bbox": [
+            99,
+            667,
+            473,
+            729
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Discussion",
+        "text_level": 1,
+        "bbox": [
+            89,
+            750,
+            205,
+            766
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Choice of VGGSound as base dataset VGGSound is commonly used to evaluate audio-visual models on the multimodal video classification task. As it is currently the most suitable testbed for audio-visual classification tasks (due to its size, diversity of categories, non-constrained setting, and relatively strong audio-visual correspondence), it serves as an optimal starting point for our substantially improved VGGSounder benchmark with a multi-label evaluation pro",
+        "bbox": [
+            89,
+            779,
+            483,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "tocol for foundation models that makes the benchmark suitable for meaningful evaluation.",
+        "bbox": [
+            511,
+            375,
+            903,
+            405
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Multi-label vs single-label classification Video content is inherently complex, often containing multiple co-occurring objects and actions both within and across modalities. This makes it unlikely that a given clip belongs to just one class as is the case in the single-label classification task. Therefore, our VGGSounder extends the VGGSound test set to multi-label classification. Unlike models trained on a narrow single-label dataset, foundation models develop versatile representations. Multi-label evaluation thus also aligns very naturally with this capability.",
+        "bbox": [
+            511,
+            409,
+            906,
+            559
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            511,
+            573,
+            633,
+            589
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We introduced an modality-aware evaluation set for audiovisual foundation models. VGGSounder builds on the widely used VGGSound dataset by adding: (a) comprehensive human annotations for missing classes, (b) specifying modality information per label, (c) introducing specialised meta-labels for frequently occurring real-world challenges, and (d) using heuristic methods to improve label quality.",
+        "bbox": [
+            511,
+            598,
+            906,
+            704
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Through our newly introduced metric, modality confusion, we observe that incorporating additional modalities does not necessarily yield better results. Models often become more confused on a substantial subset of test samples. Furthermore, finetuned embedding models tend to rely heavily on audio cues, while foundation models depend more on visual information. Additionally, our meta-label analysis highlights distinct challenges across various specialised yet commonly occurring scenarios such as background music, static images, and voice-overs. Overall, we hope that the VGGSounder benchmark will advance the evaluation and development of foundational audio-visual models.",
+        "bbox": [
+            511,
+            705,
+            908,
+            885
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "1034",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            91,
+            90,
+            256,
+            107
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The authors would like to thank Felix Förster, Sayak Mallick, and Prasanna Mayilvahananan for their help with data annotation, Felix Förster and Monica Riedler for proofreading the paper, and Thomas Klein and Shyamgopal Karthik for their help in setting up MTurk. They also thank numerous MTurk workers for labelling.",
+        "bbox": [
+            89,
+            114,
+            480,
+            205
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work was in part supported by the BMBF (FKZ: 01IS24060, 01I524085B, 01IS18039A), the DFG (SFB 1233, project number: 276693517), and the Open Philanthropy Foundation funded by the Good Ventures Foundation. WB acknowledges financial support via an Emmy Noether Grant funded by the German Research Foundation (DFG) under grant no. BR 6382/1-1. WB is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 - Project number 390727645. This research utilised compute resources at the Tübingen Machine Learning Cloud, DFG FKZ INST 37/1057-1 FUGG. The authors thank the IMPRS-IS for supporting TW.",
+        "bbox": [
+            89,
+            205,
+            482,
+            387
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            91,
+            402,
+            187,
+            417
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016.",
+            "[2] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE TPAMI, 2018.",
+            "[3] Triantafyllos Afouras, Yuki M Asano, Francois Fagan, Andrea Vedaldi, and Florian Metze. Self-supervised object detection from audio-visual correspondence. In ECCV, 2020.",
+            "[4] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In ICASSP, 2020.",
+            "[5] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In ECCV, 2020.",
+            "[6] Humam Alwassel, Dhruv Mahajan, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, 2020.",
+            "[7] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Facetalk: Audio-driven motion diffusion for neural parametric head models. In CVPR, 2024.",
+            "[8] Relja Arandjelović and Andrew Zisserman. Look, listen and learn. In ICCV, 2017.",
+            "[9] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In ECCV, 2018.",
+            "[10] Yuki M. Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. In NeurIPS, 2020.",
+            "[11] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Sound-net: Learning sound representations from unlabeled video. In NeurIPS, 2016."
+        ],
+        "bbox": [
+            94,
+            426,
+            482,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[12] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020.",
+            "[13] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, 2020.",
+            "[14] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Audio-visual synchronisation in the wild. In BMVC, 2021.",
+            "[15] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In CVPR, 2021.",
+            "[16] Peihao Chen, Yang Zhang, Mingkui Tan, Hongdong Xiao, Deng Huang, and Chuang Gan. Generating visually aligned sound from videos. IEEE Transactions on Image Processing, 2020.",
+            "[17] Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, and Furu Wei. Beats: Audio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058, 2022.",
+            "[18] Yanbei Chen, Yongqin Xian, A. Sophia Koepke, Ying Shan, and Zeynep Akata. Distilling audio-visual knowledge by compositional contrastive learning. In CVPR, 2021.",
+            "[19] Ziyang Chen, Prem Seetharaman, Bryan Russell, Oriol Nieto, David Bourgin, Andrew Owens, and Justin Salamon. Video-guided foley sound generation with multimodal controls. In CVPR, 2025.",
+            "[20] Ying Cheng, Ruize Wang, Zhihao Pan, Rui Feng, and Yuejie Zhang. Look, listen, and attend: Co-attention network for self-supervised audio-visual representation learning. In ACM MM, 2020.",
+            "[21] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024.",
+            "[22] Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with  $90\\%$  * chatgpt quality. 2023.",
+            "[23] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In ACCV, 2016.",
+            "[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009.",
+            "[25] Joshua P Ebeneze, Yongjun Wu, Hai Wei, Sriram Sethuraman, and Zongyi Liu. Detection of audio-video synchronization errors via event detection. In ICASSP, 2021.",
+            "[26] Chuang Gan, Deng Huang, Peihao Chen, Joshua B Tenenbaum, and Antonio Torralba. Foley music: Learning to generate music from videos. In ECCV, 2020.",
+            "[27] Ruohan Gao and Kristen Grauman. Co-separating sounds of visual objects. In ICCV, 2019.",
+            "[28] Aryo Pradipta Gema, Joshua Ong Jun Leang, Giwon Hong, Alessio Devoto, Alberto Carlo Maria Mancino, Rohit Saxena, Xuanli He, Yu Zhao, Xiaotang Du, Mohammad"
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "1035",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Reza Ghasemi Madani, Claire Barale, Robert McHardy, Joshua Harris, Jean Kaddour, Emile van Krieken, and Pasquale Minervini. Are we done with mmlu? arXiv preprint arXiv:2406.04127, 2024.",
+            "[29] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In ICASSP, 2017.",
+            "[30] 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 CVPR, 2023.",
+            "[31] Shir Goldstein and Yael Moses. Guitar music transcription from silent video. In BMVC, 2018.",
+            "[32] Yuan Gong, Alexander H Liu, Andrew Rouditchenko, and James Glass. Uavm: Towards unifying audio and visual models. IEEE Signal Processing Letters, 2022.",
+            "[33] Yuan Gong, Andrew Rouditchenko, Alexander H Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James Glass. Contrastive audio-visual masked autoencoder. In ICLR, 2023.",
+            "[34] Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, et al. A survey on llm-as-a-judge. arXiv preprint arXiv:2411.15594, 2024.",
+            "[35] Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt. Aligning ai with shared human values. In ICLR, 2021.",
+            "[36] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. In ICLR, 2021.",
+            "[37] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Sparse in space and time: Audio-visual synchronisation with trainable selectors. In BMVC, 2022.",
+            "[38] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Sparse in space and time: Audio-visual synchronisation with trainable selectors. In BMVC, 2022.",
+            "[39] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Synchformer: Efficient synchronization from sparse cues. In ICASSP, 2024.",
+            "[40] Amir Jamaludin, Joon Son Chung, and Andrew Zisserman. You said that?: Synthesising talking faces from audio. IJCV, 2019.",
+            "[41] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017.",
+            "[42] Naji Khosravan, Shervin Ardeshir, and Rohit Puri. On attention modules for audio-visual synchronization. In CVPR Workshop, 2019.",
+            "[43] Jongsuk Kim, Hyeongkeun Lee, Kyeongha Rho, Junmo Kim, and Joon Son Chung. Equipav: Leveraging equivariance for audio-visual contrastive learning. In ICML, 2024.",
+            "[44] A. Sophia Koepke, Olivia Wiles, and Andrew Zisserman. Visual pitch estimation. In SMC, 2019."
+        ],
+        "bbox": [
+            91,
+            90,
+            483,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[45] A Sophia Koepke, Olivia Wiles, Yael Moses, and Andrew Zisserman. Sight to sound: An end-to-end approach for visual piano transcription. In ICASSP, 2020.",
+            "[46] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In NeurlPS, 2018.",
+            "[47] Sangho Lee, Jiwan Chung, Youngjae Yu, Gunhee Kim, Thomas Breuel, Gal Chechik, and Yale Song. Acav100m: Automatic curation of large-scale datasets for audio-visual video representation learning. In ICCV, 2021.",
+            "[48] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In CVPR, 2022.",
+            "[49] Yan-Bo Lin and Gedas Bertasius. Siamese vision transformers are scalable audio-visual learners. In ECCV, 2024.",
+            "[50] Yan-Bo Lin and Yu-Chiang Frank Wang. Audiovisual transformer with instance attention for audio-visual event localization. In ACCV, 2020.",
+            "[51] Yan-Bo Lin, Yu-Jhe Li, and Yu-Chiang Frank Wang. Dual-modality seq2seq network for audio-visual event localization. In ICASSP, 2019.",
+            "[52] Zuyan Liu, Yuhao Dong, Jiahui Wang, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Ola: Pushing the frontiers of omni-modal language model. arXiv preprint arXiv:2502.04328, 2025.",
+            "[53] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In CVPR, 2024.",
+            "[54] Jie Ma, Min Hu, Pinghui Wang, Wangchun Sun, Lingyun Song, Hongbin Pei, Jun Liu, and Youtian Du. Look, listen, and answer: Overcoming biases for audio-visual question answering. NeurIPS, 2024.",
+            "[55] 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.",
+            "[56] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In ECCV, 2022.",
+            "[57] Shentong Mo and Pedro Morgado. Unveiling the power of audio-visual early fusion transformers with dense interactions through masked modeling. In CVPR, 2024.",
+            "[58] Liliane Momeni, Triantafyllos Afouras, Themos Stafylakis, Samuel Albanie, and Andrew Zisserman. Seeing wake words: Audio-visual keyword spotting. In BMVC, 2020.",
+            "[59] Arsha Nagrani, Joon Son Chung, Samuel Albanie, and Andrew Zisserman. Disentangled speech embeddings using cross-modal self-supervision. In ICASSP, 2020.",
+            "[60] Arsha Nagrani, Shan Yang, Anurag Arnab, Aren Jansen, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In NeurIPS, 2021.",
+            "[61] Medhini Narasimhan, Shiry Ginosar, Andrew Owens, Alexei A Efros, and Trevor Darrell. Strumming to the beat: Audio-conditioned contrastive video textures. arXiv preprint arXiv:2104.02687, 2021."
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "1036",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[62] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In ECCV, 2018.",
+            "[63] Andrew Owens, Jiajun Wu, Josh H McDermott, William T Freeman, and Antonio Torralba. Ambient sound provides supervision for visual learning. In ECCV, 2016.",
+            "[64] Andrew Owens, Jiajun Wu, Josh H McDermott, William T Freeman, and Antonio Torralba. Learning sight from sound: Ambient sound provides supervision for visual learning. IJCV, 2018.",
+            "[65] Mandela Patrick, Yuki M Asano, Ruth Fong, João F Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multi-modal self-supervision from generalized data transformations. In NeurIPS, 2020.",
+            "[66] KR Prajwal, Liliane Momeni, Triantafyllos Afouras, and Andrew Zisserman. Visual keyword spotting with attention. In BMVC, 2021.",
+            "[67] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In ECCV, 2020.",
+            "[68] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICLR, 2023.",
+            "[69] Kun Su, Xiulong Liu, and Eli Shlizerman. Multi-instrumentalist net: Unsupervised generation of music from body movements. arXiv preprint arXiv:2012.03478, 2020.",
+            "[70] Kun Su, Xiulong Liu, and Eli Shlizerman. How does it sound? In NeurIPS, 2021.",
+            "[71] Kun Su, Xiulong Liu, and Eli Shlizerman. From vision to audio and beyond: A unified model for audio-visual representation and generation. In ICML, 2024.",
+            "[72] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023.",
+            "[73] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023.",
+            "[74] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, 2018.",
+            "[75] Efthymios Tzinis, Scott Wisdom, Aren Jansen, Shawn Hershey, Tal Remez, Daniel PW Ellis, and John R Hershey. Into the wild with audioscope: Unsupervised audio-visual separation of on-screen sounds. In ICLR, 2021.",
+            "[76] Ilpo Viertola, Vladimir Iashin, and Esa Rahtu. Temporally aligned audio for video with autoregression. In ICASSP, 2025.",
+            "[77] Olivia Wiles, A. Sophia Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In ECCV, 2018.",
+            "[78] Yu Wu, Linchao Zhu, Yan Yan, and Yi Yang. Dual attention matching for audio-visual event localization. In CVPR, 2019.",
+            "[79] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020."
+        ],
+        "bbox": [
+            91,
+            92,
+            480,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[80] Haoming Xu, Runhao Zeng, Qingyao Wu, Mingkui Tan, and Chuang Gan. Cross-modal relation-aware networks for audio-visual event localization. In ACM MM, 2020.",
+            "[81] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In ECCV, 2024.",
+            "[82] An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025.",
+            "[83] Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. Avqa: A dataset for audio-visual question answering on videos. In ACM MM, 2022.",
+            "[84] Heeseung Yun, Youngjae Yu, Wonsuk Yang, Kangil Lee, and Gunhee Kim. Pano-avqa: Grounded audio-visual question answering on 360deg videos. In ICCV, 2021.",
+            "[85] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In ECCV, 2018.",
+            "[86] Hang Zhao, Chuang Gan, Wei-Chiu Ma, and Antonio Torralba. The sound of motions. In ICCV, 2019.",
+            "[87] Hang Zhou, Ziwei Liu, Xudong Xu, Ping Luo, and Xiaogang Wang. Vision-infused deep audio inpainting. In ICCV, 2019.",
+            "[88] Yipin Zhou, Zhaowen Wang, Chen Fang, Trung Bui, and Tamara L. Berg. Visual to sound: Generating natural sound for videos in the wild. In CVPR, 2018.",
+            "[89] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023.",
+            "[90] Lingyu Zhu and Esa Rahtu. V-slowfast network for efficient visual sound separation. In WACV, 2022."
+        ],
+        "bbox": [
+            516,
+            92,
+            903,
+            561
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "1037",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 10
+    }
+]