Add Batch f778b63e-db2a-4ae0-a4d9-ac063e465670 data

.gitattributes

@@ -1897,3 +1897,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/Metadata[[:space:]]Conditioning[[:space:]]Accelerates[[:space:]]Language[[:space:]]Model[[:space:]]Pre-training/b2db629e-ea1c-49a3-b780-6786bf3888f1_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/Metastable[[:space:]]Dynamics[[:space:]]of[[:space:]]Chain-of-Thought[[:space:]]Reasoning_[[:space:]]Provable[[:space:]]Benefits[[:space:]]of[[:space:]]Search,[[:space:]]RL[[:space:]]and[[:space:]]Distillation/d5c547eb-4d52-4dd6-a1cf-27b15a57d0e7_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/MetricEmbedding_[[:space:]]Accelerate[[:space:]]Metric[[:space:]]Nearness[[:space:]]by[[:space:]]Tropical[[:space:]]Inner[[:space:]]Product/f5fbaa3b-db04-4f3d-98aa-8b0ce5e18b35_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MimicMotion_[[:space:]]High-Quality[[:space:]]Human[[:space:]]Motion[[:space:]]Video[[:space:]]Generation[[:space:]]with[[:space:]]Confidence-aware[[:space:]]Pose[[:space:]]Guidance/06747124-f5c1-49e9-89da-3b2d84cc796b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mind[[:space:]]Your[[:space:]]Step[[:space:]](by[[:space:]]Step)_[[:space:]]Chain-of-Thought[[:space:]]can[[:space:]]Reduce[[:space:]]Performance[[:space:]]on[[:space:]]Tasks[[:space:]]where[[:space:]]Thinking[[:space:]]Makes[[:space:]]Humans[[:space:]]Worse/c5ac481f-06ce-428c-a6a1-af450147d7f0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mind[[:space:]]the[[:space:]]Gap_[[:space:]]A[[:space:]]Practical[[:space:]]Attack[[:space:]]on[[:space:]]GGUF[[:space:]]Quantization/36c60854-11ea-47c6-bfdb-b6ef74c79b0d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mind[[:space:]]the[[:space:]]Gap_[[:space:]]a[[:space:]]Spectral[[:space:]]Analysis[[:space:]]of[[:space:]]Rank[[:space:]]Collapse[[:space:]]and[[:space:]]Signal[[:space:]]Propagation[[:space:]]in[[:space:]]Attention[[:space:]]Layers/bc91cb4b-3713-4802-95f4-a813a601e7f4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MindAligner_[[:space:]]Explicit[[:space:]]Brain[[:space:]]Functional[[:space:]]Alignment[[:space:]]for[[:space:]]Cross-Subject[[:space:]]Visual[[:space:]]Decoding[[:space:]]from[[:space:]]Limited[[:space:]]fMRI[[:space:]]Data/9c365be7-b2b5-41f9-a434-a3131c723e3b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MindCustomer_[[:space:]]Multi-Context[[:space:]]Image[[:space:]]Generation[[:space:]]Blended[[:space:]]with[[:space:]]Brain[[:space:]]Signal/8c335697-32bf-4d12-8441-07e837ba6e05_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MindLLM_[[:space:]]A[[:space:]]Subject-Agnostic[[:space:]]and[[:space:]]Versatile[[:space:]]Model[[:space:]]for[[:space:]]fMRI-to-text[[:space:]]Decoding/33b83fb9-ebc9-4760-81ee-af53162ee500_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Minerva_[[:space:]]A[[:space:]]Programmable[[:space:]]Memory[[:space:]]Test[[:space:]]Benchmark[[:space:]]for[[:space:]]Language[[:space:]]Models/08a6cf6a-c318-43be-96ff-f6dc835f785d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Minimalist[[:space:]]Concept[[:space:]]Erasure[[:space:]]in[[:space:]]Generative[[:space:]]Models/2eb5f228-69c7-40cb-b671-1edd595e1a77_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Minimax[[:space:]]Optimal[[:space:]]Regret[[:space:]]Bound[[:space:]]for[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]with[[:space:]]Trajectory[[:space:]]Feedback/143faa98-cad6-4d13-a5d0-563d6e4024eb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Minimum[[:space:]]Width[[:space:]]for[[:space:]]Universal[[:space:]]Approximation[[:space:]]using[[:space:]]Squashable[[:space:]]Activation[[:space:]]Functions/b9dc0d8d-1633-4637-8699-e7e07e789115_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MiraGe_[[:space:]]Editable[[:space:]]2D[[:space:]]Images[[:space:]]using[[:space:]]Gaussian[[:space:]]Splatting/081c41d5-d6c3-417a-922f-257edd076506_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mirror,[[:space:]]Mirror[[:space:]]of[[:space:]]the[[:space:]]Flow_[[:space:]]How[[:space:]]Does[[:space:]]Regularization[[:space:]]Shape[[:space:]]Implicit[[:space:]]Bias_/90fdfb7a-1827-498d-8d6e-e141ea883fe0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MissScore_[[:space:]]High-Order[[:space:]]Score[[:space:]]Estimation[[:space:]]in[[:space:]]the[[:space:]]Presence[[:space:]]of[[:space:]]Missing[[:space:]]Data/b9c23e9f-35db-496b-8f7c-1725efcd4176_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mitigating[[:space:]]Heterogeneous[[:space:]]Token[[:space:]]Overfitting[[:space:]]in[[:space:]]LLM[[:space:]]Knowledge[[:space:]]Editing/b4ea99ae-ee5b-4ceb-865f-b11b70431c7d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mitigating[[:space:]]Local[[:space:]]Cohesion[[:space:]]and[[:space:]]Global[[:space:]]Sparseness[[:space:]]in[[:space:]]Graph[[:space:]]Contrastive[[:space:]]Learning[[:space:]]with[[:space:]]Fuzzy[[:space:]]Boundaries/18620a7a-f10f-4373-ba11-4b483c22c6f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mitigating[[:space:]]Object[[:space:]]Hallucination[[:space:]]in[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models[[:space:]]via[[:space:]]Image-Grounded[[:space:]]Guidance/069b2c9f-41df-4e20-9c29-42a6dbc8c2a2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mitigating[[:space:]]Over-Squashing[[:space:]]in[[:space:]]Graph[[:space:]]Neural[[:space:]]Networks[[:space:]]by[[:space:]]Spectrum-Preserving[[:space:]]Sparsification/64b68372-9f1e-4bde-9760-a81fd4bf3156_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mitigating[[:space:]]Plasticity[[:space:]]Loss[[:space:]]in[[:space:]]Continual[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]by[[:space:]]Reducing[[:space:]]Churn/4bf05448-a2d8-4d8e-8590-8e99990f2dde_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MixBridge_[[:space:]]Heterogeneous[[:space:]]Image-to-Image[[:space:]]Backdoor[[:space:]]Attack[[:space:]]through[[:space:]]Mixture[[:space:]]of[[:space:]]Schrödinger[[:space:]]Bridges/7b099411-4558-49fd-8c51-f9cc4cd42812_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MixMin_[[:space:]]Finding[[:space:]]Data[[:space:]]Mixtures[[:space:]]via[[:space:]]Convex[[:space:]]Minimization/a10ed349-86ca-48fb-802c-84e65c0adb10_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mixed-curvature[[:space:]]decision[[:space:]]trees[[:space:]]and[[:space:]]random[[:space:]]forests/0e4e8503-a99a-4465-809e-7b1493adf75b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mixture[[:space:]]of[[:space:]]Experts[[:space:]]Made[[:space:]]Intrinsically[[:space:]]Interpretable/92b2cca5-129f-4941-82b5-9921baef2aef_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mixture[[:space:]]of[[:space:]]Experts[[:space:]]Provably[[:space:]]Detect[[:space:]]and[[:space:]]Learn[[:space:]]the[[:space:]]Latent[[:space:]]Cluster[[:space:]]Structure[[:space:]]in[[:space:]]Gradient-Based[[:space:]]Learning/0a278ebe-0a0a-4bcd-848b-deddacbdc7aa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mixture[[:space:]]of[[:space:]]Hidden-Dimensions_[[:space:]]Not[[:space:]]All[[:space:]]Hidden-States’[[:space:]]Dimensions[[:space:]]are[[:space:]]Needed[[:space:]]in[[:space:]]Transformer/b37b4f1a-6e19-4c45-b0f2-aecbb866bc72_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mixture[[:space:]]of[[:space:]]Lookup[[:space:]]Experts/760625b4-2ccd-4fd5-9b25-84dafe223a8a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MoE-SVD_[[:space:]]Structured[[:space:]]Mixture-of-Experts[[:space:]]LLMs[[:space:]]Compression[[:space:]]via[[:space:]]Singular[[:space:]]Value[[:space:]]Decomposition/c9fa16f2-8053-4f66-b95f-072a0cc5b793_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MoEQuant_[[:space:]]Enhancing[[:space:]]Quantization[[:space:]]for[[:space:]]Mixture-of-Experts[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Expert-Balanced[[:space:]]Sampling[[:space:]]and[[:space:]]Affinity[[:space:]]Guidance/65f0d822-c679-4174-a1ba-1d0bb76ffdbd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MoHAVE_[[:space:]]Mixture[[:space:]]of[[:space:]]Hierarchical[[:space:]]Audio-Visual[[:space:]]Experts[[:space:]]for[[:space:]]Robust[[:space:]]Speech[[:space:]]Recognition/bce7e8a1-4c68-4b66-a85f-79bff47faf0b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MoH_[[:space:]]Multi-Head[[:space:]]Attention[[:space:]]as[[:space:]]Mixture-of-Head[[:space:]]Attention/53226982-3e41-4335-93cb-fc619d36cbed_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MoMa_[[:space:]]Modulating[[:space:]]Mamba[[:space:]]for[[:space:]]Adapting[[:space:]]Image[[:space:]]Foundation[[:space:]]Models[[:space:]]to[[:space:]]Video[[:space:]]Recognition/8694306e-6046-4e3f-9a7c-116098086103_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MoRAgent_[[:space:]]Parameter[[:space:]]Efficient[[:space:]]Agent[[:space:]]Tuning[[:space:]]with[[:space:]]Mixture-of-Roles/c431dadd-3481-4754-8674-b22286bd49da_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Modalities[[:space:]]Contribute[[:space:]]Unequally_[[:space:]]Enhancing[[:space:]]Medical[[:space:]]Multi-modal[[:space:]]Learning[[:space:]]through[[:space:]]Adaptive[[:space:]]Modality[[:space:]]Token[[:space:]]Re-balancing/e84ab24d-46ea-4c09-8273-e489728277fa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Model[[:space:]]Immunization[[:space:]]from[[:space:]]a[[:space:]]Condition[[:space:]]Number[[:space:]]Perspective/604be868-b4fb-4a43-8435-4d778be209cb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Model[[:space:]]Steering_[[:space:]]Learning[[:space:]]with[[:space:]]a[[:space:]]Reference[[:space:]]Model[[:space:]]Improves[[:space:]]Generalization[[:space:]]Bounds[[:space:]]and[[:space:]]Scaling[[:space:]]Laws/029808c0-200a-4264-b554-214d6b0f524f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Model[[:space:]]Swarms_[[:space:]]Collaborative[[:space:]]Search[[:space:]]to[[:space:]]Adapt[[:space:]]LLM[[:space:]]Experts[[:space:]]via[[:space:]]Swarm[[:space:]]Intelligence/5d54bed7-bd1d-49d4-952c-c852f8104f17_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Model[[:space:]]Uncertainty[[:space:]]Quantification[[:space:]]by[[:space:]]Conformal[[:space:]]Prediction[[:space:]]in[[:space:]]Continual[[:space:]]Learning/a95a6766-90a8-4027-bcd2-8ca794efb041_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Model-Based[[:space:]]Exploration[[:space:]]in[[:space:]]Monitored[[:space:]]Markov[[:space:]]Decision[[:space:]]Processes/fa8e4ba2-2b89-4d63-8bb8-7be1a66a164b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Modeling[[:space:]]All-Atom[[:space:]]Glycan[[:space:]]Structures[[:space:]]via[[:space:]]Hierarchical[[:space:]]Message[[:space:]]Passing[[:space:]]and[[:space:]]Multi-Scale[[:space:]]Pre-training/63e0cf4f-f48a-41bd-9910-3dcff8aa82f6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Modeling[[:space:]]Multi-Task[[:space:]]Model[[:space:]]Merging[[:space:]]as[[:space:]]Adaptive[[:space:]]Projective[[:space:]]Gradient[[:space:]]Descent/3c09391e-27b2-432f-b9f1-20df3a4683fd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Models[[:space:]]of[[:space:]]Heavy-Tailed[[:space:]]Mechanistic[[:space:]]Universality/d4c9ea13-f8bf-426a-94b9-83446dc2693a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Modified[[:space:]]K-means[[:space:]]Algorithm[[:space:]]with[[:space:]]Local[[:space:]]Optimality[[:space:]]Guarantees/26784730-ffc8-4c60-a66e-8d248040726c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Modular[[:space:]]Duality[[:space:]]in[[:space:]]Deep[[:space:]]Learning/28ce37bc-fdf6-4f6d-ad01-61a916447083_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Modularized[[:space:]]Self-Reflected[[:space:]]Video[[:space:]]Reasoner[[:space:]]for[[:space:]]Multimodal[[:space:]]LLM[[:space:]]with[[:space:]]Application[[:space:]]to[[:space:]]Video[[:space:]]Question[[:space:]]Answering/7159a8ae-8f77-4ba8-bb09-0261b164ea71_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Modulated[[:space:]]Diffusion_[[:space:]]Accelerating[[:space:]]Generative[[:space:]]Modeling[[:space:]]with[[:space:]]Modulated[[:space:]]Quantization/f6c47750-9e9d-4ddb-8c66-ad99c9c6740c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Moirai-MoE_[[:space:]]Empowering[[:space:]]Time[[:space:]]Series[[:space:]]Foundation[[:space:]]Models[[:space:]]with[[:space:]]Sparse[[:space:]]Mixture[[:space:]]of[[:space:]]Experts/773b34ea-1d3d-4042-a724-c97ed8abf464_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Momentum-Driven[[:space:]]Adaptivity_[[:space:]]Towards[[:space:]]Tuning-Free[[:space:]]Asynchronous[[:space:]]Federated[[:space:]]Learning/5c3e3e27-95b6-4bc8-a925-2cc0c6b6f597_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Monte[[:space:]]Carlo[[:space:]]Tree[[:space:]]Diffusion[[:space:]]for[[:space:]]System[[:space:]]2[[:space:]]Planning/407de874-64ae-4528-af76-a3ed7e8b1e26_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Monte[[:space:]]Carlo[[:space:]]Tree[[:space:]]Search[[:space:]]for[[:space:]]Comprehensive[[:space:]]Exploration[[:space:]]in[[:space:]]LLM-Based[[:space:]]Automatic[[:space:]]Heuristic[[:space:]]Design/9ada0bc9-cb03-459a-a253-81205990ca7a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Monte-Carlo[[:space:]]Tree[[:space:]]Search[[:space:]]with[[:space:]]Uncertainty[[:space:]]Propagation[[:space:]]via[[:space:]]Optimal[[:space:]]Transport/00da5638-7f60-4b2d-92d3-cece48343ce5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/More[[:space:]]Than[[:space:]]Meets[[:space:]]the[[:space:]]Eye_[[:space:]]Enhancing[[:space:]]Multi-Object[[:space:]]Tracking[[:space:]]Even[[:space:]]with[[:space:]]Prolonged[[:space:]]Occlusions/39f39c4d-a984-436f-9c3a-4d43a6fea7e2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Morse_[[:space:]]Dual-Sampling[[:space:]]for[[:space:]]Lossless[[:space:]]Acceleration[[:space:]]of[[:space:]]Diffusion[[:space:]]Models/7592be98-d212-4319-9f5a-75b783ffa30a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MuLan_[[:space:]]Adapting[[:space:]]Multilingual[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]Hundreds[[:space:]]of[[:space:]]Languages[[:space:]]with[[:space:]]Negligible[[:space:]]Cost/f737471c-862b-4efd-aaf7-d4c10db6f04b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Multi-Armed[[:space:]]Bandits[[:space:]]with[[:space:]]Interference_[[:space:]]Bridging[[:space:]]Causal[[:space:]]Inference[[:space:]]and[[:space:]]Adversarial[[:space:]]Bandits/0d1a227d-c483-4fa0-827c-38b6335416c8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Multi-Domain[[:space:]]Graph[[:space:]]Foundation[[:space:]]Models_[[:space:]]Robust[[:space:]]Knowledge[[:space:]]Transfer[[:space:]]via[[:space:]]Topology[[:space:]]Alignment/1f34d310-3fd1-466b-8be0-09e398927b4e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Multi-Marginal[[:space:]]Stochastic[[:space:]]Flow[[:space:]]Matching[[:space:]]for[[:space:]]High-Dimensional[[:space:]]Snapshot[[:space:]]Data[[:space:]]at[[:space:]]Irregular[[:space:]]Time[[:space:]]Points/28ed61f0-1ac1-4a69-84de-64b278bd0315_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Multi-Modal[[:space:]]Object[[:space:]]Re-identification[[:space:]]via[[:space:]]Sparse[[:space:]]Mixture-of-Experts/d9830710-9d1a-4aa6-9ab9-3c6953202def_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Multi-Objective[[:space:]]Causal[[:space:]]Bayesian[[:space:]]Optimization/e26fdc80-39dd-44b4-b12a-e2f455eb17f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Multi-Session[[:space:]]Budget[[:space:]]Optimization[[:space:]]for[[:space:]]Forward[[:space:]]Auction-based[[:space:]]Federated[[:space:]]Learning/291c7e27-1578-4eb1-86c3-f6e26c8fb8a7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Multi-Stage[[:space:]]Manipulation[[:space:]]with[[:space:]]Demonstration-Augmented[[:space:]]Reward,[[:space:]]Policy,[[:space:]]and[[:space:]]World[[:space:]]Model[[:space:]]Learning/166fb61b-f2c5-475b-b9ec-f31549364770_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Multi-Timescale[[:space:]]Dynamics[[:space:]]Model[[:space:]]Bayesian[[:space:]]Optimization[[:space:]]for[[:space:]]Plasma[[:space:]]Stabilization[[:space:]]in[[:space:]]Tokamaks/16f84b04-0b3c-4260-8c2c-195d4fed6c63_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Multi-Turn[[:space:]]Code[[:space:]]Generation[[:space:]]Through[[:space:]]Single-Step[[:space:]]Rewards/2a064cfb-d3b5-492f-8f90-b687f89dad03_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Multi-View[[:space:]]Graph[[:space:]]Clustering[[:space:]]via[[:space:]]Node-Guided[[:space:]]Contrastive[[:space:]]Encoding/50d310dc-c958-46c2-9536-788ea61cc810_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Multi-agent[[:space:]]Architecture[[:space:]]Search[[:space:]]via[[:space:]]Agentic[[:space:]]Supernet/25a746aa-2e14-4571-b770-111e7a05e017_origin.pdf filter=lfs diff=lfs merge=lfs -text

2025/MimicMotion_ High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance/06747124-f5c1-49e9-89da-3b2d84cc796b_content_list.json

@@ -0,0 +1,2088 @@
+[
+    {
+        "type": "text",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "text_level": 1,
+        "bbox": [
+            143,
+            109,
+            828,
+            151
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yuang Zhang  $^{12}$  Jiaxi Gu  $^{1}$  Li-Wen Wang  $^{1}$  Han Wang  $^{12}$  Junqi Cheng  $^{1}$  Yuefeng Zhu  $^{1}$  Fangyuan Zou  $^{1}$",
+        "bbox": [
+            109,
+            195,
+            859,
+            212
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/19b946f1c7307a0afd877f0e6b2a0148ff7cad305b0d1d9ff9c2403a5064aaeb.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            253,
+            274,
+            311
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/59b037784c53cb0f54365680bd263d90d07472be01e047d6d79cca6983436ee7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            285,
+            253,
+            470,
+            311
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/4109f4a128db299605945f91f88bf2950f52128c1b848cd1a1534f5a59bf48aa.jpg",
+        "image_caption": [
+            "Figure 1: Pose-guided dancing and talking videos generated by MimicMotion showcase its capability to produce diverse human motions and long videos."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            89,
+            315,
+            470,
+            435
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            241,
+            498,
+            318,
+            513
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In recent years, while generative AI has advanced significantly in image generation, video generation continues to face challenges in controllability, length, and detail quality, which hinder its application. We present MimicMotion, a framework for generating high-quality human videos of arbitrary length using motion guidance. Our approach has several highlights. Firstly, we introduce confidence-aware pose guidance that ensures high frame quality and temporal smoothness. Secondly, we introduce regional loss amplification based on pose confidence, which reduces image distortion in key regions. Lastly, we propose a progressive latent fusion strategy to generate long and smooth videos. Experiments demonstrate the effectiveness of our approach in producing high-quality human motion videos. Videos and comparisons are available at https://tencent.github.io/MimicMotion.",
+        "bbox": [
+            116,
+            523,
+            444,
+            810
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Tencent  $^{2}$ Shanghai Jiao Tong University. Correspondence to: Jiaxi Gu <imjiaxi@gmail.com>.",
+        "bbox": [
+            84,
+            827,
+            475,
+            854
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Proceedings of the  $42^{nd}$  International Conference on Machine Learning, Vancouver, Canada. PMLR 267, 2025. Copyright 2025 by the author(s).",
+        "bbox": [
+            84,
+            864,
+            473,
+            905
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            498,
+            239,
+            629,
+            255
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Video generation has advanced alongside image generation. However, video generation is more challenging due to its higher inherent complexities, including the need for high-quality imagery and seamless temporal smoothness. In addition, controlling the generated content and extension to significant lengths is essential for real-world use. In this paper, we focus on pose-guided video generation with a reference image. Our goal is to generate a video that contains rich imagery details and adheres to the reference image and the pose guidance.",
+        "bbox": [
+            495,
+            265,
+            887,
+            417
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Currently, there are plenty of works focusing on this task, such as DisCo (Wang et al., 2023),AnimateAnyone (Hu, 2024), MagicDance (Chang et al., 2023), MagicAnimate (Xu et al., 2024), etc. Though various techniques have been studied, the results are unsatisfactory in several aspects. Imagery distortion especially on the regions of human hands is still a common issue which is particularly evident in videos containing large movements. Besides, to achieve good temporal smoothness, imagery details are sometimes sacrificed resulting in videos of blurred frames. In the presence of diverse appearances and motions in videos, accurate pose estimation is inherently challenging. This inaccuracy not only creates a conflict between pose alignment and temporal smoothness but also hinders the model scaling on the training schedule due to overfitting on noisy samples. In addition, due to computational limitations and model capabilities, there are still significant challenges in generating high-quality long videos containing a large number of frames. To solve these problems, we propose a series of approaches for generating long but still smooth videos based on pose guidance and image reference.",
+        "bbox": [
+            495,
+            422,
+            888,
+            741
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To alleviate the negative impact of inaccurate pose estimation, we propose an approach of confidence-aware pose guidance. By introducing the concept of confidence to the pose sequence representation, better temporal smoothness can be achieved and imagery distortion can also be eased. Confidence-based regional loss amplification can make the hand regions more accurate and clear. In addition, we propose a progressive latent fusion method for achieving long but still smooth video generation. Through generating video segments with overlapped frames with the proposed pro",
+        "bbox": [
+            495,
+            748,
+            888,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "1",
+        "bbox": [
+            480,
+            922,
+            491,
+            934
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "gressive latent fusion, our model can handle arbitrary-length pose sequence guidance. By merging the generated video segments, the final long video can be of good cross-frame smoothness and imagery richness at the same time (see Figure 1). For model training, to keep the cost of model training within an acceptable range, our method is based on a generally pre-trained video generation model. The amount of training data is not large and no special manual annotation is required.",
+        "bbox": [
+            84,
+            84,
+            475,
+            220
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In summary, there are three key contributions of this work:",
+        "bbox": [
+            84,
+            227,
+            473,
+            242
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. We improve the pose guidance by employing a confidence-aware strategy. In this way, we alleviate the negative impact of inaccurate pose estimation. This approach not only reduces the influence of noisy samples during training but also corrects erroneous pose guidance during inference.",
+            "2. Based on the confidence-aware strategy, we propose hand region enhancement to alleviate hand distortion by strengthening the loss weight of the region of human hands with high pose confidence.",
+            "3. While overlapped diffusion is a standard technique for generating long videos, we advance it with position-aware progressive latent fusion that improves temporal smoothness at segment boundaries. Extensive experiments show the effectiveness of the proposed approach."
+        ],
+        "bbox": [
+            98,
+            258,
+            475,
+            503
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            84,
+            523,
+            227,
+            539
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Diffusion Models for Image/Video Generation",
+        "text_level": 1,
+        "bbox": [
+            84,
+            549,
+            437,
+            565
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Diffusion-based models have demonstrated promising results in the fields of image (Ho et al., 2020; Song et al., 2020; Rombach et al., 2022a; Zhang et al., 2023; Wang et al., 2024a) and video generation (Singer et al., 2022; Guo et al., 2023; Ho et al., 2022; He et al., 2022; Chen et al., 2023; Blattmann et al., 2023), renowned for their capacity in generative tasks. Diffusion models operating in the pixel domain encounter challenges in generating high-resolution images due to information redundancy and high computational costs. Latent Diffusion Models (LDM) (Rombach et al., 2022a) address these issues by performing the diffusion process in low-dimensional latent spaces, significantly enhancing generation efficiency and quality while reducing computational demands. Recently, efforts have been made to improve training efficiency through noise schedule (Zheng et al., 2024a;b). Compared to image generation, video generation demands a more precise understanding of temporal motion patterns. Recent video generation models leverage diffusion models by adding temporal layers to pre-trained image generation models (Singer et al., 2022; Guo et al., 2023; Wu et al., 2023; Wang et al., 2024c), or utilizing DiT (Peebles & Xie, 2023) structures to enhance",
+        "bbox": [
+            84,
+            573,
+            475,
+            905
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "generative capabilities for videos (Yan et al., 2021; Yu et al., 2023; Ma et al., 2024a; Bao et al., 2024). Stable Video Diffusion (SVD) (Blattmann et al., 2023) is one of the most popular open-source models built upon LDM. It offers a straightforward and effective method for image-based video generation and serves as a powerful pre-trained model for this task. Our approach extends SVD for pose-guided video generation, leveraging its pre-trained generative capabilities.",
+        "bbox": [
+            496,
+            84,
+            887,
+            205
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Pose-Guided Human Motion Transfer",
+        "text_level": 1,
+        "bbox": [
+            496,
+            222,
+            795,
+            236
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Pose-to-appearance mapping aims to transfer motion from the source identity to the target identity. Methods based on paired keypoints from source and target images employ local affine transformations (Siarohin et al., 2019; 2021) or Thin-Plate Spline transformations (Zhao & Zhang, 2022) to warp the source image to match the driving image. These techniques aim to minimize distortion by applying weighted affine transformations, thereby generating poses in the output image that closely resemble those in the driving image. Similarly, methods such as (Chan et al., 2019; Tu et al., 2024; Ma et al., 2024b; Hu, 2024; Feng et al., 2023; Peng et al., 2024; Li et al., 2024) utilize pose stick figures obtained from off-the-shelf human pose detectors as motion indicators and directly generate video frames through generative models. Depth information (Feng et al., 2023) or 3D human parametric models, such as SMPL (Skinned Multi-Person Linear) (Zhu et al., 2024; Wang et al., 2024b), can also be used to represent human geometry and motion characteristics from the source video. Nevertheless, these overly dense guidance techniques can rely too much on the signal from the source video, such as the outline of the body, leading to a degradation in the quality of the generated videos, especially when the target identity differs significantly from the source. Our approach, leveraging off-the-shelf human pose detectors, is capable of capturing the motion of the human body in driving videos without introducing excessive extraneous information, thereby ensuring the overall quality of the generated video. In addition, we introduce confidence-aware pose guidance, which effectively mitigates the influence of inaccurate pose estimation in training and inference. In this way, we achieve superior portrait frame quality, especially in the hand regions.",
+        "bbox": [
+            495,
+            244,
+            888,
+            731
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3. Long video generation",
+        "text_level": 1,
+        "bbox": [
+            496,
+            744,
+            687,
+            760
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recent diffusion-based video generation algorithms are constrained to producing videos with durations of only a few seconds, significantly limiting their practical applications. As a result, substantial research efforts have been dedicated to extending the duration of generated videos, leading to the proposal of various approaches to overcome this limitation. Methods like (He et al., 2022; Voleti et al., 2022) autoregressively predict successive frames, enabling the generation of infinitely long videos. However, these methods",
+        "bbox": [
+            495,
+            768,
+            888,
+            905
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "header",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "bbox": [
+            165,
+            56,
+            807,
+            70
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "2",
+        "bbox": [
+            480,
+            922,
+            491,
+            934
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "often face quality degradation due to error accumulation and the lack of long-term temporal coherence. Hierarchical approach (Yin et al., 2023; Wang et al., 2024c) are proposed for generating long videos in a coarse-to-fine manner. It first creates a coarse storyline with keyframes using a global diffusion model, then iteratively refines the video with local diffusion models to produce intermediate frames.",
+        "bbox": [
+            84,
+            85,
+            473,
+            191
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "MultiDiffusion (Bar-Tal et al., 2023) combines multiple processes that use pre-trained text-to-image diffusion models to create high-quality images with user-defined controls. It works by applying the model to different parts of an image and using an optimization method to ensure all parts blend seamlessly. This allows users to generate images that meet specific requirements, like certain aspect ratios or spatial layouts, without needing additional training or fine-tuning. Lumiere (Bar-Tal et al., 2024) extends MultiDiffusion to video generation by dividing the video into overlapping temporal segments. Each segment is independently denoised, and an optimization algorithm then combines these denoised segments. This approach ensures high coherence in the generated video, effectively maintaining temporal smoothness across segments. However, our experiments reveal that abrupt transitions can still occur at segment boundaries.",
+        "bbox": [
+            86,
+            198,
+            475,
+            441
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Building upon this principle, we introduce a position-aware progressive latent fusion strategy that enhances temporal smoothness near segment boundaries. We adaptively assign fusion weight based on the temporal position, ensuring a smooth transition at the segment boundaries.",
+        "bbox": [
+            84,
+            446,
+            473,
+            523
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            84,
+            542,
+            174,
+            556
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            84,
+            568,
+            212,
+            580
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "A Diffusion Model (DM) learns a diffusion process that generates a probability distribution for a given dataset. In the case of visual content generation tasks, a neural network of DM is trained to reverse the process of adding noise to real data so new data can be progressively generated starting from random noise. For a data sample  $\\mathbf{x} \\sim p_{\\mathrm{data}}$  from a specific data distribution  $p_{\\mathrm{data}}$ , the forward diffusion process is defined as a fixed Markov Chain that gradually adds Gaussian noise to the data, following:  $q(\\mathbf{x_t} \\mid \\mathbf{x}_{t-1}) = \\mathcal{N}(\\mathbf{x}_t; \\sqrt{1 - \\beta_t}\\mathbf{x}_{t-1}, \\beta_t\\mathbf{I})$  for  $t = 1, \\dots, T$ , where  $T$  is the number of perturbing steps and  $\\mathbf{x}_t$  represents noisy data after adding  $t$  steps of noise on the real data  $\\mathbf{x}_0$ . This process is controlled by a schedule  $\\beta_t$ . The above diffusion process can be reformed as:",
+        "bbox": [
+            84,
+            590,
+            475,
+            801
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nq \\left(\\mathbf {x} _ {\\mathbf {t}} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            145,
+            811,
+            473,
+            829
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\bar{\\alpha}_{t} = \\prod_{i=1}^{t} \\alpha_{i}$  and  $\\alpha_{t} = 1 - \\beta_{t}$ . Following DDPM (Ho et al., 2020), a denoising function  $\\epsilon_{\\theta}$  with parameter  $\\theta$  is trained by minimizing the mean squared error:",
+        "bbox": [
+            84,
+            835,
+            475,
+            882
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), \\mathbf {x} _ {t}, \\mathbf {c}, t} [ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}; \\mathbf {c}, t) \\| _ {2} ^ {2} ] \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            158,
+            888,
+            473,
+            907
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/a761ee732ae6e82883fed18dd696e3c412dfe71c87cf190a79148f44957ef285.jpg",
+        "image_caption": [
+            "Figure 2: MimicMotion integrates an image-to-video diffusion model with confidence-aware pose guidance, utilizing a spatiotemporal U-Net and a PoseNet for introducing pose sequence as the condition. Our design features adaptive adjustment of pose influence via keypoint confidence scores and prioritizes high-confidence hand regions in the loss function."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            498,
+            82,
+            885,
+            247
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{c}$  is the condition. We then train  $\\epsilon_{\\theta}$  until convergence.",
+        "bbox": [
+            496,
+            395,
+            887,
+            410
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Data Preparation",
+        "text_level": 1,
+        "bbox": [
+            496,
+            426,
+            653,
+            441
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To train a pose-guided video diffusion model, we collect a video dataset containing various human motions. The dataset need not be excessively large, as the pre-trained model already has a good prior.",
+        "bbox": [
+            496,
+            450,
+            885,
+            511
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given a video from our dataset, the training sample is constructed with three parts: a reference image (denoted as  $I_{\\mathrm{ref}}$ ), a sequence of raw video frames, and the corresponding poses. Basic pre-processing operations of resizing and cropping are applied to the raw video to get a sequence of video frames. For a given video, a fixed number of frames are randomly sampled at equal intervals as input video frames to the diffusion model. The input reference image is independently sampled from the video and pre-processed in the same way as the video frames. Another input of the model is the pose sequence, which is extracted from the video frames with DWPose (Yang et al., 2023) frame by frame.",
+        "bbox": [
+            496,
+            518,
+            885,
+            700
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3. Pose-Guided Video Diffusion Model",
+        "text_level": 1,
+        "bbox": [
+            496,
+            715,
+            781,
+            729
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The goal of MimicMotion is to generate high-quality, pose-guided human videos from a single reference image and a sequence of poses to mimic. This task involves the synthesis of realistic motion that adheres to the provided pose sequence while maintaining visual fidelity to the reference image. We exploit the ability of a specific pre-trained video diffusion model to reduce the data requirement and computational cost of training a video diffusion model from scratch. Stable Video Diffusion (SVD) (Blattmann et al., 2023) is an open-source image-to-video diffusion model trained on a large-scale video dataset. It shows good perfor",
+        "bbox": [
+            495,
+            739,
+            887,
+            905
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "header",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "bbox": [
+            165,
+            56,
+            807,
+            70
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "3",
+        "bbox": [
+            480,
+            922,
+            491,
+            934
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "mance on both video quality and diversity compared with the other contemporary models. The model structure of MimicMotion is designed to integrate a pre-trained Stable Video Diffusion (SVD) model to leverage its image-to-video generation capabilities.",
+        "bbox": [
+            84,
+            84,
+            473,
+            161
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Learning a diffusion process in pixel space is costly, especially for high-definition videos involving many frames. We follow the Latent Diffusion Model (LDM) (Rombach et al., 2022b) to encode pixels into a lower-dimensional latent space. LDM adopts a pair of autoencoders, consisting of an encoder  $\\mathcal{E}$  and a decoder  $\\mathcal{D}$ . Given a data sample  $\\mathbf{x}$ , it is encoded into the latent space as  $\\mathbf{z} = \\mathcal{E}(\\mathbf{x})$ . Conversely, the latent vector  $\\mathbf{z}$  can be decoded into pixel space via  $\\mathbf{x} = \\mathcal{D}(\\mathbf{z})$ .",
+        "bbox": [
+            84,
+            167,
+            475,
+            305
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Figure 2 shows the structure of our model. The VAE encoder on input video frames and the corresponding decoder for getting denoised video frames are both adopted from SVD and these parameters are frozen. The VAE encoder is applied independently to each frame of the input video and the conditional reference image. In contrast, the VAE decoder processes the latent features that undergo spatiotemporal interaction from U-Net. To enhance the temporal smoothness of low-level details, the VAE decoder incorporates temporal layers alongside the spatial layers.",
+        "bbox": [
+            84,
+            311,
+            475,
+            464
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The reference image and the sequence of poses are two other inputs of the model. The reference image follows two pathways: First, it is processed through CLIP (Radford et al., 2021) to extract features that are fed into each U-Net block's cross-attention. Second, the image is encoded by the frozen VAE encoder to obtain its latent representation, which is then temporally duplicated to match the video frame dimensions. This latent representation is concatenated with the video frame features along the channel dimension before entering the U-Net.",
+        "bbox": [
+            84,
+            469,
+            475,
+            621
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For pose guidance, PoseNet consisting of multiple convolution layers is designed to extract features from input pose sequences. The reason for not using the VAE encoder is that the pixel value distribution of the pose sequence is different from that of common images on which the VAE autoencoder is trained. The features of poses are then element-wisely added to the output of the first convolution layer of U-Net. Pose guidance is not added to every U-Net block due to: a) the sequence pose being extracted without temporal interaction may confuse the spatio-temporal layers within U-Net; b) excessive involvement of the pose sequence may degrade the performance of the pre-trained image-to-video model.",
+        "bbox": [
+            84,
+            628,
+            475,
+            811
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4. Confidence-Aware Pose Guidance",
+        "text_level": 1,
+        "bbox": [
+            84,
+            825,
+            354,
+            840
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Inaccurate pose estimation has a negative impact on the model's training and inference. However, accurate pose estimation is challenging in dynamic videos. The limited",
+        "bbox": [
+            84,
+            849,
+            473,
+            895
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "capability of the pose estimation model, like DWPose (Yang et al., 2023), is only one aspect of the reason. The more significant cause is the inherent uncertainty of pose from dynamic appearances and motions. Specifically, incorrect pose guidance signals can mislead the model, resulting in the generation of inaccurate or distorted outputs, as illustrated in Figure 7. Moreover, noisy pose guidance signals can lead to overfitting on samples with incorrect poses, potentially causing training instability. This in turn may hinder the model's training.",
+        "bbox": [
+            495,
+            84,
+            885,
+            236
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For this problem, we propose confidence-aware pose guidance, which leverages the confidence scores associated with each keypoint from the pose estimation model. These scores reflect the likelihood of accurate detection, with higher values indicating higher visibility, less self-occlusion from other body parts, and motion blur. Instead of applying a fixed confidence threshold to filter the keypoints, as commonly adopted in prior works (moo, 2024; Chang et al., 2023), we utilize brightness of the keypoints and limb to represent the confidence level of pose estimation. Specifically, we multiply the color assigned to each keypoint and limb by its confidence score. Consequently, keypoints and corresponding limbs with higher confidence scores will appear more significant on the pose guidance map. This method enables the model to prioritize more reliable pose information in its guidance, thereby enhancing the overall accuracy of pose-guided generation. In this way, the uncertainty of pose estimation can be conveyed through the pose guidance, making pose guidance more informative.",
+        "bbox": [
+            495,
+            243,
+            888,
+            530
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Hand region enhancement Moreover, we employ pose estimation and the associated confidence scores to alleviate region-specific artifacts, such as hand distortion, which are prevalent in the diffusion-based image and video generation models. Specifically, we identify reliable regions via thresholding keypoint confidence scores. By setting a threshold, we can distinguish between keypoints that are confidently detected and those that may be unreliable due to self-occlusion or motion blur. Keypoints with confidence scores above the threshold are considered reliable. We implement a masking strategy that generates masks based on a confidence threshold. We unmask areas where confidence scores surpass a predefined threshold, thereby identifying reliable regions. When computing the loss of the video diffusion model, the loss values corresponding to the un-masked regions are amplified by a certain scale so they can have more effect on the model training than other masked regions.",
+        "bbox": [
+            495,
+            550,
+            887,
+            824
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Specifically, to mitigate hand distortion, we compute masks using a confidence threshold for keypoints in the hand region. Only hands with all keypoint confidence scores exceeding this threshold are considered reliable, as a higher score correlates to higher visual quality. We then construct",
+        "bbox": [
+            495,
+            830,
+            888,
+            906
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "header",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "bbox": [
+            165,
+            56,
+            805,
+            69
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "4",
+        "bbox": [
+            480,
+            922,
+            491,
+            934
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/f25fa6af7807f5f514300796cb352ffce81a791286c17c431910aab1343c1108.jpg",
+        "image_caption": [
+            "Figure 3: Overview of our approach for long video generation. The colored boxes represent latent video frames. A darker color means a higher weight. The dashed boxes represent video frame features involved in latent fusion."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            89,
+            80,
+            472,
+            209
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "a bounding box around the hand by padding the boundary of these keypoints, and the enclosed rectangle is designated as unmasked. This region is subsequently assigned a larger weight in the loss calculation during the training of the video diffusion model. This selective unmasking and weighting process biases the model's learning towards hands, especially hands with higher visual quality, effectively reducing distortion and improving the overall quality.",
+        "bbox": [
+            84,
+            314,
+            475,
+            435
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.5. Progressive Latent Fusion for Long Videos",
+        "text_level": 1,
+        "bbox": [
+            84,
+            450,
+            416,
+            467
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Limited by computation resources, generating long videos containing a large number of frames is challenging. For this problem, prior works like MultiDiffusion (Bar-Tal et al., 2023) successfully use latent fusion for panoramic image generation. A similar idea can be applied to the video generation task. A straightforward approach is directly applying MultiDiffusion in the time domain, as in Lumiere (Bar-Tal et al., 2024). However, unlike spatial discontinuities between image tiles, temporal discontinuities are more noticeable to viewers, resulting in flickering or abrupt content changes. We address this by a progressive approach for generating long videos with high temporal continuity.",
+        "bbox": [
+            84,
+            474,
+            475,
+            656
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Progressive latent fusion is training-free and is integrated into the denoising process of the latent diffusion model during inference. Figure 3 shows an overview of this process. We omit the VAE for brevity. The denoising process is done in latent space in our method. In general, there are  $T$  denoising steps in total and our latent fusion is applied within each step. For a long given pose sequence, we use a pre-defined strategy for splitting the whole sequence into segments, consisting of a fixed number of frames per segment (denoted as  $N$ ), with a certain number  $(C)$  of overlapped frames between every two adjacent segments. For the sake of generation efficiency, it is common to assume that  $C \\ll N$ . During each denoising step, video segments are firstly denoised separately with the trained model, conditioning on the same reference image and the corresponding sub-sequence of poses. Algorithm 1 shows the specific de",
+        "bbox": [
+            84,
+            662,
+            475,
+            905
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [
+            "Algorithm 1 Progressive latent fusion for long videos."
+        ],
+        "code_body": "1: Input:  \n2:  $I_{\\mathrm{ref}}$ : Reference image;  \n3:  $P_i^j$ : Pose of  $j$ -th frame in  $i$ -th video segment;  \n4:  $\\mathbf{z}_i^j$ : The latent of  $j$ -th frame in  $i$ -th video segment;  \n5:  $N$ : The number of frames in each video segment;  \n6:  $C$ : The number of overlapped frames.  \n7: Output: A long sequence of latent video frames.  \n8:  $\\mathbf{z} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$   \n9:  $\\lambda_{\\mathrm{fusion}} \\gets 1 / (C + 1)$   \n10: for  $t = T$  to 1 do  \n11: for  $i = 1, 2, \\ldots$  do  \n12:  $\\mathbf{z}_i \\gets \\mathrm{DM}(\\mathbf{z}_i, I_{\\mathrm{ref}}, P_i, t)$   \n13: end for  \n14: for  $i = 1, 2, \\ldots$  do  \n15: for  $j = 1$  to  $N$  do  \n16: if  $i > 1$  and  $j \\leq C$  then  \n17:  $\\mathbf{z}_i^j \\gets j \\lambda_{\\mathrm{fusion}} \\mathbf{z}_i^j + (1 - j \\lambda_{\\mathrm{fusion}}) \\mathbf{z}_{i-1}^{N-C+j}$   \n18: else if  $j > N - C$  then  \n19:  $\\mathbf{z}_i^j \\gets (N + 1 - j) \\lambda_{\\mathrm{fusion}} \\mathbf{z}_i^j + (C - N + j) \\lambda_{\\mathrm{fusion}} \\mathbf{z}_{i+1}^{C-N+j}$   \n20: end if  \n21: end for  \n22: end for  \n23: end for  \n24: return  $\\{\\mathbf{z}_i^{1:N-C}\\}_{i=1,2,\\ldots}$",
+        "bbox": [
+            500,
+            103,
+            888,
+            494
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "tails of progressive latent fusion. As inputs, the reference image is denoted as  $I_{\\mathrm{ref}}$ , and the pose frame corresponding to  $j$ -th frame in  $i$ -th video segment is denoted as  $P_i^j$ . We use  $\\mathbf{z}_i^j$  to denote the latent feature of  $j$ -th frame in  $i$ -th video segment. The denoising process starts from the maximum time step  $T$  and the latent features are initialized with a normal distribution  $\\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ . Within each denoising step at time step  $t$ , the reversed diffusion process defined by the trained model (DM) is applied to the latent features of each video segment numbered  $i$  separately, with  $\\mathbf{z}_i$ ,  $I_{\\mathrm{ref}}$ ,  $P_i$  and  $t$  as inputs. During the latent fusion stage, for every two adjacent video segments, the involved video frames are then fused. To avoid the corruption of temporal smoothness near video segment boundaries after latent fusion, we propose progressive latent fusion. For a video frame involved in latent fusion, its fusion weight is determined by its relative position in the video segment it belongs. Specifically, if a frame is close to the segment it belongs to, it will be assigned a heavier weight. For implementation, a fusion scale is pre-defined as  $\\lambda_{\\mathrm{fusion}} = 1 / (C + 1)$  for controlling the level of latent fusion.",
+        "bbox": [
+            495,
+            532,
+            887,
+            852
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After applying  $T$  iterative denoising steps, we aggregate the final long sequence by combining the denoised overlapped video segments in latent space. The resulting aggregated",
+        "bbox": [
+            495,
+            859,
+            885,
+            906
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "header",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "bbox": [
+            165,
+            56,
+            805,
+            70
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "5",
+        "bbox": [
+            480,
+            922,
+            491,
+            934
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "long video in latent space is then fed to the temporal VAE decoder to reconstruct the long video.",
+        "bbox": [
+            84,
+            84,
+            473,
+            114
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            84,
+            133,
+            218,
+            152
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            84,
+            160,
+            282,
+            175
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Due to the lack of suitable high-quality academic datasets, we collect 4,436 human dancing videos from the internet for training. The average length is 20.1s. We adopt the pre-trained weights from SVD (Blattmann et al., 2023). The PoseNet is trained from scratch. We train our model on 8 NVIDIA A100 GPUs for 20 epochs, with a batch size of 8 and 16 frames per clip. The loss weight of the hand region is 10. The learning rate is  $10^{-5}$  with a linear warmup of 500 iterations. We tune all parameters in the UNet and PoseNet.",
+        "bbox": [
+            84,
+            183,
+            475,
+            320
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Comparison to State-of-the-Art Methods",
+        "text_level": 1,
+        "bbox": [
+            84,
+            337,
+            403,
+            351
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We compare our method with latest state-of-the-art pose-guided human video generation methods, including MagicAnymate (Xu et al., 2024), MagicPose (Chang et al., 2023), Moore-AnymateAnyone (moo, 2024), and MuseV (Xia et al., 2024). For the testing protocol of previous works (Wang et al., 2023; Chang et al., 2023), we adopt the TikTok (Jafarian & Park, 2021) dataset and use sequence 335 to 340 for our evaluation.",
+        "bbox": [
+            84,
+            359,
+            475,
+            479
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We conduct qualitative and quantitative comparisons, complemented by a user study. Each method has a different input aspect ratio. To ensure a fair comparison, we only consider the central square region of the videos. Specifically, we individually apply a center crop to the reference image and pose sequence to match each method's input aspect ratio. Then, we extract the center squares from the generated videos for comparison across different methods.",
+        "bbox": [
+            84,
+            488,
+            475,
+            608
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Qualitative evaluation We conduct qualitative comparisons between the selected baselines and our method. In Figure 4, we showcase sample frames to highlight the superior quality of individual frames produced by our method. Additionally, in Figure 5, we illustrate the temporal differences, demonstrating the enhanced temporal smoothness of our approach compared to existing methods.",
+        "bbox": [
+            84,
+            633,
+            475,
+            739
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Figure 4 presents a comparison of the generated frames, where each row represents a distinct example. The first row demonstrates the superior hand quality achieved by our approach, while the second row showcases the improved adherence to pose guidance. These improvements directly result from our confidence-aware pose guidance and hand region enhancement design.",
+        "bbox": [
+            84,
+            747,
+            475,
+            852
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Importantly, our method shows superior temporal smoothness, characterized by smooth motion and minimal flickering. Figure 5 demonstrates this through pixel-wise",
+        "bbox": [
+            84,
+            859,
+            475,
+            905
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/585445b80438a52457ea93a33f67c3da8eee20a52b14f0bdf9f999851486f15c.jpg",
+        "image_caption": [
+            "Figure 4: Qualitative comparison to the state-of-the-art methods on TikTok test set. Our method achieves better hand generation quality and adheres better to the reference pose. Note that our method is not trained on the TikTok dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            498,
+            80,
+            885,
+            194
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/27bc3acfc4348dcf414cd321147f2f3f1fc813024444d569287b7329fc30f6c2.jpg",
+        "image_caption": [
+            "Figure 5: Comparison of temporal smoothness via frame differences. MagicPose exhibits abrupt transitions, whereas Moore and MuseV show flickering texture and text. In contrast, our method maintains stable frame differences."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            498,
+            297,
+            885,
+            498
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "differences between consecutive frames. While Magic-Pose (Chang et al., 2023) shows abrupt transitions, Moore-AnymateAnyone (moo, 2024) exhibits unstable clothing textures, and MuseV (Xia et al., 2024) struggles with consistent text rendering, our method maintains stable frame-to-frame transitions without visible artifacts. Videos in the supplementary material further demonstrate this improvement. This enhancement is enabled by our confidence-aware pose guidance, which mitigates the impact of inaccurate pose inputs and the associated temporal noise. By weighting pose signals by confidence, our method ensures high temporal smoothness in the presence of pose guidance noise.",
+        "bbox": [
+            496,
+            599,
+            887,
+            781
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Quantitative evaluation Our method outperforms existing approaches across all evaluation metrics, as shown in Table 1. We evaluated performance on test sequences from the TikTok dataset (Jafarian & Park, 2021) using both video-level metrics FID-VID (Balaji et al., 2019) and FVD (Unterthiner et al., 2018), as well as frame-level metrics PSNR and SSIM (Wang et al., 2004) following (Wang et al., 2023).",
+        "bbox": [
+            496,
+            800,
+            888,
+            905
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "header",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "bbox": [
+            165,
+            56,
+            807,
+            70
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "6",
+        "bbox": [
+            480,
+            922,
+            491,
+            934
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/f1117c48341dfbb9164d9c414e812c9914d3921a0b470db42ab6b43559c8cacd.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>FID-VID↓</td><td>FVD↓</td><td>SSIM↑</td><td>PSNR↑</td></tr><tr><td>MagicAnymate</td><td>16.2</td><td>848</td><td>0.740</td><td>17.5</td></tr><tr><td>MagicPose</td><td>13.3</td><td>916</td><td>0.776</td><td>18.8</td></tr><tr><td>Moore</td><td>12.4</td><td>728</td><td>0.758</td><td>18.7</td></tr><tr><td>MuseV</td><td>14.6</td><td>754</td><td>0.766</td><td>17.6</td></tr><tr><td>MimicMotion (ours)</td><td>9.3</td><td>594</td><td>0.795</td><td>20.1</td></tr></table>",
+        "bbox": [
+            91,
+            82,
+            467,
+            172
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/1400efdb2edd84681571215f10068b4bc9f36e7b172af2582a8ca8c30af78754.jpg",
+        "image_caption": [
+            "Figure 6: Preference of videos generated by MimicMotion (ours) over baseline methods on the TikTok dataset test split. Users consistently prefer MimicMotion over other methods."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            138,
+            276,
+            431,
+            354
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The results demonstrate our method's superior performance compared to current state-of-the-art approaches.",
+        "bbox": [
+            84,
+            440,
+            473,
+            470
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "User study To supplement our quantitative and qualitative evaluations, we conduct a user study to assess the subjective preferences of participants regarding the generated videos on the TikTok dataset test split. The study involves showing two video clips—one generated by our method and the other by one of the baseline methods—to a diverse group of users. Participants are instructed to select the video that they perceived as having higher quality, considering factors including image quality and temporal smoothness of characters and clothing. We collected data from 36 participants, with each participant evaluating 6 video pairs for our method against each baseline method. As shown in Figure 6, the results indicate a strong preference for MimicMotion over the baseline methods. In comparison to MagicPose and Moore, the participants almost favored all videos produced by our method. Despite MuseV showing higher image quality compared to other baselines, the preference for videos produced by our method still reached  $75.5\\%$ . These findings align with our qualitative and quantitative evaluation, reinforcing the effectiveness of our method in meeting user expectations for high-quality human video generation.",
+        "bbox": [
+            86,
+            488,
+            475,
+            805
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            84,
+            821,
+            225,
+            835
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table 2, confidence-aware pose guiding and hand region enhancement improve all metrics. Progressive latent fusion notably improves FVD scores, indicating better temporal coherence, though other metrics may not reflect",
+        "bbox": [
+            84,
+            845,
+            473,
+            905
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/7248a6dc4568f1de633e36aae8fc0732debfe367be5c2113d3addac58dd80d03.jpg",
+        "table_caption": [
+            "Table 1: Quantitative comparison to state-of-the-art methods: MagicAnymate (Xu et al., 2024), MagicPose (Chang et al., 2023), Moore (moo, 2024), MuseV (Xia et al., 2024). We evaluate on the TikTok dataset test split. MimicMotion is the best in all metrics."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Hand</td><td>Conf.</td><td>Prog.</td><td>FID-VID↓</td><td>FVD↓</td><td>SSIM↑</td><td>PSNR↑</td></tr><tr><td></td><td></td><td></td><td>14.6</td><td>776</td><td>0.760</td><td>18.0</td></tr><tr><td></td><td></td><td>✓</td><td>15.0</td><td>678</td><td>0.758</td><td>17.9</td></tr><tr><td></td><td>✓</td><td>✓</td><td>12.2</td><td>623</td><td>0.787</td><td>18.4</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>9.3</td><td>594</td><td>0.795</td><td>20.1</td></tr></table>",
+        "bbox": [
+            504,
+            82,
+            880,
+            161
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Ablation studies of hand region augmentation (hand), confidence-aware pose guiding (conf.), and progressive latent fusion (prog.).",
+        "bbox": [
+            496,
+            171,
+            885,
+            218
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "this improvement. We analyze each component below.",
+        "bbox": [
+            496,
+            247,
+            856,
+            263
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Confidence-aware pose guiding Figure 7 shows the effectiveness of confidence-aware pose guidance. Each row corresponds to one sample. The left side shows three images to extract the pose. On the right side, we plot the corresponding pose guiding signals, both with and without confidence-aware pose guiding. From the guiding signals, we can see that there are errors in the pose estimated by DW- Pose. Nevertheless, our confidence-aware design minimizes the impact of incorrect pose estimation in guidance signals.",
+        "bbox": [
+            495,
+            282,
+            887,
+            419
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Specifically, in the case of Pose 1, the estimation exhibits a duplicate detection issue, which leads to the inclusion of duplicate keypoints. In the case of Pose 2, there is one hand obscured, and the keypoints of this hand are incorrectly estimated on the other hand; In the case of Pose 3, the right elbow is obscured, but it is still detected with confidence above the threshold thus falsely remains in the guidance signal. These problems lead to confusing hand guidance signals and ultimately lead to distortions such as deformed hands or wrong spatial relationships in the generated frames.",
+        "bbox": [
+            495,
+            425,
+            888,
+            577
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In contrast, by integrating confidence scores into pose guidance, our method effectively mitigates these issues. The confidence scores provide a measure of reliability for each keypoint, allowing our model to weigh the guidance signals accordingly. Specifically, keypoint with lower confidence, which typically correspond to inaccurate keypoints caused by self-occlusion or motion blur, will be of less significance in the guidance. This approach leads to clearer and richer pose guidance, as the influence of potentially erroneous keypoints is reduced. The corresponding generation results demonstrate how our method enhances the robustness of generation against false guiding signals (Pose 1 and Pose 2) and offers visibility hints to resolve the front-back ambiguity of 2D pose estimation (Pose 3).",
+        "bbox": [
+            495,
+            584,
+            887,
+            795
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Hand region enhancement In conjunction with confidence-aware pose guiding, we further improve hand generation quality by assigning a higher weight to the hand region in the training loss. Figure 8 compares the generation result with and without hand region enhancement, using the same reference image and pose guidance. All experiments",
+        "bbox": [
+            495,
+            814,
+            885,
+            905
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "header",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "bbox": [
+            165,
+            56,
+            805,
+            70
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "7",
+        "bbox": [
+            480,
+            922,
+            491,
+            934
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/42fee89479cfe40139d2fd943a8e556d277fc145ace02b12f22fd0049236b312.jpg",
+        "image_caption": [
+            "Figure 7: Confidence-aware pose guiding enhances generation robustness to false guiding signals (Pose 1&2) and provides hints to handle self-occlusion (Pose 3)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            80,
+            472,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/c159c5a6b82939f889572d74f86dd933afa618ea2d6626240289b67910c69d6c.jpg",
+        "image_caption": [
+            "Figure 8: Hand region enhancement consistently reduces hand distortion and improves visual appeal, using the same reference image and pose guidance."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            364,
+            473,
+            467
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "incorporate confidence-aware pose guidance. The hands in the first row are cropped from the generated video frames of a model trained without hand region enhancement, which exhibits noticeable distortions, such as irregular and misplaced fingers. In contrast, the results of the model trained with hand region enhancement (the second row) show consistent improvements in hand generation quality and a reduction in hand distortion. These results show the effectiveness of the proposed hand region enhancement design, which substantially mitigates hand distortion, which is a prevalent challenge in diffusion-based models. As humans often focus on hands, this design aligns the training process with human preferences, thereby enhancing the overall visual quality of the results. There is no noticeable degradation in the quality of the remaining parts.",
+        "bbox": [
+            84,
+            554,
+            473,
+            781
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Progressive latent fusion To achieve seamless transitions between video segments, we introduce progressive latent fusion, a technique that gradually blends frames in the overlapped regions of consecutive video segments. Figure 9 presents a comparison of a sample transition with an overlap of 6 frames between two 16-frame segments. We visualize 4 sample frames. The original MultiDiffusion approach",
+        "bbox": [
+            84,
+            800,
+            475,
+            905
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b7f0f31dd40c62b140b4f94b3189b9d63f22637e5f75416b07fe4e93862a1e1c.jpg",
+        "image_caption": [
+            "Figure 9: Progressive latent fusion ensures smooth segment transitions by avoiding temporal discontinuities of weights in simple averaging, which reduces artifacts in Y-T slices."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            498,
+            80,
+            883,
+            258
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "employs a simple averaging of frames within the overlap region. This method assigns equal weight to all frames in the overlap region, irrespective of their temporal position (whether they are closer to the preceding or subsequent segment). It lacks a gradual transition from one segment to another, causing noticeable flickers at denoising window boundaries. This is evident in the y-t slice shown on the left of Figure 9a, where the transition at segment boundaries is abrupt. The right side of the figure shows four frames at the segment boundary. Note that the background in the top-left corner (enlarged) is initially clear in segment 1. It suddenly becomes blurry in the overlapped region and then suddenly reverts to clearer in the main part of segment 2. This artifact is not observed when progressive latent fusion is applied.",
+        "bbox": [
+            495,
+            354,
+            885,
+            566
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The proposed progressive latent fusion approach (see Figure 9b) mitigates these issues. The left y-t slice shows smooth transitions across segment boundaries, avoiding abrupt changes. The right side of the figure demonstrates that this approach eliminates a sudden blurring, mitigating flicking artifacts, thus improving the overall visual temporal coherence for long video generation.",
+        "bbox": [
+            495,
+            573,
+            885,
+            679
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            496,
+            698,
+            614,
+            713
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this study, we introduce MimicMotion, a pose-guided human video generation model that leverages confidence-aware pose guidance and progressive latent fusion for producing high-quality, long videos with human motion guided by pose. Through extensive experiments and ablation studies, we show that our model achieves superior adaptation to noisy pose estimation, enhancing hand quality and ensuring temporal smoothness. The integration of confidence scores into pose guidance, the enhancement of hand region loss, and the implementation of progressive latent fusion are crucial in achieving these improvements, resulting in more visually compelling and realistic human video generation.",
+        "bbox": [
+            495,
+            724,
+            885,
+            905
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "header",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "bbox": [
+            165,
+            56,
+            807,
+            70
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "8",
+        "bbox": [
+            480,
+            922,
+            491,
+            934
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Impact Statement",
+        "text_level": 1,
+        "bbox": [
+            86,
+            83,
+            240,
+            99
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "MimicMotion advances human motion video generation technology with potential applications in creative content production, digital human, and entertainment. The framework enables the creation of high-quality, controllable human motion videos. However, like other video generation technologies, there are potential risks of misuse for creating deceptive content that could be used for misinformation. Responsible development and deployment of such technology should include appropriate safeguards, such as watermarking, detection methods, and clear usage guidelines. We encourage the development of technical capabilities and ethical frameworks for video generation technologies.",
+        "bbox": [
+            84,
+            109,
+            475,
+            291
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            86,
+            309,
+            181,
+            325
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "MooreThreads/Moore-AnimateAnyone, May 2024. URL https://github.com/MooreThreads/Moore-AnimateAnyone. original-date: 2024-01-12T07:55:21Z.",
+            "Balaji, Y., Min, M. R., Bai, B., Chellappa, R., and Graf, H. P. Conditional gan with discriminative filter generation for text-to-video synthesis. In *IJCAI*, volume 1, pp. 2, 2019.",
+            "Bao, F., Xiang, C., Yue, G., He, G., Zhu, H., Zheng, K., Zhao, M., Liu, S., Wang, Y., and Zhu, J. Vidu: a highly consistent, dynamic and skilled text-to-video generator with diffusion models. arXiv preprint arXiv:2405.04233, 2024.",
+            "Bar-Tal, O., Yariv, L., Lipman, Y., and Dekel, T. Multi-diffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023.",
+            "Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Li, Y., Michaeli, T., et al. Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024.",
+            "Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023.",
+            "Chan, C., Ginosar, S., Zhou, T., and Efros, A. A. Everybody dance now. In Proceedings of the IEEE/CVF international conference on computer vision, pp. 5933-5942, 2019.",
+            "Chang, D., Shi, Y., Gao, Q., Xu, H., Fu, J., Song, G., Yan, Q., Zhu, Y., Yang, X., and Soleymani, M. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In *Forty-first International Conference on Machine Learning*, 2023."
+        ],
+        "bbox": [
+            86,
+            333,
+            477,
+            905
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Chen, H., Xia, M., He, Y., Zhang, Y., Cun, X., Yang, S., Xing, J., Liu, Y., Chen, Q., Wang, X., et al. Videocaster1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023.",
+            "Chen, H., Zhang, Y., Cun, X., Xia, M., Wang, X., Weng, C., and Shan, Y. Videocaster2: Overcoming data limitations for high-quality video diffusion models. arXiv preprint arXiv:2401.09047, 2024.",
+            "Feng, M., Liu, J., Yu, K., Yao, Y., Hui, Z., Guo, X., Lin, X., Xue, H., Shi, C., Li, X., et al. Dreamoving: A human video generation framework based on diffusion models, 2023.",
+            "Guo, Y., Yang, C., Rao, A., Wang, Y., Qiao, Y., Lin, D., and Dai, B. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023.",
+            "He, Y., Yang, T., Zhang, Y., Shan, Y., and Chen, Q. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022.",
+            "Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020.",
+            "Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., and Fleet, D. J. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022.",
+            "Hu, L. Animate anyone: Consistent and controllable imaged-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8153-8163, 2024.",
+            "Jafarian, Y. and Park, H. S. Learning high fidelity depths of dressed humans by watching social media dance videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 12753-12762, June 2021.",
+            "Karras, T., Aittala, M., Aila, T., and Laine, S. Elucidating the Design Space of Diffusion-Based Generative Models, October 2022. URL http://arxiv.org/abs/2206.00364.arXiv:2206.00364 [cs, stat].",
+            "Li, H., Li, Y., Yang, Y., Cao, J., Zhu, Z., Cheng, X., and Long, C. Dispose: Disentangling pose guidance for controllable human image animation. arXiv preprint arXiv:2412.09349, 2024.",
+            "Ma, X., Wang, Y., Jia, G., Chen, X., Liu, Z., Li, Y.-F., Chen, C., and Qiao, Y. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024a."
+        ],
+        "bbox": [
+            498,
+            84,
+            887,
+            904
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "header",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "bbox": [
+            165,
+            56,
+            807,
+            70
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "9",
+        "bbox": [
+            480,
+            922,
+            491,
+            934
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ma, Y., He, Y., Cun, X., Wang, X., Chen, S., Li, X., and Chen, Q. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp. 4117-4125, 2024b.",
+            "Peebles, W. and Xie, S. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pp. 4195-4205, 2023.",
+            "Peng, B., Wang, J., Zhang, Y., Li, W., Yang, M.-C., and Jia, J. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024.",
+            "Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp. 8748-8763. PMLR, 2021.",
+            "Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 10684-10695, 2022a.",
+            "Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10684-10695, 2022b.",
+            "Siarohin, A., Lathuilière, S., Tulyakov, S., Ricci, E., and Sebe, N. First order motion model for image animation. Advances in neural information processing systems, 32, 2019.",
+            "Siarohin, A., Woodford, O. J., Ren, J., Chai, M., and Tulyakov, S. Motion representations for articulated animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13653-13662, 2021.",
+            "Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.",
+            "Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020.",
+            "Tu, S., Xing, Z., Han, X., Cheng, Z.-Q., Dai, Q., Luo, C., and Wu, Z. Stableimator: High-quality identity-preserving human image animation. arXiv preprint arXiv:2411.17697, 2024."
+        ],
+        "bbox": [
+            86,
+            84,
+            475,
+            904
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., and Gelly, S. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018.",
+            "Voleti, V., Jolicoeur-Martineau, A., and Pal, C. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in neural information processing systems, 35:23371-23385, 2022.",
+            "Wang, Q., Bai, X., Wang, H., Qin, Z., and Chen, A. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024a.",
+            "Wang, Q., Jiang, Z., Xu, C., Zhang, J., Wang, Y., Zhang, X., Cao, Y., Cao, W., Wang, C., and Fu, Y. Vividpose: Advancing stable video diffusion for realistic human image animation. arXiv preprint arXiv:2405.18156v1, 2024b.",
+            "Wang, T., Li, L., Lin, K., Zhai, Y., Lin, C.-C., Yang, Z., Zhang, H., Liu, Z., and Wang, L. Disco: Disentangled control for realistic human dance generation. arXiv preprint arXiv:2307.00040, 2023.",
+            "Wang, W., Liu, J., Lin, Z., Yan, J., Chen, S., Low, C., Hoang, T., Wu, J., Liew, J. H., Yan, H., et al. Magicvideo v2: Multi-stage high-aesthetic video generation. arXiv preprint arXiv:2401.04468, 2024c.",
+            "Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004.",
+            "Wu, J. Z., Ge, Y., Wang, X., Lei, S. W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., and Shou, M. Z. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 7623-7633, 2023.",
+            "Xia, Z., Chen, Z., Wu, B., Li, C., Hung, K.-W., Zhan, C., He, Y., and Zhou, W. Musev: Infinite-length and high fidelity virtual human video generation with visual conditioned parallel denoising. arxiv, 2024.",
+            "Xu, Z., Zhang, J., Liew, J. H., Yan, H., Liu, J.-W., Zhang, C., Feng, J., and Shou, M. Z. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1481-1490, 2024.",
+            "Yan, W., Zhang, Y., Abbeel, P., and Srinivas, A. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021."
+        ],
+        "bbox": [
+            500,
+            84,
+            885,
+            905
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "header",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "bbox": [
+            165,
+            56,
+            807,
+            69
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "10",
+        "bbox": [
+            477,
+            922,
+            495,
+            934
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yang, Z., Zeng, A., Yuan, C., and Li, Y. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 4210-4220, 2023.",
+            "Yin, S., Wu, C., Yang, H., Wang, J., Wang, X., Ni, M., Yang, Z., Li, L., Liu, S., Yang, F., et al. Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346, 2023.",
+            "Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A. G., Yang, M.-H., Hao, Y., Essa, I., et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10459-10469, 2023.",
+            "Zhang, L., Rao, A., and Agrawala, M. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 3836-3847, 2023.",
+            "Zhao, J. and Zhang, H. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 3657-3666, 2022.",
+            "Zheng, T., Geng, C., Jiang, P., Wan, B., Zhang, H., Chen, J., Wang, J., and Li, B. Non-uniform timestep sampling: Towards faster diffusion model training. In Cai, J., Kankanhalli, M. S., Prabhakaran, B., Boll, S., Subramanian, R., Zheng, L., Singh, V. K., Cesar, P., Xie, L., and Xu, D. (eds.), Proceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 - 1 November 2024, pp. 7036-7045. ACM, 2024a.",
+            "Zheng, T., Jiang, P., Wan, B., Zhang, H., Chen, J., Wang, J., and Li, B. Beta-tuned timestep diffusion model. In Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., and Varol, G. (eds.), Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part III, volume 15061 of Lecture Notes in Computer Science, pp. 114-130. Springer, 2024b.",
+            "Zhu, S., Chen, J. L., Dai, Z., Xu, Y., Cao, X., Yao, Y., Zhu, H., and Zhu, S. Champ: Controllable and consistent human image animation with 3d parametric guidance, 2024."
+        ],
+        "bbox": [
+            86,
+            84,
+            475,
+            787
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "header",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "bbox": [
+            165,
+            56,
+            807,
+            70
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "11",
+        "bbox": [
+            477,
+            922,
+            493,
+            934
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A. Training Details",
+        "text_level": 1,
+        "bbox": [
+            84,
+            83,
+            250,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "MimicMotion is tuned from Stable Video Diffusion (Blattmann et al., 2023). The video samples are resized to the resolution of  $576 \\times 1024$ . The training dataset contains 4,436 human dancing videos collected from the internet. We train the model for 20 epochs at a batch size of 8. We follow Stable video diffusion and adopt the noise distribution, i.e.  $\\log \\sigma \\sim \\mathcal{N}(P_{\\mathrm{mean}}, P_{\\mathrm{std}}^2)$ , proposed by Karras et al (Karras et al., 2022) with parameter  $P_{\\mathrm{mean}} = 0.5$  and  $P_{\\mathrm{std}} = 1.4$ . We train our model on 8 NVIDIA A100 GPUs with a batch size of 8 and 16 frames per clip. The loss weight assigned to the hand region is set at 10, while the remaining part is left to 1. The learning rate is  $10^{-5}$  with a linear warmup schedule of 500 iterations. We tune all parameters in the UNet and PoseNet.",
+        "bbox": [
+            84,
+            109,
+            887,
+            214
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B. PoseNet Model Architecture",
+        "text_level": 1,
+        "bbox": [
+            84,
+            234,
+            349,
+            250
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The PoseNet consists of convolutional and activation layers. The detailed architecture is listed in Table 3.",
+        "bbox": [
+            84,
+            260,
+            774,
+            275
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/bc5cc09746b9394ff5fb61d12f31f11ed604044e65ce13074dcb0afbab878c33.jpg",
+        "table_caption": [
+            "Table 3: Detailed Architecture of PoseNet"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>#</td><td>Type</td><td>Channels</td><td>Kernel size</td><td>Stride</td><td>Pad</td></tr><tr><td>1</td><td>Input</td><td>3</td><td>-</td><td>-</td><td>-</td></tr><tr><td>2</td><td>Conv2d</td><td>3</td><td>3x3</td><td>1</td><td>1</td></tr><tr><td>3</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>4</td><td>Conv2d</td><td>16</td><td>4x4</td><td>2</td><td>1</td></tr><tr><td>5</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>6</td><td>Conv2d</td><td>16</td><td>3x3</td><td>1</td><td>1</td></tr><tr><td>7</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>8</td><td>Conv2d</td><td>32</td><td>4x4</td><td>2</td><td>1</td></tr><tr><td>9</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>10</td><td>Conv2d</td><td>32</td><td>3x3</td><td>1</td><td>1</td></tr><tr><td>11</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>12</td><td>Conv2d</td><td>64</td><td>4x4</td><td>2</td><td>1</td></tr><tr><td>13</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>14</td><td>Conv2d</td><td>64</td><td>3x3</td><td>1</td><td>1</td></tr><tr><td>15</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>16</td><td>Conv2d</td><td>128</td><td>3x3</td><td>1</td><td>1</td></tr><tr><td>17</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>18</td><td>Conv2d</td><td>320</td><td>1x1</td><td>1</td><td>0</td></tr></table>",
+        "bbox": [
+            302,
+            315,
+            671,
+            571
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C. Additional Details on User Study",
+        "text_level": 1,
+        "bbox": [
+            84,
+            598,
+            388,
+            616
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The user study was designed to evaluate the subjective preferences of participants regarding the quality of videos generated by our method, MimicMotion, compared to several baseline methods on the TikTok dataset test split. Participants were instructed to select the video they perceived as having higher quality. The data collection process was conducted through an online interface, as depicted in Figure 10, which allowed participants to easily view and compare the video clips and submit their preferences. The results of our method will randomly appear in the video on the left or the right, with the other video being the comparative method.",
+        "bbox": [
+            84,
+            625,
+            887,
+            715
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D. Cross-Domain Results: Cartoon and Animal Videos",
+        "text_level": 1,
+        "bbox": [
+            84,
+            734,
+            545,
+            750
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "To demonstrate the versatility of MimicMotion, we evaluate its performance on cross-domain scenarios, including 4 cartoon and 4 animal dancing videos. The videos are available on our anonymous project page. As shown in Figure 11, the model generates cartoon dancing videos while maintaining the artistic style of the characters. For animal subjects (Figure 12), our method produces reasonable results despite their significantly different appearance from humans.",
+        "bbox": [
+            84,
+            760,
+            887,
+            821
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "These results highlight MimicMotion's generalizability, which stems from two key factors: first, the pose feature space of humans, cartoons, and animals may share a common subspace that is preserved in our training, human control signals can also guide cartoons and animals. Second, we map the limb lengths of the pose template to the reference character following MuseV (Xia et al., 2024), making the pose guidance signal closer to the reference character. This helps the model maintain the character's appearance while adapting poses across varying body proportions.",
+        "bbox": [
+            84,
+            828,
+            887,
+            904
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "header",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "bbox": [
+            163,
+            56,
+            807,
+            70
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "12",
+        "bbox": [
+            477,
+            922,
+            495,
+            934
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Play all",
+        "bbox": [
+            210,
+            320,
+            251,
+            330
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/f7efef5875d9b396a0e0145f15e688ccb8f18a0e35598b16e3318537f6b7d124.jpg",
+        "image_caption": [
+            "Figure 10: Snapshot of the user preference selection interface that prompts users to evaluate image quality and temporal smoothness. Note that the two videos play automatically, with our method randomly appearing on one side."
+        ],
+        "image_footnote": [
+            "Please pay attention to the image quality, whether there is flickering, the temporal consistency of the characters and clothing, and select the higher quality video (you can rate the video after it finishes playing)."
+        ],
+        "bbox": [
+            207,
+            330,
+            763,
+            546
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left Right",
+        "bbox": [
+            210,
+            608,
+            277,
+            621
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "header",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "bbox": [
+            166,
+            56,
+            805,
+            70
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "13",
+        "bbox": [
+            477,
+            922,
+            495,
+            934
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/67ec9036ebe99358a14dce64d98e027a513b114bdfb968a2f472cd3d9799f677.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            259,
+            170,
+            406,
+            335
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/0931b0f9bb61c8436a65bab1db7b445a002d0ec9a48573aac016077c4983e3e1.jpg",
+        "image_caption": [
+            "Reference images"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            429,
+            171,
+            547,
+            335
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/cd1e595574d8de41b001de6a7cbf8fc48b85b3c0caea2b790cb56ed0071df2ff.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            584,
+            172,
+            705,
+            338
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/0a4785c0ce9561c58295a2bd6a52c4f38e855066593022e2eb078a3cbced4cf5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            733,
+            172,
+            864,
+            338
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/b84936adffe1390da1e0cca9f282e46ad116e1d6edc2ba459ab9d7e78b216a89.jpg",
+        "image_caption": [
+            "Reference pose"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            98,
+            367,
+            243,
+            563
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/9f178ebd4d7a98933e636ea89a24bacba36ed33cd474f8241bb0d5506be62da3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            261,
+            366,
+            406,
+            563
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/840e441f1373437bafcfaf9c1ec9ef64036cf91c2ff08c5057db021d28fa0bc9.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            444,
+            367,
+            553,
+            563
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/6953977940b6372ef66a5dd2ac23ce63dfe69215cf592b0dcdff4352e8c5c235.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            570,
+            367,
+            717,
+            563
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/916eb2ad69576d807d969c40703861dbaacb607065f0979f7738294fdac48970.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            730,
+            369,
+            870,
+            563
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/298fcd3e3041b08a461a2e3b3d0195ac3124b36164c6d1222eed9abd5e83a282.jpg",
+        "image_caption": [
+            "Figure 11: Generated cartoon dancing videos. Our method successfully preserves the artistic style of cartoon characters while generating natural motions that follow the input poses. Top: reference frame; Left: reference pose guidance; Bottom: generated frames at different timestamps."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            98,
+            573,
+            243,
+            771
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/39940397f7c77b65c14a842d75c4e25df342ebb39cb8f1a8557dc253dbd14af9.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            261,
+            573,
+            406,
+            770
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/b799e14f26eb31aacd56159c50d967baf8d08261669c730972c523a26fca05c2.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            442,
+            574,
+            560,
+            770
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/f0e73a331087383708f4f3c71a4d6e4d3ca01faf9dd90e5594548ba9ee562deb.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            570,
+            573,
+            715,
+            770
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/b83746673bc3cfc89c689e80b43b0ab0c027ffb013f3d6c06ebfbffe5a639ae6.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            732,
+            574,
+            870,
+            770
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "header",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "bbox": [
+            165,
+            56,
+            807,
+            70
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "14",
+        "bbox": [
+            477,
+            922,
+            495,
+            934
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/7f88303698c2b7b00706038d71e3a843c0a5397c2cda077085ab993d322ea6c7.jpg",
+        "image_caption": [
+            "Reference images"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            253,
+            170,
+            867,
+            343
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/ae48b8d5ca551bd42ee5c18ba805787666292cba97036318da8fd30549b581c1.jpg",
+        "image_caption": [
+            "Reference pose"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            98,
+            367,
+            243,
+            563
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/c23541ea3b05a15428e704a190a7efb841023adda176684006ddff03dc49aa5c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            261,
+            366,
+            406,
+            563
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/3be0c670674894f95ec83556ce7c6f7d9acee0bf715f9618b6fa0729a0ac201d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            416,
+            366,
+            560,
+            561
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/b63ac8a98ef0a6b29882054afbebf8251684b70a668eb661611c984b9cd3ec81.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            601,
+            388,
+            687,
+            534
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/03cf025c93bb9f7a96b079b309c6728bb8132b87e1cdce3055e750adf9de9b67.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            725,
+            368,
+            870,
+            551
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/7bcaca0935a8bfac1d8cb3e7215ab0089f9a1f25b5a0dba6efa6ade3bad0f391.jpg",
+        "image_caption": [
+            "Figure 12: Generated animal dancing videos. MimicMotion can generalize to animal dancing zero-shot, producing plausible motions despite the significant appearance differences from humans. Top: reference frame; Left: reference pose guidance; Bottom: generated frames at different timestamps."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            98,
+            573,
+            243,
+            771
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/0b8a5eaacfca497ec4cedcabafda5330880c8a3778546479f5c3aae08be73f6e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            261,
+            573,
+            406,
+            770
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/3c865da86ee1080c8f35f507d29cf695a11c86cc6a0fb6430db13c06dd94a676.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            416,
+            573,
+            560,
+            770
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/280f92deb0f07a582eca28e5b921e0f85170d66add85e36a4d04d73b779f1343.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            606,
+            595,
+            689,
+            739
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/1d536d6cef967e4d9da0ce1c7d88cbd5e371e660efc33d807979397e286bee80.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            725,
+            575,
+            870,
+            758
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "header",
+        "text": "MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance",
+        "bbox": [
+            165,
+            56,
+            807,
+            70
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "15",
+        "bbox": [
+            477,
+            922,
+            493,
+            934
+        ],
+        "page_idx": 14
+    }
+]

2025/MimicMotion_ High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance/06747124-f5c1-49e9-89da-3b2d84cc796b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8e0f88cfc1a48afaee8bb9f4dab2ae85d50e686da66e8b5a5c57f37fbb8b665
+size 8295388

2025/MimicMotion_ High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance/full.md

@@ -0,0 +1,377 @@
+# MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance
+
+Yuang Zhang  $^{12}$  Jiaxi Gu  $^{1}$  Li-Wen Wang  $^{1}$  Han Wang  $^{12}$  Junqi Cheng  $^{1}$  Yuefeng Zhu  $^{1}$  Fangyuan Zou  $^{1}$
+
+![](images/19b946f1c7307a0afd877f0e6b2a0148ff7cad305b0d1d9ff9c2403a5064aaeb.jpg)
+
+![](images/59b037784c53cb0f54365680bd263d90d07472be01e047d6d79cca6983436ee7.jpg)
+
+![](images/4109f4a128db299605945f91f88bf2950f52128c1b848cd1a1534f5a59bf48aa.jpg)  
+Figure 1: Pose-guided dancing and talking videos generated by MimicMotion showcase its capability to produce diverse human motions and long videos.
+
+# Abstract
+
+In recent years, while generative AI has advanced significantly in image generation, video generation continues to face challenges in controllability, length, and detail quality, which hinder its application. We present MimicMotion, a framework for generating high-quality human videos of arbitrary length using motion guidance. Our approach has several highlights. Firstly, we introduce confidence-aware pose guidance that ensures high frame quality and temporal smoothness. Secondly, we introduce regional loss amplification based on pose confidence, which reduces image distortion in key regions. Lastly, we propose a progressive latent fusion strategy to generate long and smooth videos. Experiments demonstrate the effectiveness of our approach in producing high-quality human motion videos. Videos and comparisons are available at https://tencent.github.io/MimicMotion.
+
+$^{1}$ Tencent  $^{2}$ Shanghai Jiao Tong University. Correspondence to: Jiaxi Gu <imjiaxi@gmail.com>.
+
+Proceedings of the  $42^{nd}$  International Conference on Machine Learning, Vancouver, Canada. PMLR 267, 2025. Copyright 2025 by the author(s).
+
+# 1. Introduction
+
+Video generation has advanced alongside image generation. However, video generation is more challenging due to its higher inherent complexities, including the need for high-quality imagery and seamless temporal smoothness. In addition, controlling the generated content and extension to significant lengths is essential for real-world use. In this paper, we focus on pose-guided video generation with a reference image. Our goal is to generate a video that contains rich imagery details and adheres to the reference image and the pose guidance.
+
+Currently, there are plenty of works focusing on this task, such as DisCo (Wang et al., 2023),AnimateAnyone (Hu, 2024), MagicDance (Chang et al., 2023), MagicAnimate (Xu et al., 2024), etc. Though various techniques have been studied, the results are unsatisfactory in several aspects. Imagery distortion especially on the regions of human hands is still a common issue which is particularly evident in videos containing large movements. Besides, to achieve good temporal smoothness, imagery details are sometimes sacrificed resulting in videos of blurred frames. In the presence of diverse appearances and motions in videos, accurate pose estimation is inherently challenging. This inaccuracy not only creates a conflict between pose alignment and temporal smoothness but also hinders the model scaling on the training schedule due to overfitting on noisy samples. In addition, due to computational limitations and model capabilities, there are still significant challenges in generating high-quality long videos containing a large number of frames. To solve these problems, we propose a series of approaches for generating long but still smooth videos based on pose guidance and image reference.
+
+To alleviate the negative impact of inaccurate pose estimation, we propose an approach of confidence-aware pose guidance. By introducing the concept of confidence to the pose sequence representation, better temporal smoothness can be achieved and imagery distortion can also be eased. Confidence-based regional loss amplification can make the hand regions more accurate and clear. In addition, we propose a progressive latent fusion method for achieving long but still smooth video generation. Through generating video segments with overlapped frames with the proposed pro
+
+gressive latent fusion, our model can handle arbitrary-length pose sequence guidance. By merging the generated video segments, the final long video can be of good cross-frame smoothness and imagery richness at the same time (see Figure 1). For model training, to keep the cost of model training within an acceptable range, our method is based on a generally pre-trained video generation model. The amount of training data is not large and no special manual annotation is required.
+
+In summary, there are three key contributions of this work:
+
+1. We improve the pose guidance by employing a confidence-aware strategy. In this way, we alleviate the negative impact of inaccurate pose estimation. This approach not only reduces the influence of noisy samples during training but also corrects erroneous pose guidance during inference.  
+2. Based on the confidence-aware strategy, we propose hand region enhancement to alleviate hand distortion by strengthening the loss weight of the region of human hands with high pose confidence.  
+3. While overlapped diffusion is a standard technique for generating long videos, we advance it with position-aware progressive latent fusion that improves temporal smoothness at segment boundaries. Extensive experiments show the effectiveness of the proposed approach.
+
+# 2. Related Work
+
+# 2.1. Diffusion Models for Image/Video Generation
+
+Diffusion-based models have demonstrated promising results in the fields of image (Ho et al., 2020; Song et al., 2020; Rombach et al., 2022a; Zhang et al., 2023; Wang et al., 2024a) and video generation (Singer et al., 2022; Guo et al., 2023; Ho et al., 2022; He et al., 2022; Chen et al., 2023; Blattmann et al., 2023), renowned for their capacity in generative tasks. Diffusion models operating in the pixel domain encounter challenges in generating high-resolution images due to information redundancy and high computational costs. Latent Diffusion Models (LDM) (Rombach et al., 2022a) address these issues by performing the diffusion process in low-dimensional latent spaces, significantly enhancing generation efficiency and quality while reducing computational demands. Recently, efforts have been made to improve training efficiency through noise schedule (Zheng et al., 2024a;b). Compared to image generation, video generation demands a more precise understanding of temporal motion patterns. Recent video generation models leverage diffusion models by adding temporal layers to pre-trained image generation models (Singer et al., 2022; Guo et al., 2023; Wu et al., 2023; Wang et al., 2024c), or utilizing DiT (Peebles & Xie, 2023) structures to enhance
+
+generative capabilities for videos (Yan et al., 2021; Yu et al., 2023; Ma et al., 2024a; Bao et al., 2024). Stable Video Diffusion (SVD) (Blattmann et al., 2023) is one of the most popular open-source models built upon LDM. It offers a straightforward and effective method for image-based video generation and serves as a powerful pre-trained model for this task. Our approach extends SVD for pose-guided video generation, leveraging its pre-trained generative capabilities.
+
+# 2.2. Pose-Guided Human Motion Transfer
+
+Pose-to-appearance mapping aims to transfer motion from the source identity to the target identity. Methods based on paired keypoints from source and target images employ local affine transformations (Siarohin et al., 2019; 2021) or Thin-Plate Spline transformations (Zhao & Zhang, 2022) to warp the source image to match the driving image. These techniques aim to minimize distortion by applying weighted affine transformations, thereby generating poses in the output image that closely resemble those in the driving image. Similarly, methods such as (Chan et al., 2019; Tu et al., 2024; Ma et al., 2024b; Hu, 2024; Feng et al., 2023; Peng et al., 2024; Li et al., 2024) utilize pose stick figures obtained from off-the-shelf human pose detectors as motion indicators and directly generate video frames through generative models. Depth information (Feng et al., 2023) or 3D human parametric models, such as SMPL (Skinned Multi-Person Linear) (Zhu et al., 2024; Wang et al., 2024b), can also be used to represent human geometry and motion characteristics from the source video. Nevertheless, these overly dense guidance techniques can rely too much on the signal from the source video, such as the outline of the body, leading to a degradation in the quality of the generated videos, especially when the target identity differs significantly from the source. Our approach, leveraging off-the-shelf human pose detectors, is capable of capturing the motion of the human body in driving videos without introducing excessive extraneous information, thereby ensuring the overall quality of the generated video. In addition, we introduce confidence-aware pose guidance, which effectively mitigates the influence of inaccurate pose estimation in training and inference. In this way, we achieve superior portrait frame quality, especially in the hand regions.
+
+# 2.3. Long video generation
+
+Recent diffusion-based video generation algorithms are constrained to producing videos with durations of only a few seconds, significantly limiting their practical applications. As a result, substantial research efforts have been dedicated to extending the duration of generated videos, leading to the proposal of various approaches to overcome this limitation. Methods like (He et al., 2022; Voleti et al., 2022) autoregressively predict successive frames, enabling the generation of infinitely long videos. However, these methods
+
+often face quality degradation due to error accumulation and the lack of long-term temporal coherence. Hierarchical approach (Yin et al., 2023; Wang et al., 2024c) are proposed for generating long videos in a coarse-to-fine manner. It first creates a coarse storyline with keyframes using a global diffusion model, then iteratively refines the video with local diffusion models to produce intermediate frames.
+
+MultiDiffusion (Bar-Tal et al., 2023) combines multiple processes that use pre-trained text-to-image diffusion models to create high-quality images with user-defined controls. It works by applying the model to different parts of an image and using an optimization method to ensure all parts blend seamlessly. This allows users to generate images that meet specific requirements, like certain aspect ratios or spatial layouts, without needing additional training or fine-tuning. Lumiere (Bar-Tal et al., 2024) extends MultiDiffusion to video generation by dividing the video into overlapping temporal segments. Each segment is independently denoised, and an optimization algorithm then combines these denoised segments. This approach ensures high coherence in the generated video, effectively maintaining temporal smoothness across segments. However, our experiments reveal that abrupt transitions can still occur at segment boundaries.
+
+Building upon this principle, we introduce a position-aware progressive latent fusion strategy that enhances temporal smoothness near segment boundaries. We adaptively assign fusion weight based on the temporal position, ensuring a smooth transition at the segment boundaries.
+
+# 3. Method
+
+# 3.1. Preliminaries
+
+A Diffusion Model (DM) learns a diffusion process that generates a probability distribution for a given dataset. In the case of visual content generation tasks, a neural network of DM is trained to reverse the process of adding noise to real data so new data can be progressively generated starting from random noise. For a data sample  $\mathbf{x} \sim p_{\mathrm{data}}$  from a specific data distribution  $p_{\mathrm{data}}$ , the forward diffusion process is defined as a fixed Markov Chain that gradually adds Gaussian noise to the data, following:  $q(\mathbf{x_t} \mid \mathbf{x}_{t-1}) = \mathcal{N}(\mathbf{x}_t; \sqrt{1 - \beta_t}\mathbf{x}_{t-1}, \beta_t\mathbf{I})$  for  $t = 1, \dots, T$ , where  $T$  is the number of perturbing steps and  $\mathbf{x}_t$  represents noisy data after adding  $t$  steps of noise on the real data  $\mathbf{x}_0$ . This process is controlled by a schedule  $\beta_t$ . The above diffusion process can be reformed as:
+
+$$
+q \left(\mathbf {x} _ {\mathbf {t}} \mid \mathbf {x} _ {0}\right) = \mathcal {N} \left(\mathbf {x} _ {t}; \sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0}, (1 - \bar {\alpha} _ {t}) \mathbf {I}\right) \tag {1}
+$$
+
+where  $\bar{\alpha}_{t} = \prod_{i=1}^{t} \alpha_{i}$  and  $\alpha_{t} = 1 - \beta_{t}$ . Following DDPM (Ho et al., 2020), a denoising function  $\epsilon_{\theta}$  with parameter  $\theta$  is trained by minimizing the mean squared error:
+
+$$
+\mathbb {E} _ {\epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I}), \mathbf {x} _ {t}, \mathbf {c}, t} [ \| \epsilon - \epsilon_ {\theta} (\mathbf {x} _ {t}; \mathbf {c}, t) \| _ {2} ^ {2} ] \tag {2}
+$$
+
+![](images/a761ee732ae6e82883fed18dd696e3c412dfe71c87cf190a79148f44957ef285.jpg)  
+Figure 2: MimicMotion integrates an image-to-video diffusion model with confidence-aware pose guidance, utilizing a spatiotemporal U-Net and a PoseNet for introducing pose sequence as the condition. Our design features adaptive adjustment of pose influence via keypoint confidence scores and prioritizes high-confidence hand regions in the loss function.
+
+where  $\mathbf{c}$  is the condition. We then train  $\epsilon_{\theta}$  until convergence.
+
+# 3.2. Data Preparation
+
+To train a pose-guided video diffusion model, we collect a video dataset containing various human motions. The dataset need not be excessively large, as the pre-trained model already has a good prior.
+
+Given a video from our dataset, the training sample is constructed with three parts: a reference image (denoted as  $I_{\mathrm{ref}}$ ), a sequence of raw video frames, and the corresponding poses. Basic pre-processing operations of resizing and cropping are applied to the raw video to get a sequence of video frames. For a given video, a fixed number of frames are randomly sampled at equal intervals as input video frames to the diffusion model. The input reference image is independently sampled from the video and pre-processed in the same way as the video frames. Another input of the model is the pose sequence, which is extracted from the video frames with DWPose (Yang et al., 2023) frame by frame.
+
+# 3.3. Pose-Guided Video Diffusion Model
+
+The goal of MimicMotion is to generate high-quality, pose-guided human videos from a single reference image and a sequence of poses to mimic. This task involves the synthesis of realistic motion that adheres to the provided pose sequence while maintaining visual fidelity to the reference image. We exploit the ability of a specific pre-trained video diffusion model to reduce the data requirement and computational cost of training a video diffusion model from scratch. Stable Video Diffusion (SVD) (Blattmann et al., 2023) is an open-source image-to-video diffusion model trained on a large-scale video dataset. It shows good perfor
+
+mance on both video quality and diversity compared with the other contemporary models. The model structure of MimicMotion is designed to integrate a pre-trained Stable Video Diffusion (SVD) model to leverage its image-to-video generation capabilities.
+
+Learning a diffusion process in pixel space is costly, especially for high-definition videos involving many frames. We follow the Latent Diffusion Model (LDM) (Rombach et al., 2022b) to encode pixels into a lower-dimensional latent space. LDM adopts a pair of autoencoders, consisting of an encoder  $\mathcal{E}$  and a decoder  $\mathcal{D}$ . Given a data sample  $\mathbf{x}$ , it is encoded into the latent space as  $\mathbf{z} = \mathcal{E}(\mathbf{x})$ . Conversely, the latent vector  $\mathbf{z}$  can be decoded into pixel space via  $\mathbf{x} = \mathcal{D}(\mathbf{z})$ .
+
+Figure 2 shows the structure of our model. The VAE encoder on input video frames and the corresponding decoder for getting denoised video frames are both adopted from SVD and these parameters are frozen. The VAE encoder is applied independently to each frame of the input video and the conditional reference image. In contrast, the VAE decoder processes the latent features that undergo spatiotemporal interaction from U-Net. To enhance the temporal smoothness of low-level details, the VAE decoder incorporates temporal layers alongside the spatial layers.
+
+The reference image and the sequence of poses are two other inputs of the model. The reference image follows two pathways: First, it is processed through CLIP (Radford et al., 2021) to extract features that are fed into each U-Net block's cross-attention. Second, the image is encoded by the frozen VAE encoder to obtain its latent representation, which is then temporally duplicated to match the video frame dimensions. This latent representation is concatenated with the video frame features along the channel dimension before entering the U-Net.
+
+For pose guidance, PoseNet consisting of multiple convolution layers is designed to extract features from input pose sequences. The reason for not using the VAE encoder is that the pixel value distribution of the pose sequence is different from that of common images on which the VAE autoencoder is trained. The features of poses are then element-wisely added to the output of the first convolution layer of U-Net. Pose guidance is not added to every U-Net block due to: a) the sequence pose being extracted without temporal interaction may confuse the spatio-temporal layers within U-Net; b) excessive involvement of the pose sequence may degrade the performance of the pre-trained image-to-video model.
+
+# 3.4. Confidence-Aware Pose Guidance
+
+Inaccurate pose estimation has a negative impact on the model's training and inference. However, accurate pose estimation is challenging in dynamic videos. The limited
+
+capability of the pose estimation model, like DWPose (Yang et al., 2023), is only one aspect of the reason. The more significant cause is the inherent uncertainty of pose from dynamic appearances and motions. Specifically, incorrect pose guidance signals can mislead the model, resulting in the generation of inaccurate or distorted outputs, as illustrated in Figure 7. Moreover, noisy pose guidance signals can lead to overfitting on samples with incorrect poses, potentially causing training instability. This in turn may hinder the model's training.
+
+For this problem, we propose confidence-aware pose guidance, which leverages the confidence scores associated with each keypoint from the pose estimation model. These scores reflect the likelihood of accurate detection, with higher values indicating higher visibility, less self-occlusion from other body parts, and motion blur. Instead of applying a fixed confidence threshold to filter the keypoints, as commonly adopted in prior works (moo, 2024; Chang et al., 2023), we utilize brightness of the keypoints and limb to represent the confidence level of pose estimation. Specifically, we multiply the color assigned to each keypoint and limb by its confidence score. Consequently, keypoints and corresponding limbs with higher confidence scores will appear more significant on the pose guidance map. This method enables the model to prioritize more reliable pose information in its guidance, thereby enhancing the overall accuracy of pose-guided generation. In this way, the uncertainty of pose estimation can be conveyed through the pose guidance, making pose guidance more informative.
+
+Hand region enhancement Moreover, we employ pose estimation and the associated confidence scores to alleviate region-specific artifacts, such as hand distortion, which are prevalent in the diffusion-based image and video generation models. Specifically, we identify reliable regions via thresholding keypoint confidence scores. By setting a threshold, we can distinguish between keypoints that are confidently detected and those that may be unreliable due to self-occlusion or motion blur. Keypoints with confidence scores above the threshold are considered reliable. We implement a masking strategy that generates masks based on a confidence threshold. We unmask areas where confidence scores surpass a predefined threshold, thereby identifying reliable regions. When computing the loss of the video diffusion model, the loss values corresponding to the un-masked regions are amplified by a certain scale so they can have more effect on the model training than other masked regions.
+
+Specifically, to mitigate hand distortion, we compute masks using a confidence threshold for keypoints in the hand region. Only hands with all keypoint confidence scores exceeding this threshold are considered reliable, as a higher score correlates to higher visual quality. We then construct
+
+![](images/f25fa6af7807f5f514300796cb352ffce81a791286c17c431910aab1343c1108.jpg)  
+Figure 3: Overview of our approach for long video generation. The colored boxes represent latent video frames. A darker color means a higher weight. The dashed boxes represent video frame features involved in latent fusion.
+
+a bounding box around the hand by padding the boundary of these keypoints, and the enclosed rectangle is designated as unmasked. This region is subsequently assigned a larger weight in the loss calculation during the training of the video diffusion model. This selective unmasking and weighting process biases the model's learning towards hands, especially hands with higher visual quality, effectively reducing distortion and improving the overall quality.
+
+# 3.5. Progressive Latent Fusion for Long Videos
+
+Limited by computation resources, generating long videos containing a large number of frames is challenging. For this problem, prior works like MultiDiffusion (Bar-Tal et al., 2023) successfully use latent fusion for panoramic image generation. A similar idea can be applied to the video generation task. A straightforward approach is directly applying MultiDiffusion in the time domain, as in Lumiere (Bar-Tal et al., 2024). However, unlike spatial discontinuities between image tiles, temporal discontinuities are more noticeable to viewers, resulting in flickering or abrupt content changes. We address this by a progressive approach for generating long videos with high temporal continuity.
+
+Progressive latent fusion is training-free and is integrated into the denoising process of the latent diffusion model during inference. Figure 3 shows an overview of this process. We omit the VAE for brevity. The denoising process is done in latent space in our method. In general, there are  $T$  denoising steps in total and our latent fusion is applied within each step. For a long given pose sequence, we use a pre-defined strategy for splitting the whole sequence into segments, consisting of a fixed number of frames per segment (denoted as  $N$ ), with a certain number  $(C)$  of overlapped frames between every two adjacent segments. For the sake of generation efficiency, it is common to assume that  $C \ll N$ . During each denoising step, video segments are firstly denoised separately with the trained model, conditioning on the same reference image and the corresponding sub-sequence of poses. Algorithm 1 shows the specific de
+
+Algorithm 1 Progressive latent fusion for long videos.  
+1: Input:  
+2:  $I_{\mathrm{ref}}$ : Reference image;  
+3:  $P_i^j$ : Pose of  $j$ -th frame in  $i$ -th video segment;  
+4:  $\mathbf{z}_i^j$ : The latent of  $j$ -th frame in  $i$ -th video segment;  
+5:  $N$ : The number of frames in each video segment;  
+6:  $C$ : The number of overlapped frames.  
+7: Output: A long sequence of latent video frames.  
+8:  $\mathbf{z} \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$   
+9:  $\lambda_{\mathrm{fusion}} \gets 1 / (C + 1)$   
+10: for  $t = T$  to 1 do  
+11: for  $i = 1, 2, \ldots$  do  
+12:  $\mathbf{z}_i \gets \mathrm{DM}(\mathbf{z}_i, I_{\mathrm{ref}}, P_i, t)$   
+13: end for  
+14: for  $i = 1, 2, \ldots$  do  
+15: for  $j = 1$  to  $N$  do  
+16: if  $i > 1$  and  $j \leq C$  then  
+17:  $\mathbf{z}_i^j \gets j \lambda_{\mathrm{fusion}} \mathbf{z}_i^j + (1 - j \lambda_{\mathrm{fusion}}) \mathbf{z}_{i-1}^{N-C+j}$   
+18: else if  $j > N - C$  then  
+19:  $\mathbf{z}_i^j \gets (N + 1 - j) \lambda_{\mathrm{fusion}} \mathbf{z}_i^j + (C - N + j) \lambda_{\mathrm{fusion}} \mathbf{z}_{i+1}^{C-N+j}$   
+20: end if  
+21: end for  
+22: end for  
+23: end for  
+24: return  $\{\mathbf{z}_i^{1:N-C}\}_{i=1,2,\ldots}$
+
+tails of progressive latent fusion. As inputs, the reference image is denoted as  $I_{\mathrm{ref}}$ , and the pose frame corresponding to  $j$ -th frame in  $i$ -th video segment is denoted as  $P_i^j$ . We use  $\mathbf{z}_i^j$  to denote the latent feature of  $j$ -th frame in  $i$ -th video segment. The denoising process starts from the maximum time step  $T$  and the latent features are initialized with a normal distribution  $\mathcal{N}(\mathbf{0}, \mathbf{I})$ . Within each denoising step at time step  $t$ , the reversed diffusion process defined by the trained model (DM) is applied to the latent features of each video segment numbered  $i$  separately, with  $\mathbf{z}_i$ ,  $I_{\mathrm{ref}}$ ,  $P_i$  and  $t$  as inputs. During the latent fusion stage, for every two adjacent video segments, the involved video frames are then fused. To avoid the corruption of temporal smoothness near video segment boundaries after latent fusion, we propose progressive latent fusion. For a video frame involved in latent fusion, its fusion weight is determined by its relative position in the video segment it belongs. Specifically, if a frame is close to the segment it belongs to, it will be assigned a heavier weight. For implementation, a fusion scale is pre-defined as  $\lambda_{\mathrm{fusion}} = 1 / (C + 1)$  for controlling the level of latent fusion.
+
+After applying  $T$  iterative denoising steps, we aggregate the final long sequence by combining the denoised overlapped video segments in latent space. The resulting aggregated
+
+long video in latent space is then fed to the temporal VAE decoder to reconstruct the long video.
+
+# 4. Experiments
+
+# 4.1. Implementation Details
+
+Due to the lack of suitable high-quality academic datasets, we collect 4,436 human dancing videos from the internet for training. The average length is 20.1s. We adopt the pre-trained weights from SVD (Blattmann et al., 2023). The PoseNet is trained from scratch. We train our model on 8 NVIDIA A100 GPUs for 20 epochs, with a batch size of 8 and 16 frames per clip. The loss weight of the hand region is 10. The learning rate is  $10^{-5}$  with a linear warmup of 500 iterations. We tune all parameters in the UNet and PoseNet.
+
+# 4.2. Comparison to State-of-the-Art Methods
+
+We compare our method with latest state-of-the-art pose-guided human video generation methods, including MagicAnymate (Xu et al., 2024), MagicPose (Chang et al., 2023), Moore-AnymateAnyone (moo, 2024), and MuseV (Xia et al., 2024). For the testing protocol of previous works (Wang et al., 2023; Chang et al., 2023), we adopt the TikTok (Jafarian & Park, 2021) dataset and use sequence 335 to 340 for our evaluation.
+
+We conduct qualitative and quantitative comparisons, complemented by a user study. Each method has a different input aspect ratio. To ensure a fair comparison, we only consider the central square region of the videos. Specifically, we individually apply a center crop to the reference image and pose sequence to match each method's input aspect ratio. Then, we extract the center squares from the generated videos for comparison across different methods.
+
+Qualitative evaluation We conduct qualitative comparisons between the selected baselines and our method. In Figure 4, we showcase sample frames to highlight the superior quality of individual frames produced by our method. Additionally, in Figure 5, we illustrate the temporal differences, demonstrating the enhanced temporal smoothness of our approach compared to existing methods.
+
+Figure 4 presents a comparison of the generated frames, where each row represents a distinct example. The first row demonstrates the superior hand quality achieved by our approach, while the second row showcases the improved adherence to pose guidance. These improvements directly result from our confidence-aware pose guidance and hand region enhancement design.
+
+Importantly, our method shows superior temporal smoothness, characterized by smooth motion and minimal flickering. Figure 5 demonstrates this through pixel-wise
+
+![](images/585445b80438a52457ea93a33f67c3da8eee20a52b14f0bdf9f999851486f15c.jpg)  
+Figure 4: Qualitative comparison to the state-of-the-art methods on TikTok test set. Our method achieves better hand generation quality and adheres better to the reference pose. Note that our method is not trained on the TikTok dataset.
+
+![](images/27bc3acfc4348dcf414cd321147f2f3f1fc813024444d569287b7329fc30f6c2.jpg)  
+Figure 5: Comparison of temporal smoothness via frame differences. MagicPose exhibits abrupt transitions, whereas Moore and MuseV show flickering texture and text. In contrast, our method maintains stable frame differences.
+
+differences between consecutive frames. While Magic-Pose (Chang et al., 2023) shows abrupt transitions, Moore-AnymateAnyone (moo, 2024) exhibits unstable clothing textures, and MuseV (Xia et al., 2024) struggles with consistent text rendering, our method maintains stable frame-to-frame transitions without visible artifacts. Videos in the supplementary material further demonstrate this improvement. This enhancement is enabled by our confidence-aware pose guidance, which mitigates the impact of inaccurate pose inputs and the associated temporal noise. By weighting pose signals by confidence, our method ensures high temporal smoothness in the presence of pose guidance noise.
+
+Quantitative evaluation Our method outperforms existing approaches across all evaluation metrics, as shown in Table 1. We evaluated performance on test sequences from the TikTok dataset (Jafarian & Park, 2021) using both video-level metrics FID-VID (Balaji et al., 2019) and FVD (Unterthiner et al., 2018), as well as frame-level metrics PSNR and SSIM (Wang et al., 2004) following (Wang et al., 2023).
+
+<table><tr><td>Method</td><td>FID-VID↓</td><td>FVD↓</td><td>SSIM↑</td><td>PSNR↑</td></tr><tr><td>MagicAnymate</td><td>16.2</td><td>848</td><td>0.740</td><td>17.5</td></tr><tr><td>MagicPose</td><td>13.3</td><td>916</td><td>0.776</td><td>18.8</td></tr><tr><td>Moore</td><td>12.4</td><td>728</td><td>0.758</td><td>18.7</td></tr><tr><td>MuseV</td><td>14.6</td><td>754</td><td>0.766</td><td>17.6</td></tr><tr><td>MimicMotion (ours)</td><td>9.3</td><td>594</td><td>0.795</td><td>20.1</td></tr></table>
+
+![](images/1400efdb2edd84681571215f10068b4bc9f36e7b172af2582a8ca8c30af78754.jpg)  
+Figure 6: Preference of videos generated by MimicMotion (ours) over baseline methods on the TikTok dataset test split. Users consistently prefer MimicMotion over other methods.
+
+The results demonstrate our method's superior performance compared to current state-of-the-art approaches.
+
+User study To supplement our quantitative and qualitative evaluations, we conduct a user study to assess the subjective preferences of participants regarding the generated videos on the TikTok dataset test split. The study involves showing two video clips—one generated by our method and the other by one of the baseline methods—to a diverse group of users. Participants are instructed to select the video that they perceived as having higher quality, considering factors including image quality and temporal smoothness of characters and clothing. We collected data from 36 participants, with each participant evaluating 6 video pairs for our method against each baseline method. As shown in Figure 6, the results indicate a strong preference for MimicMotion over the baseline methods. In comparison to MagicPose and Moore, the participants almost favored all videos produced by our method. Despite MuseV showing higher image quality compared to other baselines, the preference for videos produced by our method still reached  $75.5\%$ . These findings align with our qualitative and quantitative evaluation, reinforcing the effectiveness of our method in meeting user expectations for high-quality human video generation.
+
+# 4.3. Ablation Study
+
+As shown in Table 2, confidence-aware pose guiding and hand region enhancement improve all metrics. Progressive latent fusion notably improves FVD scores, indicating better temporal coherence, though other metrics may not reflect
+
+Table 1: Quantitative comparison to state-of-the-art methods: MagicAnymate (Xu et al., 2024), MagicPose (Chang et al., 2023), Moore (moo, 2024), MuseV (Xia et al., 2024). We evaluate on the TikTok dataset test split. MimicMotion is the best in all metrics.  
+
+<table><tr><td>Hand</td><td>Conf.</td><td>Prog.</td><td>FID-VID↓</td><td>FVD↓</td><td>SSIM↑</td><td>PSNR↑</td></tr><tr><td></td><td></td><td></td><td>14.6</td><td>776</td><td>0.760</td><td>18.0</td></tr><tr><td></td><td></td><td>✓</td><td>15.0</td><td>678</td><td>0.758</td><td>17.9</td></tr><tr><td></td><td>✓</td><td>✓</td><td>12.2</td><td>623</td><td>0.787</td><td>18.4</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>9.3</td><td>594</td><td>0.795</td><td>20.1</td></tr></table>
+
+Table 2: Ablation studies of hand region augmentation (hand), confidence-aware pose guiding (conf.), and progressive latent fusion (prog.).
+
+this improvement. We analyze each component below.
+
+Confidence-aware pose guiding Figure 7 shows the effectiveness of confidence-aware pose guidance. Each row corresponds to one sample. The left side shows three images to extract the pose. On the right side, we plot the corresponding pose guiding signals, both with and without confidence-aware pose guiding. From the guiding signals, we can see that there are errors in the pose estimated by DW- Pose. Nevertheless, our confidence-aware design minimizes the impact of incorrect pose estimation in guidance signals.
+
+Specifically, in the case of Pose 1, the estimation exhibits a duplicate detection issue, which leads to the inclusion of duplicate keypoints. In the case of Pose 2, there is one hand obscured, and the keypoints of this hand are incorrectly estimated on the other hand; In the case of Pose 3, the right elbow is obscured, but it is still detected with confidence above the threshold thus falsely remains in the guidance signal. These problems lead to confusing hand guidance signals and ultimately lead to distortions such as deformed hands or wrong spatial relationships in the generated frames.
+
+In contrast, by integrating confidence scores into pose guidance, our method effectively mitigates these issues. The confidence scores provide a measure of reliability for each keypoint, allowing our model to weigh the guidance signals accordingly. Specifically, keypoint with lower confidence, which typically correspond to inaccurate keypoints caused by self-occlusion or motion blur, will be of less significance in the guidance. This approach leads to clearer and richer pose guidance, as the influence of potentially erroneous keypoints is reduced. The corresponding generation results demonstrate how our method enhances the robustness of generation against false guiding signals (Pose 1 and Pose 2) and offers visibility hints to resolve the front-back ambiguity of 2D pose estimation (Pose 3).
+
+Hand region enhancement In conjunction with confidence-aware pose guiding, we further improve hand generation quality by assigning a higher weight to the hand region in the training loss. Figure 8 compares the generation result with and without hand region enhancement, using the same reference image and pose guidance. All experiments
+
+![](images/42fee89479cfe40139d2fd943a8e556d277fc145ace02b12f22fd0049236b312.jpg)  
+Figure 7: Confidence-aware pose guiding enhances generation robustness to false guiding signals (Pose 1&2) and provides hints to handle self-occlusion (Pose 3).
+
+![](images/c159c5a6b82939f889572d74f86dd933afa618ea2d6626240289b67910c69d6c.jpg)  
+Figure 8: Hand region enhancement consistently reduces hand distortion and improves visual appeal, using the same reference image and pose guidance.
+
+incorporate confidence-aware pose guidance. The hands in the first row are cropped from the generated video frames of a model trained without hand region enhancement, which exhibits noticeable distortions, such as irregular and misplaced fingers. In contrast, the results of the model trained with hand region enhancement (the second row) show consistent improvements in hand generation quality and a reduction in hand distortion. These results show the effectiveness of the proposed hand region enhancement design, which substantially mitigates hand distortion, which is a prevalent challenge in diffusion-based models. As humans often focus on hands, this design aligns the training process with human preferences, thereby enhancing the overall visual quality of the results. There is no noticeable degradation in the quality of the remaining parts.
+
+Progressive latent fusion To achieve seamless transitions between video segments, we introduce progressive latent fusion, a technique that gradually blends frames in the overlapped regions of consecutive video segments. Figure 9 presents a comparison of a sample transition with an overlap of 6 frames between two 16-frame segments. We visualize 4 sample frames. The original MultiDiffusion approach
+
+![](images/b7f0f31dd40c62b140b4f94b3189b9d63f22637e5f75416b07fe4e93862a1e1c.jpg)  
+Figure 9: Progressive latent fusion ensures smooth segment transitions by avoiding temporal discontinuities of weights in simple averaging, which reduces artifacts in Y-T slices.
+
+employs a simple averaging of frames within the overlap region. This method assigns equal weight to all frames in the overlap region, irrespective of their temporal position (whether they are closer to the preceding or subsequent segment). It lacks a gradual transition from one segment to another, causing noticeable flickers at denoising window boundaries. This is evident in the y-t slice shown on the left of Figure 9a, where the transition at segment boundaries is abrupt. The right side of the figure shows four frames at the segment boundary. Note that the background in the top-left corner (enlarged) is initially clear in segment 1. It suddenly becomes blurry in the overlapped region and then suddenly reverts to clearer in the main part of segment 2. This artifact is not observed when progressive latent fusion is applied.
+
+The proposed progressive latent fusion approach (see Figure 9b) mitigates these issues. The left y-t slice shows smooth transitions across segment boundaries, avoiding abrupt changes. The right side of the figure demonstrates that this approach eliminates a sudden blurring, mitigating flicking artifacts, thus improving the overall visual temporal coherence for long video generation.
+
+# 5. Conclusion
+
+In this study, we introduce MimicMotion, a pose-guided human video generation model that leverages confidence-aware pose guidance and progressive latent fusion for producing high-quality, long videos with human motion guided by pose. Through extensive experiments and ablation studies, we show that our model achieves superior adaptation to noisy pose estimation, enhancing hand quality and ensuring temporal smoothness. The integration of confidence scores into pose guidance, the enhancement of hand region loss, and the implementation of progressive latent fusion are crucial in achieving these improvements, resulting in more visually compelling and realistic human video generation.
+
+# Impact Statement
+
+MimicMotion advances human motion video generation technology with potential applications in creative content production, digital human, and entertainment. The framework enables the creation of high-quality, controllable human motion videos. However, like other video generation technologies, there are potential risks of misuse for creating deceptive content that could be used for misinformation. Responsible development and deployment of such technology should include appropriate safeguards, such as watermarking, detection methods, and clear usage guidelines. We encourage the development of technical capabilities and ethical frameworks for video generation technologies.
+
+# References
+
+MooreThreads/Moore-AnimateAnyone, May 2024. URL https://github.com/MooreThreads/Moore-AnimateAnyone. original-date: 2024-01-12T07:55:21Z.  
+Balaji, Y., Min, M. R., Bai, B., Chellappa, R., and Graf, H. P. Conditional gan with discriminative filter generation for text-to-video synthesis. In *IJCAI*, volume 1, pp. 2, 2019.  
+Bao, F., Xiang, C., Yue, G., He, G., Zhu, H., Zheng, K., Zhao, M., Liu, S., Wang, Y., and Zhu, J. Vidu: a highly consistent, dynamic and skilled text-to-video generator with diffusion models. arXiv preprint arXiv:2405.04233, 2024.  
+Bar-Tal, O., Yariv, L., Lipman, Y., and Dekel, T. Multi-diffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023.  
+Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Li, Y., Michaeli, T., et al. Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024.  
+Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023.  
+Chan, C., Ginosar, S., Zhou, T., and Efros, A. A. Everybody dance now. In Proceedings of the IEEE/CVF international conference on computer vision, pp. 5933-5942, 2019.  
+Chang, D., Shi, Y., Gao, Q., Xu, H., Fu, J., Song, G., Yan, Q., Zhu, Y., Yang, X., and Soleymani, M. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In *Forty-first International Conference on Machine Learning*, 2023.
+
+Chen, H., Xia, M., He, Y., Zhang, Y., Cun, X., Yang, S., Xing, J., Liu, Y., Chen, Q., Wang, X., et al. Videocaster1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023.  
+Chen, H., Zhang, Y., Cun, X., Xia, M., Wang, X., Weng, C., and Shan, Y. Videocaster2: Overcoming data limitations for high-quality video diffusion models. arXiv preprint arXiv:2401.09047, 2024.  
+Feng, M., Liu, J., Yu, K., Yao, Y., Hui, Z., Guo, X., Lin, X., Xue, H., Shi, C., Li, X., et al. Dreamoving: A human video generation framework based on diffusion models, 2023.  
+Guo, Y., Yang, C., Rao, A., Wang, Y., Qiao, Y., Lin, D., and Dai, B. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023.  
+He, Y., Yang, T., Zhang, Y., Shan, Y., and Chen, Q. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022.  
+Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020.  
+Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., and Fleet, D. J. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022.  
+Hu, L. Animate anyone: Consistent and controllable imaged-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8153-8163, 2024.  
+Jafarian, Y. and Park, H. S. Learning high fidelity depths of dressed humans by watching social media dance videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 12753-12762, June 2021.  
+Karras, T., Aittala, M., Aila, T., and Laine, S. Elucidating the Design Space of Diffusion-Based Generative Models, October 2022. URL http://arxiv.org/abs/2206.00364.arXiv:2206.00364 [cs, stat].  
+Li, H., Li, Y., Yang, Y., Cao, J., Zhu, Z., Cheng, X., and Long, C. Dispose: Disentangling pose guidance for controllable human image animation. arXiv preprint arXiv:2412.09349, 2024.  
+Ma, X., Wang, Y., Jia, G., Chen, X., Liu, Z., Li, Y.-F., Chen, C., and Qiao, Y. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024a.
+
+Ma, Y., He, Y., Cun, X., Wang, X., Chen, S., Li, X., and Chen, Q. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp. 4117-4125, 2024b.  
+Peebles, W. and Xie, S. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pp. 4195-4205, 2023.  
+Peng, B., Wang, J., Zhang, Y., Li, W., Yang, M.-C., and Jia, J. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024.  
+Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp. 8748-8763. PMLR, 2021.  
+Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 10684-10695, 2022a.  
+Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10684-10695, 2022b.  
+Siarohin, A., Lathuilière, S., Tulyakov, S., Ricci, E., and Sebe, N. First order motion model for image animation. Advances in neural information processing systems, 32, 2019.  
+Siarohin, A., Woodford, O. J., Ren, J., Chai, M., and Tulyakov, S. Motion representations for articulated animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13653-13662, 2021.  
+Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.  
+Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020.  
+Tu, S., Xing, Z., Han, X., Cheng, Z.-Q., Dai, Q., Luo, C., and Wu, Z. Stableimator: High-quality identity-preserving human image animation. arXiv preprint arXiv:2411.17697, 2024.
+
+Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., and Gelly, S. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018.  
+Voleti, V., Jolicoeur-Martineau, A., and Pal, C. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in neural information processing systems, 35:23371-23385, 2022.  
+Wang, Q., Bai, X., Wang, H., Qin, Z., and Chen, A. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024a.  
+Wang, Q., Jiang, Z., Xu, C., Zhang, J., Wang, Y., Zhang, X., Cao, Y., Cao, W., Wang, C., and Fu, Y. Vividpose: Advancing stable video diffusion for realistic human image animation. arXiv preprint arXiv:2405.18156v1, 2024b.  
+Wang, T., Li, L., Lin, K., Zhai, Y., Lin, C.-C., Yang, Z., Zhang, H., Liu, Z., and Wang, L. Disco: Disentangled control for realistic human dance generation. arXiv preprint arXiv:2307.00040, 2023.  
+Wang, W., Liu, J., Lin, Z., Yan, J., Chen, S., Low, C., Hoang, T., Wu, J., Liew, J. H., Yan, H., et al. Magicvideo v2: Multi-stage high-aesthetic video generation. arXiv preprint arXiv:2401.04468, 2024c.  
+Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004.  
+Wu, J. Z., Ge, Y., Wang, X., Lei, S. W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., and Shou, M. Z. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 7623-7633, 2023.  
+Xia, Z., Chen, Z., Wu, B., Li, C., Hung, K.-W., Zhan, C., He, Y., and Zhou, W. Musev: Infinite-length and high fidelity virtual human video generation with visual conditioned parallel denoising. arxiv, 2024.  
+Xu, Z., Zhang, J., Liew, J. H., Yan, H., Liu, J.-W., Zhang, C., Feng, J., and Shou, M. Z. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1481-1490, 2024.  
+Yan, W., Zhang, Y., Abbeel, P., and Srinivas, A. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021.
+
+Yang, Z., Zeng, A., Yuan, C., and Li, Y. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 4210-4220, 2023.  
+Yin, S., Wu, C., Yang, H., Wang, J., Wang, X., Ni, M., Yang, Z., Li, L., Liu, S., Yang, F., et al. Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346, 2023.  
+Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A. G., Yang, M.-H., Hao, Y., Essa, I., et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10459-10469, 2023.  
+Zhang, L., Rao, A., and Agrawala, M. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 3836-3847, 2023.  
+Zhao, J. and Zhang, H. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 3657-3666, 2022.  
+Zheng, T., Geng, C., Jiang, P., Wan, B., Zhang, H., Chen, J., Wang, J., and Li, B. Non-uniform timestep sampling: Towards faster diffusion model training. In Cai, J., Kankanhalli, M. S., Prabhakaran, B., Boll, S., Subramanian, R., Zheng, L., Singh, V. K., Cesar, P., Xie, L., and Xu, D. (eds.), Proceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 - 1 November 2024, pp. 7036-7045. ACM, 2024a.  
+Zheng, T., Jiang, P., Wan, B., Zhang, H., Chen, J., Wang, J., and Li, B. Beta-tuned timestep diffusion model. In Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., and Varol, G. (eds.), Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part III, volume 15061 of Lecture Notes in Computer Science, pp. 114-130. Springer, 2024b.  
+Zhu, S., Chen, J. L., Dai, Z., Xu, Y., Cao, X., Yao, Y., Zhu, H., and Zhu, S. Champ: Controllable and consistent human image animation with 3d parametric guidance, 2024.
+
+# A. Training Details
+
+MimicMotion is tuned from Stable Video Diffusion (Blattmann et al., 2023). The video samples are resized to the resolution of  $576 \times 1024$ . The training dataset contains 4,436 human dancing videos collected from the internet. We train the model for 20 epochs at a batch size of 8. We follow Stable video diffusion and adopt the noise distribution, i.e.  $\log \sigma \sim \mathcal{N}(P_{\mathrm{mean}}, P_{\mathrm{std}}^2)$ , proposed by Karras et al (Karras et al., 2022) with parameter  $P_{\mathrm{mean}} = 0.5$  and  $P_{\mathrm{std}} = 1.4$ . We train our model on 8 NVIDIA A100 GPUs with a batch size of 8 and 16 frames per clip. The loss weight assigned to the hand region is set at 10, while the remaining part is left to 1. The learning rate is  $10^{-5}$  with a linear warmup schedule of 500 iterations. We tune all parameters in the UNet and PoseNet.
+
+# B. PoseNet Model Architecture
+
+The PoseNet consists of convolutional and activation layers. The detailed architecture is listed in Table 3.
+
+Table 3: Detailed Architecture of PoseNet  
+
+<table><tr><td>#</td><td>Type</td><td>Channels</td><td>Kernel size</td><td>Stride</td><td>Pad</td></tr><tr><td>1</td><td>Input</td><td>3</td><td>-</td><td>-</td><td>-</td></tr><tr><td>2</td><td>Conv2d</td><td>3</td><td>3x3</td><td>1</td><td>1</td></tr><tr><td>3</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>4</td><td>Conv2d</td><td>16</td><td>4x4</td><td>2</td><td>1</td></tr><tr><td>5</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>6</td><td>Conv2d</td><td>16</td><td>3x3</td><td>1</td><td>1</td></tr><tr><td>7</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>8</td><td>Conv2d</td><td>32</td><td>4x4</td><td>2</td><td>1</td></tr><tr><td>9</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>10</td><td>Conv2d</td><td>32</td><td>3x3</td><td>1</td><td>1</td></tr><tr><td>11</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>12</td><td>Conv2d</td><td>64</td><td>4x4</td><td>2</td><td>1</td></tr><tr><td>13</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>14</td><td>Conv2d</td><td>64</td><td>3x3</td><td>1</td><td>1</td></tr><tr><td>15</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>16</td><td>Conv2d</td><td>128</td><td>3x3</td><td>1</td><td>1</td></tr><tr><td>17</td><td>SiLU</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>18</td><td>Conv2d</td><td>320</td><td>1x1</td><td>1</td><td>0</td></tr></table>
+
+# C. Additional Details on User Study
+
+The user study was designed to evaluate the subjective preferences of participants regarding the quality of videos generated by our method, MimicMotion, compared to several baseline methods on the TikTok dataset test split. Participants were instructed to select the video they perceived as having higher quality. The data collection process was conducted through an online interface, as depicted in Figure 10, which allowed participants to easily view and compare the video clips and submit their preferences. The results of our method will randomly appear in the video on the left or the right, with the other video being the comparative method.
+
+# D. Cross-Domain Results: Cartoon and Animal Videos
+
+To demonstrate the versatility of MimicMotion, we evaluate its performance on cross-domain scenarios, including 4 cartoon and 4 animal dancing videos. The videos are available on our anonymous project page. As shown in Figure 11, the model generates cartoon dancing videos while maintaining the artistic style of the characters. For animal subjects (Figure 12), our method produces reasonable results despite their significantly different appearance from humans.
+
+These results highlight MimicMotion's generalizability, which stems from two key factors: first, the pose feature space of humans, cartoons, and animals may share a common subspace that is preserved in our training, human control signals can also guide cartoons and animals. Second, we map the limb lengths of the pose template to the reference character following MuseV (Xia et al., 2024), making the pose guidance signal closer to the reference character. This helps the model maintain the character's appearance while adapting poses across varying body proportions.
+
+Play all
+
+Figure 10: Snapshot of the user preference selection interface that prompts users to evaluate image quality and temporal smoothness. Note that the two videos play automatically, with our method randomly appearing on one side.  
+![](images/f7efef5875d9b396a0e0145f15e688ccb8f18a0e35598b16e3318537f6b7d124.jpg)  
+Please pay attention to the image quality, whether there is flickering, the temporal consistency of the characters and clothing, and select the higher quality video (you can rate the video after it finishes playing).
+
+Left Right
+
+![](images/67ec9036ebe99358a14dce64d98e027a513b114bdfb968a2f472cd3d9799f677.jpg)
+
+![](images/0931b0f9bb61c8436a65bab1db7b445a002d0ec9a48573aac016077c4983e3e1.jpg)  
+Reference images
+
+![](images/cd1e595574d8de41b001de6a7cbf8fc48b85b3c0caea2b790cb56ed0071df2ff.jpg)
+
+![](images/0a4785c0ce9561c58295a2bd6a52c4f38e855066593022e2eb078a3cbced4cf5.jpg)
+
+![](images/b84936adffe1390da1e0cca9f282e46ad116e1d6edc2ba459ab9d7e78b216a89.jpg)  
+Reference pose
+
+![](images/9f178ebd4d7a98933e636ea89a24bacba36ed33cd474f8241bb0d5506be62da3.jpg)
+
+![](images/840e441f1373437bafcfaf9c1ec9ef64036cf91c2ff08c5057db021d28fa0bc9.jpg)
+
+![](images/6953977940b6372ef66a5dd2ac23ce63dfe69215cf592b0dcdff4352e8c5c235.jpg)
+
+![](images/916eb2ad69576d807d969c40703861dbaacb607065f0979f7738294fdac48970.jpg)
+
+![](images/298fcd3e3041b08a461a2e3b3d0195ac3124b36164c6d1222eed9abd5e83a282.jpg)  
+Figure 11: Generated cartoon dancing videos. Our method successfully preserves the artistic style of cartoon characters while generating natural motions that follow the input poses. Top: reference frame; Left: reference pose guidance; Bottom: generated frames at different timestamps.
+
+![](images/39940397f7c77b65c14a842d75c4e25df342ebb39cb8f1a8557dc253dbd14af9.jpg)
+
+![](images/b799e14f26eb31aacd56159c50d967baf8d08261669c730972c523a26fca05c2.jpg)
+
+![](images/f0e73a331087383708f4f3c71a4d6e4d3ca01faf9dd90e5594548ba9ee562deb.jpg)
+
+![](images/b83746673bc3cfc89c689e80b43b0ab0c027ffb013f3d6c06ebfbffe5a639ae6.jpg)
+
+![](images/7f88303698c2b7b00706038d71e3a843c0a5397c2cda077085ab993d322ea6c7.jpg)  
+Reference images
+
+![](images/ae48b8d5ca551bd42ee5c18ba805787666292cba97036318da8fd30549b581c1.jpg)  
+Reference pose
+
+![](images/c23541ea3b05a15428e704a190a7efb841023adda176684006ddff03dc49aa5c.jpg)
+
+![](images/3be0c670674894f95ec83556ce7c6f7d9acee0bf715f9618b6fa0729a0ac201d.jpg)
+
+![](images/b63ac8a98ef0a6b29882054afbebf8251684b70a668eb661611c984b9cd3ec81.jpg)
+
+![](images/03cf025c93bb9f7a96b079b309c6728bb8132b87e1cdce3055e750adf9de9b67.jpg)
+
+![](images/7bcaca0935a8bfac1d8cb3e7215ab0089f9a1f25b5a0dba6efa6ade3bad0f391.jpg)  
+Figure 12: Generated animal dancing videos. MimicMotion can generalize to animal dancing zero-shot, producing plausible motions despite the significant appearance differences from humans. Top: reference frame; Left: reference pose guidance; Bottom: generated frames at different timestamps.
+
+![](images/0b8a5eaacfca497ec4cedcabafda5330880c8a3778546479f5c3aae08be73f6e.jpg)
+
+![](images/3c865da86ee1080c8f35f507d29cf695a11c86cc6a0fb6430db13c06dd94a676.jpg)
+
+![](images/280f92deb0f07a582eca28e5b921e0f85170d66add85e36a4d04d73b779f1343.jpg)
+
+![](images/1d536d6cef967e4d9da0ce1c7d88cbd5e371e660efc33d807979397e286bee80.jpg)

2025/MimicMotion_ High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f60a20bde82e31d51879725dd1cb474a12ae8eff53bb7169eedeababf1b09efd
+size 832778

2025/Mind Your Step (by Step)_ Chain-of-Thought can Reduce Performance on Tasks where Thinking Makes Humans Worse/c5ac481f-06ce-428c-a6a1-af450147d7f0_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5be639d17472b687fd97a1480c3d16b76431a094b8544f89c76b0b788c0ae81f
+size 5858934

2025/Mind Your Step (by Step)_ Chain-of-Thought can Reduce Performance on Tasks where Thinking Makes Humans Worse/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd650c0c62f2b9d001a541e60bf10b94b32cf8249773a4c29b3de871cf8103a4
+size 798385

2025/Mind the Gap_ A Practical Attack on GGUF Quantization/36c60854-11ea-47c6-bfdb-b6ef74c79b0d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b027190aafab2e0d4b6a988f1952fb40208c3b6670394c5dd42783bab061ab4c
+size 671339

2025/Mind the Gap_ A Practical Attack on GGUF Quantization/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a6414bec64c5f85beff0801bd74f6ea8f5a2cd3ef1fdf8977b702b0e612ae72
+size 1749089

2025/Mind the Gap_ a Spectral Analysis of Rank Collapse and Signal Propagation in Attention Layers/bc91cb4b-3713-4802-95f4-a813a601e7f4_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e3e683a24f1988e9e078d334aa95ed7bc998745d530eb6f1d18d8896d56ae0e
+size 1497831

2025/Mind the Gap_ a Spectral Analysis of Rank Collapse and Signal Propagation in Attention Layers/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6001546a936adca0595452caeb296379f2c6d26af841c340c92bf39b4d5db014
+size 1228590

2025/MindAligner_ Explicit Brain Functional Alignment for Cross-Subject Visual Decoding from Limited fMRI Data/9c365be7-b2b5-41f9-a434-a3131c723e3b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca6858fba1293f7ae5d72877e63f34daba164c5fb7d1df643816113e7f8667a6
+size 10024684

2025/MindAligner_ Explicit Brain Functional Alignment for Cross-Subject Visual Decoding from Limited fMRI Data/full.md

@@ -0,0 +1,356 @@
+# MindAligner: Explicit Brain Functional Alignment for Cross-Subject Visual Decoding from Limited fMRI Data
+
+Yuqin Dai\*2,1,4 Zhouheng Yao\*1 Chunfeng Song\* Qihao Zheng\* Weijian Mai\* Kunyu Peng5 Shuai Lu4 Wanli Ouyang1,3 Jian Yang\*2 Jiamin Wu\*1,3
+
+https://github.com/Dalyuqin/MindAligner
+
+# Abstract
+
+Brain decoding aims to reconstruct visual perception of human subject from fMRI signals, which is crucial for understanding brain's perception mechanisms. Existing methods are confined to the single-subject paradigm due to substantial brain variability, which leads to weak generalization across individuals and incurs high training costs, exacerbated by limited availability of fMRI data. To address these challenges, we propose MindAligner, an explicit functional alignment framework for cross-subject brain decoding from limited fMRI data. The proposed MindAligner enjoys several merits. First, we learn a Brain Transfer Matrix (BTM) that projects the brain signals of an arbitrary new subject to one of the known subjects, enabling seamless use of pretrained decoding models. Second, to facilitate reliable BTM learning, a Brain Functional Alignment module is proposed to perform soft cross-subject brain alignment under different visual stimuli with a multi-level brain alignment loss, uncovering fine-grained functional correspondences with high interpretability. Experiments indicate that MindAligner not only outperforms existing methods in visual decoding under data-limited conditions, but also provides valuable neuroscience insights in cross-subject functional analysis.
+
+# 1. Introduction
+
+The brain serves as the center of human cognition and unraveling its underlying mechanisms holds profound aca
+
+*Equal contribution. 1Shanghai Artificial Intelligence Laboratory 2Nanjing University of Science and Technology 3The Chinese University of Hong Kong 4Tsinghua University 5Karlsruhe Institut für Technologie. Correspondence to: Jian Yang <csjyang@njust.edu.cn>, Jiamin Wu <jiaminwu@cuhk.edu.hk>.
+
+Proceedings of the  $42^{nd}$  International Conference on Machine Learning, Vancouver, Canada. PMLR 267, 2025. Copyright 2025 by the author(s).
+
+demic significance (Naselaris et al., 2011). To investigate the brain's perceptual mechanisms, functional magnetic resonance imaging (fMRI) (Naselaris et al., 2011) has been widely used due to its noninvasive acquisition and precise localization of the functional regions. The advances in fMRI facilitate the research on brain visual decoding, which aims to recover visual stimuli seen by humans from their brain activity, contributing to the progress of cognitive science research and Brain-Computer Interfaces (BCI) (Qian et al., 2020; Horikawa & Kamitani, 2017).
+
+Despite the success in fMRI-based visual decoding, the majority methods (Seeliger et al., 2018; Lu et al., 2023; Ozcelik & VanRullen, 2023) are confined to the less practical single-subject paradigm, where a customized decoding model is trained for each person subject. Due to substantial brain differences among subjects, the decoding model trained on one subject cannot be effectively transferred to others, limiting its practicality in BCI and clinical applications. In fact, variations in individual cognitive patterns and brain structures, result in significant fMRI differences (Naselaris et al., 2011). Moreover, the high acquisition cost of fMRI limits the data availability for new subjects. Thus, adapting brain decoding models to new subjects with limited data is crucial.
+
+To address the cross-subject issue, several methods (Scotti et al., 2024b; Li et al., 2025; Wang et al., 2024) adopt brain alignment techniques in an implicit manner. They align fMRI signals from different subjects to a latent space that is assumed to capture common cognition patterns across subjects by learning subject-specific parameters. However, this implicit alignment approach has two limitations. (1) Insufficient brain alignment: learning a shared space that effectively aligns all subjects remains challenging due to noisy and limited fMRI data. As individuals have vast brain differences even when viewing identical stimuli, enforcing all subjects to be simultaneously aligned to a single latent space is prone to suboptimal alignment and compromised representation. Even with extensive multi-subject fMRI training, the shared latent space shows limited generalizability to unseen subjects. For instance, Unibrain (Lei et al., 2023) shows  $50\%$  performance drop when transferring the
+
+![](images/c45fbd195bab091d02ebea7ab87813a2a9b5aa5fc17a706a91f986708eddc284.jpg)  
+Figure 1. Different approaches to functional alignment in brain decoding: Prior works (Scotti et al., 2024b; Li et al., 2025) adopt implicit alignment approach that aligns all subjects into a single latent space, which may lead to suboptimal alignment. Differently, MindAligner employs an explicit alignment strategy, mapping novel subject signals to seen ones by establishing fine-grained functional correspondences. MindAligner not only enables high-quality visual reconstruction from fMRI signals but also facilitates brain functional analysis across subjects.
+
+decoding model to new subjects. (2) Lack of functional interpretability: existing latent alignment methods (Scotti et al., 2024b; Wang et al., 2024; Li et al., 2025) fail to explicitly account for cognitive pattern relationships between subjects. Their alignment is incapable of localizing the brain regions for functional differences and commonalities. This lack of interpretability not only limits cross-subject knowledge transfer in new subject adaptation, but also hinders analysis of neural functional mechanisms underlying human perception process. Given these limitations, an important question arises: can we create a brain alignment framework enabling effective new-subject adaptation and brain functional analysis under data scarcity?
+
+To answer this question, our motivation is to establish an explicit brain functional alignment framework that maps the novel subject's signal to a seen subject's signal. Given an arbitrary new subject, our explicit alignment can establish fine-grained functional correspondences between the new subject and seen subjects in the original brain space, as shown in Fig. 1. The aligned fMRI signal can not only seamlessly be integrated into the pre-trained decoding model of seen subjects but also reveals brain region-level cross-subject variability. However, achieving such brain alignment is challenging, as it requires paired fMRI from subjects performing the same task (i.e., viewing identical visual stimuli (Bazeille et al., 2021)), a condition not met by the existing dataset (Allen et al., 2022).
+
+Based on the above observations, we propose MindAligner, an explicit functional alignment framework for cross-subject visual decoding with limited fMRI data. The core of our method is to train a cross-subject Brain Transfer Matrix (BTM) that projects the brain signals of a new subject to
+
+one of the known subjects in the voxel-level. To overcome the lack of strictly paired fMRI signals, we propose a Brain Functional Alignment module (BFA) to perform soft cross-subject alignment between fMRI signals from different but similar visual stimuli, facilitating the mapping of functionally equivalent cortical areas. Specifically, BFA first decomposes the brain transfer matrix into two low-rank linear layers, enhancing parameter efficiency to facilitate effective adaptation with limited data. In the latent space of the BTM, a cross-stimulus neural mapper is designed to transform the fMRI under different visual stimuli, with stimulus differences as mapping condition. To achieve sufficient and fine-grained alignment, we design a multi-level brain alignment losses that incorporates a signal-level reconstruction loss and a latent alignment loss guided by visual semantic similarities. In this way, the resulting brain transfer matrix not only facilitates fine-grained alignment without shared stimuli constraint, but also encodes cross-subject brain relations for enhanced functional interpretability and neuroscience analysis.
+
+In summary, our contributions are as follows:
+
+- We propose MindAligner, the first explicit brain alignment framework that enables cross-subject visual decoding and brain functional analysis in the data-limited setting.  
+- We propose a brain transfer matrix to establish fine-grained functional correspondences between arbitrary subjects. This matrix is optimized through a brain functional alignment module, which employs a multi-level alignment loss to enable soft cross-subject mapping.  
+- Experiments demonstrate that MindAligner outper
+
+forms state-of-the-art methods in visual decoding with only  $6\%$  of the whole model's parameters learned.
+
+- We conduct cross-subject brain functional visualization and discover that the early visual cortex shows similar activities across subjects, while the higher visual cortex related to memory and spatial navigation exhibits significant inter-subject variability.
+
+# 2. Related Work
+
+# 2.1. fMRI-Based Brain Decoding
+
+Brain decoding seeks to reconstruct the visual stimuli perceived by subjects based on their brain activity, offering a deeper understanding of the brain's mechanisms for processing external information (Naselaris et al., 2011). Earlier work (Horikawa & Kamitani, 2017) reveals a correlation between Deep Neural Networks (DNNs) image representations and neural activity in the visual cortex using sparse linear regression. With the advent of generative models (Goodfellow et al., 2020; Ho et al., 2020) and extensive fMRI datasets (Allen et al., 2022), visual decoding has shifted towards mapping brain signals to the latent spaces of large models, facilitating the reconstruction of diverse visual stimuli (Gu et al., 2022; Ozcelik et al., 2022; Gao et al., 2024a; Zhou et al., 2024; Mai et al., 2024; Gao et al., 2024b). This approach has proven effective in utilizing latent diffusion models for image reconstruction (Lin et al., 2022; Takagi & Nishimoto, 2023; Mai & Zhang, 2023; Scotti et al., 2024a; Chen et al., 2023), addressing inter-subject differences by either training separate models for individual subjects or employing partially unified models with subject-specific parameters. However, influenced by neural variability, cross-subject brain signals in the latent space are prone to semantic conflicts, which can lead to convergence at suboptimal points. MindAligner addresses this by leveraging an explicit functional alignment framework across brains, this approach more effectively utilizes shared functionalities among subjects, thereby mitigating semantic conflicts.
+
+# 2.2. Cross-Subject Functional Alignment
+
+As brains differ both in size and processing mechanisms (Allen et al., 2022; Finn et al., 2017), the resulting variability in fMRI signals has spurred research into brain alignment methods. The ideal condition for functional alignment methods often depends on shared stimuli, requiring paired data from multiple subjects exposed to identical visual inputs, with alignment achieved through reconstruction loss optimization (Dadi et al., 2020; Rastegarnia et al., 2023). Relying on this ideal shared stimuli condition, a new paradigm (Bazeille et al., 2019; Thual et al., 2022; 2023; Ferrante et al., 2024) has emerged, enabling explicit alignment under shared stimuli by transforming one subject's
+
+signal to that of another, thereby preserving functionality and facilitating knowledge transfer across subjects. However, such ideal large-scale shared stimuli are often absent. To better adapt to these scenarios, current methods focus on either anatomical alignment (Bao et al., 2025; Jiang et al., 2024; Shen et al., 2024) or functional alignment in latent space (Scotti et al., 2024a;b; Wang et al., 2024; Gong et al., 2025). MindEye2 (Scotti et al., 2024b) employs ridge regression to align different subjects into a shared latent space, followed by a shared decoding module. MindBridge (Wang et al., 2024) generates pseudo stimuli to create shared stimuli for brain alignment. However, these alignment methods either rely on shared stimuli, restricting their applicability to datasets without such conditions, or utilize latent space alignment, which impedes their ability to uncover intersubject neural variability. We introduce an explicit brain functional alignment model to conquer the restriction of the shared stimuli and enhance the interpretation of inter-subject neural variability.
+
+# 3. Preliminary
+
+We begin with the illustration of the problem definition, and preliminary on cross-subject brain decoding baseline that reconstructs visual stimuli in the data-limited setting.
+
+Problem Definition. The acquisition of fMRI data is both time-intensive and costly, leading to brain decoding scenarios frequently constrained by limited data. Therefore, this study focuses on investigating cross-subject brain decoding in a data-limited setting. We follow MindEye2 (Scotti et al., 2024b) to build this setting on the Natural Scenes Dataset (NSD) (Allen et al., 2022). Specifically, the decoding model is first pre-trained for one or several subjects  $S_{K}$  using their full 40 scanning sessions of fMRI signals. Subsequently, the pre-trained decoding model is transferred to a new subject  $S_{N}$ , using only a single session of scanned fMRI (approximately 1 hour of data, representing just  $2.5\%$  of the full dataset). Finally, the adapted model is tested on 1000 images shared across all subjects for subject  $S_{N}$ .
+
+Cross-subject Brain Decoding Baseline. Here we introduce our cross-subject decoding baseline method (Scotti et al., 2024b). To reduce inter-subject differences, the baseline model first employs linear layers to map brain signals from different subjects into a shared latent space. Then the fMRI embeddings are aligned with the latent space of a CLIP model (Radford et al., 2021) through a diffusion prior (Ramesh et al., 2022), thereby leveraging generative models' capabilities for visual reconstruction. The output embeddings are then fed through a low-level submodule and a retrieval submodule. Two corresponding losses are utilized: a low-level reconstruction loss  $\mathcal{L}_{low - level}$  between the blurry images generated by the low-level submodule and the ground truth, and a bidirectional MixCo loss  $\mathcal{L}_{BiMixCo}$
+
+![](images/972eed1e101ff83f79e94a27ed65154a2c08696530785b47de9e483ffbfe2d8d.jpg)
+
+![](images/242472c573aaa552b5fed53c834e5a4ce4054c58db55a477a7cd143940e8d8f4.jpg)  
+Figure 2. Overview of MindAligner. To achieve explicit brain functional alignment, given a pre-trained brain decoding model, we design a Brain Functional Alignment Module (BFA) that learns a Brain Transfer Matrix (BTM)  $\mathcal{M}$  for fMRI mapping between the known and novel subjects. BTM is decomposed into two low-rank matrices  $\mathbf{A}$  and  $\mathbf{B}$  to create latent space for further alignment. The Cross-Stimulus Neural Mapper is proposed to create fMRI pairs under shared stimuli. In addition to the alignment losses  $\mathcal{L}_{rec}$  and  $\mathcal{L}_{KL}$  between generated and real fMRI, a latent alignment loss  $\mathcal{L}_{latent}$  guides functional alignment based on stimulus similarities. In the inference stage, only the BTM is utilized for functional mapping, enabling cross-subject brain decoding.
+
+![](images/9af0384665b3ec3137c12af72fa9cf47ad50a2637a4f17093d45abe03eb7b532.jpg)
+
+to perform contrastive optimization between the retrieval module's output and the CLIP image embeddings. The final loss for the decoding model's training is formulated as:  $\mathcal{L}_{dec} = \mathcal{L}_{prior} + \alpha_1\mathcal{L}_{low-level} + \alpha_2\mathcal{L}_{BiMixCo}$ , where  $\mathcal{L}_{prior}$  denotes the diffusion prior loss that measures discrepancies between the CLIP image embedding and the outputs produced by the diffusion prior. More details can be found in (Scotti et al., 2024b).
+
+# 4. MindAligner
+
+# 4.1. Overview
+
+Building on the pre-trained brain decoding model, MindAligner utilizes a Brain Transfer Matrix (BTM) to transform signals from novel subjects into the signal space of a known subject with limited data, thereby enabling cross-subject brain decoding. To achieve reliable and fine-grained brain alignment, we design the Brain Functional Alignment Module (BFA). Notably, the alignment module is utilized only during the training phase to assist BTM learning; during the inference phase, only the lightweight BTM is retained. Next, we provide a detailed illustration of each module of MindAligner.
+
+# 4.2. Brain Transfer Matrix
+
+Given the fMRI signal  $\mathcal{F}_N$  of an arbitrary novel subject  $S_N$  as input, the Brain Transfer Matrix (BTM)  $\mathcal{M}$  aims to transform it into the fMRI signal of a subject  $S_K$  seen
+
+during pre-training through a linear transformation:
+
+$$
+\hat {\mathcal {F}} _ {K} = \boldsymbol {\mathcal {M}} \times \mathcal {F} _ {N}, \tag {1}
+$$
+
+where  $\hat{\mathcal{F}}_K$  denotes the fMRI signal projected to the brain space of the seen subject  $S_K$ . To improve parameter efficiency in the data-limited setting, we decompose  $\mathcal{M}$  into two low-rank matrices  $\mathbf{A}$  and  $\mathbf{B}$ ,
+
+$$
+\boldsymbol {\mathcal {M}} = \mathbf {A} \times \mathbf {B}, \tag {2}
+$$
+
+where  $\mathbf{A} \in \mathbb{R}^{n \times h}$  and  $\mathbf{B} \in \mathbb{R}^{h \times k}$ ,  $n$  and  $k$  are the voxel dimensions of the novel and known subjects' fMRI, and  $h$  is the hidden dimension. The matrix decomposition creates a shared latent space between two subjects for subsequent alignment.  $\mathcal{M}$  encodes transfer weights that capture region-level inter-subject brain correlations and can be utilized for functional alignment during inference. Moreover, these correlations provide valuable insights for analyzing inter-subject variability.
+
+# 4.3. Brain Functional Alignment Module
+
+To learn a reliable brain transfer matrix, the Brain Functional Alignment Module (BFA) conducts soft cross-subject alignment in both the brain space and the shared latent space of the BTM. As no strictly-paired fMRI data under identical stimuli is provided, we employ a cross-stimulus neural mapper to facilitate stimulus transformation, rendering fMRI-pairs under visually similar stimuli. Based on these fMRI-pairs, a multi-level brain alignment loss is employed to achieve final alignment.
+
+Table 1. Visual decoding performance comparison. 1h means 1 hour of data. Bold indicates the best performance.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="4">Low-Level</td><td colspan="4">High-Level</td><td colspan="2">Retrieval</td></tr><tr><td>PixCorr↑</td><td>SSIM↑</td><td>Alex(2)↑</td><td>Alex(5)↑</td><td>Incep↑</td><td>CLIP↑</td><td>Eff↓</td><td>SwAV↓</td><td>Image↑</td><td>Brain↑</td></tr><tr><td>(1 h) MindBridge</td><td>0.109</td><td>0.324</td><td>79.0%</td><td>86.1%</td><td>78.7%</td><td>82.7%</td><td>0.866</td><td>0.528</td><td>-</td><td>-</td></tr><tr><td>(1 h) MindEye2</td><td>0.195</td><td>0.419</td><td>84.2%</td><td>90.6%</td><td>81.2%</td><td>79.2%</td><td>0.810</td><td>0.468</td><td>79.0%</td><td>57.4%</td></tr><tr><td>(1 h) Ours</td><td>0.206</td><td>0.414</td><td>85.6%</td><td>91.6%</td><td>83.0%</td><td>81.2%</td><td>0.802</td><td>0.463</td><td>79.0%</td><td>75.3%</td></tr><tr><td>(subj1) MindBridge</td><td>0.124</td><td>0.333</td><td>80.55%</td><td>86.51%</td><td>78.72%</td><td>83.56%</td><td>0.868</td><td>0.526</td><td>-</td><td>-</td></tr><tr><td>(subj1) MindEye2</td><td>0.235</td><td>0.428</td><td>88.02%</td><td>93.33%</td><td>83.56%</td><td>80.75%</td><td>0.798</td><td>0.459</td><td>93.96%</td><td>77.63%</td></tr><tr><td>(subj1) Ours</td><td>0.226</td><td>0.415</td><td>88.19%</td><td>93.26%</td><td>83.48%</td><td>81.76%</td><td>0.800</td><td>0.459</td><td>90.90%</td><td>86.88%</td></tr><tr><td>(subj2) MindBridge</td><td>0.107</td><td>0.330</td><td>78.21%</td><td>85.85%</td><td>77.96%</td><td>81.04%</td><td>0.874</td><td>0.532</td><td>-</td><td>-</td></tr><tr><td>(subj2) MindEye2</td><td>0.200</td><td>0.433</td><td>85.00%</td><td>92.13%</td><td>81.86%</td><td>79.39%</td><td>0.807</td><td>0.467</td><td>90.53%</td><td>67.18%</td></tr><tr><td>(subj2) Ours</td><td>0.218</td><td>0.426</td><td>88.08%</td><td>93.33%</td><td>84.13%</td><td>82.47%</td><td>0.791</td><td>0.452</td><td>90.04%</td><td>85.61%</td></tr><tr><td>(subj5) MindBridge</td><td>0.105</td><td>0.326</td><td>80.18%</td><td>87.70%</td><td>81.95%</td><td>85.38%</td><td>0.840</td><td>0.507</td><td>-</td><td>-</td></tr><tr><td>(subj5) MindEye2</td><td>0.175</td><td>0.405</td><td>83.11%</td><td>91.00%</td><td>84.33%</td><td>82.53%</td><td>0.781</td><td>0.444</td><td>66.94%</td><td>46.96%</td></tr><tr><td>(subj5) Ours</td><td>0.197</td><td>0.409</td><td>84.69%</td><td>91.61%</td><td>84.63%</td><td>82.76%</td><td>0.784</td><td>0.454</td><td>70.62%</td><td>65.95%</td></tr><tr><td>(subj7) MindBridge</td><td>0.101</td><td>0.308</td><td>76.99%</td><td>84.51%</td><td>76.31%</td><td>80.98%</td><td>0.882</td><td>0.546</td><td>-</td><td>-</td></tr><tr><td>(subj7) MindEye2</td><td>0.170</td><td>0.408</td><td>80.70%</td><td>85.90%</td><td>74.90%</td><td>74.29%</td><td>0.854</td><td>0.504</td><td>64.44%</td><td>37.77%</td></tr><tr><td>(subj7) Ours</td><td>0.183</td><td>0.407</td><td>81.45%</td><td>88.31%</td><td>79.92%</td><td>77.82%</td><td>0.834</td><td>0.487</td><td>64.18%</td><td>62.58%</td></tr></table>
+
+Cross-stimulus Neural Mapper. Due to the lack of stimuli-pair where two subjects view the same image, we turn to select cross-subject fMRI pairs with similar stimuli  $\mathbf{I}_N$  and  $\mathbf{I}_K$  viewed by two subjects. The cross-stimulus neural mapper aims to transform the fMRI embedding  $z_N$  under stimuli  $\mathbf{I}_N$  into those corresponding to stimuli  $\mathbf{I}_K$  of the known subject  $S_K$ . However, due to the absence of brain prior knowledge, directly generating fMRI signals is still challenging. Therefore, we leverage the differences between  $\mathbf{I}_N$  and  $\mathbf{I}_K$  as conditions for generating fMRI. Based on the fMRI embedding  $z_N$  projected by the low-rank matrix  $\mathbf{A}$ , i.e.,  $z_N = \mathbf{A} \times \mathcal{F}_N$ , we use the visual stimuli difference  $E_{\mathrm{diff}}$  viewed by the two subjects as a condition to perform linear modulation to generate the stimuli embedding  $z_K$  corresponding to the stimuli  $\mathbf{I}_K$  of the known subject:
+
+$$
+\boldsymbol {E} _ {\text {d i f f}} = \mathcal {E} _ {\text {i m a g e}} (\boldsymbol {I} _ {N}) - \mathcal {E} _ {\text {i m a g e}} (\boldsymbol {I} _ {K}), \tag {3}
+$$
+
+$$
+\boldsymbol {z} _ {\text {d i f f}} = \boldsymbol {E} _ {\text {d i f f}} \times \boldsymbol {\mathcal {M}} _ {\text {d i f f}}, \tag {4}
+$$
+
+$$
+\boldsymbol {z} _ {K} = \mathcal {M} _ {C} \left(\boldsymbol {z} _ {N}, \boldsymbol {z} _ {\text {d i f f}}\right), \tag {5}
+$$
+
+where  $\mathcal{E}_{image}$  is the image encoder of pretrained CLIP (Radford et al., 2021). The cross-stimulus neural mapper  $\mathcal{M}_C(\cdot)$  is a linear modulation that splits the condition  $z_{\mathrm{diff}}$  to scale and shift parameters using  $\mathcal{M}_{\mathrm{diff}} \in \mathbb{R}^{a \times 2h}$ .  $a$  is the clip embedding's dimension. These parameters can be used to modulate the input  $z_N$ , thereby facilitating the cross-subject stimulus transformation in the latent space. The transformed embedding  $z_K$  is then projected to the known subject's space by the low-rank matrix  $\mathbf{B}$ , rendering a synthesized fMRI embedding  $\hat{\mathcal{F}}_K$  to be aligned with the known subject's real fMRI embedding  $\mathcal{F}_K$ :
+
+$$
+\hat {\mathcal {F}} _ {K} = \boldsymbol {z} _ {K} \times \mathbf {B}. \tag {6}
+$$
+
+The effectiveness of linear structures for cross-subject alignment is justified by both theoretical and empirical evidence (Appendix F).
+
+Multi-level Brain Alignment Loss. The brain alignment loss integrates both signal-level reconstruction loss between the generated and real fMRI signals and an embedding-level alignment loss to achieve more refined alignment across different visual stimuli. To ensure the quality of synthesized fMRI  $\hat{\mathcal{F}}_K$ , an fMRI reconstruction loss is designed to enforce the consistency between  $\hat{\mathcal{F}}_K$  and the real fMRI  $\mathcal{F}_K$  of the known subject:
+
+$$
+\mathcal {L} _ {\text {r e c}} = \left| \left| \tilde {\mathcal {F}} _ {K} - \mathcal {F} _ {K} \right| \right| _ {2} ^ {2}. \tag {7}
+$$
+
+To further enhance the alignment performance, we use the Kullback-Leibler (KL) Divergence loss to enforce the consistency between distributions of the generated and real fMRI signals:
+
+$$
+\mathcal {L} _ {K L} = \mathcal {K L} \left(\tilde {\mathcal {F}} _ {K}, \mathcal {F} _ {K}\right). \tag {8}
+$$
+
+To enable fine-grained functional mapping under different stimuli, we leverage intrinsic correlations between visual semantics to guide the alignment of the corresponding brain activities in the latent space. Specifically, we design a latent alignment loss  $\mathcal{L}_{\text {latent }}$  by enforcing the consistency between fMRI embedding pairs and stimuli pairs:
+
+$$
+\boldsymbol {\mathcal {L}} _ {\text {l a t e n t}} = \left| \left| \left(\boldsymbol {\mathcal {R}} \left(\mathcal {E} _ {f} \left(\boldsymbol {z} _ {N}\right), \mathcal {E} _ {f} \left(\boldsymbol {z} _ {K}\right)\right) - \boldsymbol {\mathcal {R}} \left(\boldsymbol {E} _ {N}, \boldsymbol {E} _ {K}\right) \right| \right| _ {2} ^ {2}, \right. \tag {9}
+$$
+
+where  $\pmb{E}_N$  and  $\pmb{E}_K$  denote the CLIP embeddings of  $\pmb{I}_N$  and  $\pmb{I}_K$ .  $\mathcal{R}(\cdot)$  calculates the dissimilarity matrix between embeddings pairs.  $\mathcal{E}_f$  denotes a functional embedder for fMRI embeddings for better dissimilarity calculation. Hence, the final brain alignment loss  $\mathcal{L}_{\text{align}}$  is formulated as:
+
+$$
+\mathcal {L} _ {\text {a l i g n}} = \mathcal {L} _ {\text {d e c}} + \alpha_ {\text {r e c}} \mathcal {L} _ {\text {r e c}} + \alpha_ {K L} \mathcal {L} _ {K L} + \alpha_ {l a} \mathcal {L} _ {\text {l a t e n t}}, \tag {10}
+$$
+
+where  $\alpha_{rec},\alpha_{kl},\alpha_{la}$  are the loss coefficients, and  $\mathcal{L}_{dec}$  denotes the decoding loss in the baseline method. The combination of these losses can improve the semantic accuracy of visual reconstruction and also reduce the reliance on same-stimulus data.
+
+# 4.4. Inference
+
+During inference, only the trained BTM is used for functional alignment (Eq. 1). The generated  $\hat{\mathcal{F}}_K$  is then directly fed into the pre-trained model for brain decoding. MindAligner is a lightweight functional alignment module that enables efficient cross-subject visual decoding.
+
+# 5. Experiments
+
+In this section, we present the implementation details, followed by fMRI-to-image reconstruction results and brain functional alignment analysis. The Appendix includes additional metrics, qualitative and quantitative results, model efficiency, and further visualizations.
+
+# 5.1. Implementation Details
+
+The BTM is composed of two bias-free linear layers with a hidden size  $h = 4096$ . The input dimension  $n$  and output dimension  $k$  of BTM are determined by the specific subject transfer pairs. For subjects 1, 2, 5, and 7, the dimensions are 15,724, 14,278, 13,039, and 12,682, respectively. The cross-stimulus neural mapper is implemented using the Feature-wise Linear Modulation model (Perez et al., 2018), where the input dimension of  $\mathcal{M}_{\mathrm{diff}}$  is  $a = 768$ , matching the CLIP embedding dimension, and the output is  $2h = 8192$ . The functional embedder is a linear layer with input and output sizes of  $h = 4096$ . The loss coefficients are set to  $\alpha_{rec} = 1$ ,  $\alpha_{la} = \alpha_{KL} = 0.001$ ,  $\alpha_{1} = 0.033$ , and  $\alpha_{2} = 0.016$ . The learning rates for the brain transfer matrix, cross-stimulus neural mapper, and functional embedder are all set to 1e-5. We use a batch size of 16 and optimize using Adam. Training on a single NVIDIA A100 GPU achieves convergence in approximately 12 hours.
+
+# 5.2. Dataset
+
+We use the Natural Scenes Dataset (NSD) (Allen et al., 2022), the largest publicly available set, widely used for brain visual decoding. It includes neural responses from subjects viewing complex images from the MSCOCO-2017 dataset(Lin et al., 2014). In line with MindEye2's data-limited setting, our approach uses only a single session of neural recordings, corresponding to one hour of data.
+
+# 5.3. Metrics
+
+To evaluate fMRI-to-image reconstruction performance, we assess both low- and high-level properties of the reconstructed images. Low-level properties capture fundamental visual elements like pixel similarity and edges, while high-level properties reflect semantic information. Following previous works (Scotti et al., 2024a;b), we adopt the PixCorr, SSIM, AlexNet(2), and AlexNet(5) (Krizhevsky et al., 2012)
+
+to evaluate low-level properties and use Inception (Szegedy et al., 2016), CLIP (Radford et al., 2021), EffNet-B (Tan & Le, 2019) and SwAV (Caron et al., 2020) to evaluate high-level properties. These metrics assess the fidelity of the reconstructed images by comparing them with the ground truth. Metric details can be found in Appendix A.
+
+To evaluate functional alignment, we use two metrics: fMRI Spatial Correlation (fSC) (Conroy et al., 2009) and Transfer Quantity (TQ). fSC measures the Pearson correlation between corresponding brain regions of two subjects  $(i\neq j)$ , assessing global alignment consistency. TQ captures voxel-level differences by analyzing the weights of the BTM  $\mathcal{M}$ , which maps voxels between subjects. For a source voxel indexed by  $i$ , TQ is defined as  $\mathrm{TQ}_i = \sum_{0\leq j < p}||\mathcal{M}_{i,j}||$  where  $p$  is the number of voxels in the target brain. High TQ values indicate regions with greater activation differences and more intricate functional alignment requirements.
+
+# 5.4. fMRI-based Visual Decoding
+
+We evaluate the visual decoding performance of MindAligner in both qualitative and quantitative manners. The compared state-of-the-art methods include our baseline MindEye2 (Scotti et al., 2024b) and MindBridge (Wang et al., 2024).
+
+Qualitative Comparison. We train our model using data from a single session (1 hour of data) and visualize the results in Fig. 3. MindAligner delivers decoding results that are more consistent with the original visual stimuli semantics compared to the baseline, highlighting its effectiveness. The performance improvement can be attributed to the effective learning of brain transfer matrix that accurately aligns the novel subject with the known subject, thus well transferring the pre-trained decoding model to the new subject with limited data.
+
+Quantitative Comparison. The first row (1h) of Tab. 1 represents the average of the results from the 2nd to the 5th rows, providing a fair comparison with other methods to demonstrate our superiority. “(subj1) Ours” refers to the average result obtained by aligning the novel subject (subj 1) to every subject in the known subject list (subj 2, 5, 7). Detailed results are provided in Appendix Tab. 7. As summarized in Tab. 1, MindAligner surpasses state-of-the-art methods across almost all metrics. Notably, it achieves a  $17.9\%$  improvement in brain retrieval performance. The observed improvement can be attributed to the inherent challenges in implicit alignment strategies employed in previous methods. Aligning multiple subjects with substantial individual differences remains a complex task that may result in information loss during the alignment process. Our method addresses this challenge by adopting an explicit alignment strategy that aligns one subject at a time, avoid conflicts arising from multi-subject alignment. Our model focuses on
+
+![](images/d69132cd748036fc795a67f83f8aa3c1b99edd80f92959b297d848b3bb1b886c.jpg)  
+Figure 3. Visualization of MindAligner's decoding results from training on one hour of data.
+
+![](images/4247c43226e85a503581ad95771406cd0c95368dc571f50ff2885b286c4b885f.jpg)  
+Figure 4. Visualization results of aligning a new subject with different known subjects.
+
+fine-grained cross-subject brain mapping, thereby achieving better decoding performance with high fidelity.
+
+Ablation Study. To evaluate the effectiveness of each model design in MindAligner, we perform an ablation study using Subject 2 as the novel subject and Subject 1 as the known subject. The results exclude the refinement step of MindEye2 for generated images. As shown in Tab. 2, training MindAligner with only the visual decoding loss  $\mathcal{L}_{dec}$  yields suboptimal cross-subject reconstruction performance, underscoring the difficulty of directly generalizing pre-trained models to new subjects without effective alignment. Adding signal reconstruction loss  $\mathcal{L}_{rec}$  significantly enhances performance as it leads to more accurate brain activity reconstructions. The incorporation of  $\mathcal{L}_{KL}$  further strengthens alignment by enforcing consistency between the distributions of the generated and real signals. Lastly,
+
+![](images/1b8617f6b1605d0a594c14c8b3d55ee3abcb3134363a6ae3dad0fc77cfbe6a6c.jpg)  
+Figure 5. Visualization of transfer quantity in brain heatmaps.
+
+![](images/6c2a930e9540a2df8f4044c0bf3d2b14402ef7a8229bada2ae38cfaea76e432f.jpg)  
+Figure 6. Comparison of fSC results between MindAligner and the baseline.
+
+$\mathcal{L}_{\text {latent }}$  exploits the correlation of visual stimuli and fMRI embeddings to guide the brain alignment in the latent space, thereby improving model's ability to capture visual semantics in brain activity and enhancing low-level reconstruction performance. These losses together work in synergy to refine alignment and improve cross-subject decoding fidelity.
+
+Impact of Aligning to Different Subjects. We visualize the results of fixing a new subject and aligning it with different known subjects in Fig. 4. The reconstruction performance
+
+Table 2. Ablation study results. The combination of  ${\mathcal{L}}_{dec} + {\mathcal{L}}_{rec} + {\mathcal{L}}_{KL} + {\mathcal{L}}_{latent}$  is our final model setting.  
+
+<table><tr><td>Method</td><td>PixCorr↑</td><td>SSIM↑</td><td>Alex(2)↑</td><td>Alex(5)↑</td><td>Incep↑</td><td>CLIP↑</td><td>Eff↓</td><td>SwAV↓</td></tr><tr><td>+Ldec</td><td>0.072</td><td>0.318</td><td>63.50%</td><td>71.44%</td><td>66.07%</td><td>62.59%</td><td>0.905</td><td>0.550</td></tr><tr><td>+Ldec+Lrec</td><td>0.186</td><td>0.340</td><td>86.83%</td><td>93.51%</td><td>84.55%</td><td>82.42%</td><td>0.811</td><td>0.465</td></tr><tr><td>+Ldec+Lrec+LKL</td><td>0.191</td><td>0.407</td><td>87.98%</td><td>92.99%</td><td>86.61%</td><td>82.16%</td><td>0.780</td><td>0.453</td></tr><tr><td>+Ldec+Lrec+KL+Llatent</td><td>0.195</td><td>0.408</td><td>88.25%</td><td>93.51%</td><td>86.24%</td><td>82.72%</td><td>0.782</td><td>0.454</td></tr></table>
+
+are stable when aligned with different known subjects, as the generated images are nearly identical. This demonstrates the robustness of our brain functional alignment. The choice of the known subject has minimal impact on visual decoding performance. This is because MindAligner leverages the intrinsic correlation between visual semantics to guide the alignment of corresponding brain activities, thereby facilitating robust alignment performance. We provide more detailed cross-subject results in Appendix C.
+
+Table 3. Efficiency comparison results. "Tr. Param." refers to the model's trainable parameters when adding a new subject.  
+
+<table><tr><td>Method</td><td>Tr. Param.</td><td>Total Param.</td><td>Inference</td></tr><tr><td>MindEye2</td><td>2.21G</td><td>2.21G</td><td>5.000 s</td></tr><tr><td>MindAligner</td><td>139.23M</td><td>2.21G</td><td>5.056 s</td></tr></table>
+
+Computational Efficiency. We compare the computational efficiency between our model and baseline MindEye2 w.r.t. parameter count and inference time per image. As shown in Tab. 3, MindAligner achieves superior decoding performance while significantly reduces the fine-tuning requirement, with just  $6\%$  of MindEye2's learnable parameters, demonstrating its efficiency. Moreover, the addition of our alignment module only slightly increases the inference time.
+
+# 5.5. Brain Functional Alignment Analysis
+
+To deepen the understanding of the brain functional alignment process, we provide detailed visualizations of brain regions along with corresponding functional analysis, offering insights into cross-subject variability and underlying neuroscience mechanisms.
+
+Region-level Functional Mapping. We apply the Transfer Quantity (TQ) metric on MindAligner's brain transfer matrix to assess cross-subject associations and visualize the results through brain heatmaps. As shown in Fig. 5, the visualization results highlight two key neuroscience findings. 1) The visual system exhibits a hierarchical pattern of inter-subject variability. The early visual region (labeled as "EarlyVis" in Fig. 5) presents lower inter-subject variability while higher visual regions (including OPA, FFA, PPA, and EBA) show larger variability. This graded variability aligns with established neuroscientific principles. The early visual region processing fundamental features like lines/textures show more conserved neural mechanisms, sharing larger
+
+![](images/43d03643aaf7ccc4003ee2fdbd71ca84952227ac7bc3372639cc9ceaed830178.jpg)  
+Figure 7. Visualization of transfer quantity in brain heatmaps from MindAligner training using 1 hour and full datasets.
+
+commonality across subjects. In contrast, higher visual regions handle more complex cognitive processes, including categorical perception and semantic understanding, leading to higher variability across individuals. 2) The ventral pathway exhibits the greatest inter-subject variability. The ventral pathway - anatomically positioned on the brain's ventral surface (lower section in Fig. 5) and encompassing functional regions like PPA and FFA - demonstrates the highest variability among visual pathways. This variability arises from its important role in high-level visual processing, such as object recognition, face perception, and semantic interpretation. The ventral stream integrates sensory input with prior knowledge, experiences, and cognitive biases. This results in greater individual differences, as factors like familiarity, attention, and personal experiences shape how visual information is interpreted and understood.
+
+Cross-subject Correlation Analysis. To assess the alignment effect of MindAligner, we measure the fMRI Spatial Correlation (fSC) for different subject pairs, comparing our functional alignment strategy with the baseline in reducing inter-subject differences, as shown in Fig. 6. We use  $1 \rightarrow 2$  to denote aligning Subject 1 to Subject 2. The results demonstrate that our method significantly outperforms the existing baseline in fSC in all transfer configurations, demonstrating the superiority of our explicit alignment manner against the implicit alignment. By establishing fine-grained voxel correspondences between subjects, MindAligner significantly enhances alignment performance even without paired fMRI signals under shared stimuli, leading to a better visual decoding performance.
+
+Brain Alignment with More Data. Furthermore, to evaluate MindAligner's robustness in limited data scenarios, we compare its performance using only 1 hour of fMRI data (2.5% of the full dataset) to using the full scanning sessions. As shown in Fig. 7, even with limited data, the TQ distribution closely resembles that of the full dataset, effectively identifying regions with significant inter-subject variability. This highlights the robustness of our explicit brain alignment strategy under data scarcity.
+
+# 6. Conclusion
+
+We present MindAligner, a functional alignment framework for cross-subject brain visual decoding. Unlike existing methods, it addresses insufficient alignment and lack of interpretability by learning a brain transfer matrix for voxel-level correspondences and proposing a brain functional alignment module for cross-subject mapping. Experiments validate the effectiveness of our method.
+
+# Impact Statement
+
+MindAligner enables high-quality visual perception reconstruction from a single fMRI session, potentially advancing the clinical diagnosis and brain computer interface applications. This approach holds significant potential for enabling alignment across diverse data formats and uncovering commonalities in brain organization across species, such as between humans and monkeys. Moreover, it could play a pivotal role in advancing the creation of unified brain atlases. The datasets used are publicly available, ensuring transparency and participant privacy.
+
+# Acknowledgements
+
+This work was done during the internship at Shanghai Artificial Intelligence Laboratory. This work is supported by Shanghai Artificial Intelligence Laboratory and the NSFC under Grant Nos. U24A20330 and 62361166670.
+
+# References
+
+Allen, E. J., St-Yves, G., Wu, Y., Breedlove, J. L., Prince, J. S., Dowdle, L. T., Nau, M., Caron, B., Pestilli, F., Charest, I., et al. A massive 7t fmri dataset to bridge cognitive neuroscience and artificial intelligence. Nature Neuroscience, 25(1):116-126, 2022.  
+Bao, G., Zhang, Q., Gong, Z., Zhou, J., Fan, W., Yi, K., Naseem, U., Hu, L., and Miao, D. Wills aligner: Multisubject collaborative brain visual decoding. In Proceedings of the AAAI Conference on Artificial Intelligence, pp. 14194-14202, 2025.  
+Bazeille, T., Richard, H., Janati, H., and Thirion, B. Local
+
+optimal transport for functional brain template estimation. In Information Processing in Medical Imaging, pp. 237-248. Springer, 2019.  
+Bazeille, T., DuPre, E., Richard, H., Poline, J., and Thirion, B. An empirical evaluation of functional alignment using inter-subject decoding. NeuroImage, 245: 118683, 2021. doi: 10.1016/j.neuroimage.2021.118683. URL https://www.sciencedirect.com/science/article/pii/S1053811921009563.  
+Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., and Joulin, A. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020.  
+Chen, Z., Qing, J., Xiang, T., Yue, W., and Zhou, J. Seeing beyond the brain: Conditional diffusion model with sparse masked modeling for vision decoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 22710-22720, 2023.  
+Conroy, B., Singer, B., Haxby, J., and Ramadge, P. J. fmri-based inter-subject cortical alignment using functional connectivity. Advances in Neural Information Processing Systems, 22, 2009.  
+Dadi, K., Varoquaux, G., Machlouzarides-Shalit, A., Gorgolewski, K. J., Wassermann, D., Thirion, B., and Mensch, A. Fine-grain atlases of functional modes for fmri analysis. NeuroImage, 221:117126, 2020.  
+Ferrante, M., Boccato, T., Ozcelik, F., VanRullen, R., and Toschi, N. Through their eyes: multi-subject brain decoding with simple alignment techniques. Imaging Neuroscience, 2:1-21, 2024.  
+Finn, E. S., Scheinost, D., Finn, D. M., Shen, X., Papademetris, X., and Constable, R. T. Can brain state be manipulated to emphasize individual differences in functional connectivity? NeuroImage, 160:140-151, 2017.  
+Gao, J., Fu, Y., Wang, Y., Qian, X., Feng, J., and Fu, Y. Mind-3d: Reconstruct high-quality 3d objects in human brain. In European Conference on Computer Vision, pp. 312-329. Springer, 2024a.  
+Gao, J., Fu, Y., Wang, Y., Qian, X., Feng, J., and Fu, Y. fmri-3d: A comprehensive dataset for enhancing fmri-based 3d reconstruction. arXiv preprint arXiv:2409.11315, 2024b.  
+Gong, Z., Zhang, Q., Bao, G., Zhu, L., Xu, R., Liu, K., Hu, L., and Miao, D. Mindtuner: Cross-subject visual decoding with visual fingerprint and semantic correction. In Proceedings of the AAAI Conference on Artificial Intelligence, pp. 14247-14255, 2025.
+
+Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020.  
+Gu, Z., Jamison, K., Kuceyeski, A., and Sabuncu, M. Decoding natural image stimuli from fmri data with a surface-based convolutional network. arXiv preprint arXiv:2212.02409, 2022.  
+Haxby, J. V., Guntupalli, J. S., Connolly, A. C., Halchenko, Y. O., Conroy, B. R., Gobbini, M. I., Hanke, M., and Ramadge, P. J. A common, high-dimensional model of the representational space in human ventral temporal cortex. Neuron, 72(2):404-416, 2011.  
+Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020.  
+Horikawa, T. and Kamitani, Y. Generic decoding of seen and imagined objects using hierarchical visual features. Nature Communications, 8(1):15037, 2017.  
+Jiang, S., Meng, Z., Liu, D., Li, H., Su, F., and Zhao, Z. Mindshot: Brain decoding framework using only one image. arXiv preprint arXiv:2405.15278, 2024.  
+Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 2012.  
+Lei, J., Dai, L., Jiang, H., Wu, C., Zhang, X., Zhang, Y., Yao, J., Xie, W., Zhang, Y., Li, Y., et al. Unibrain: Universal brain mri diagnosis with hierarchical knowledge-enhanced pre-training. arXiv preprint arXiv:2309.06828, 2023.  
+Li, C., Qian, X., Wang, Y., Huo, J., Xue, X., Fu, Y., and Feng, J. Enhancing cross-subject fmri-to-video decoding with global-local functional alignment. In European Conference on Computer Vision, pp. 353-369. Springer, 2025.  
+Lin, S., Sprague, T., and Singh, A. K. Mind reader: Reconstructing complex images from brain activities. Advances in Neural Information Processing Systems, 35: 29624-29636, 2022.  
+Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., and Zitnick, C. L. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pp. 740-755. Springer, 2014.  
+Lu, Y., Du, C., Zhou, Q., Wang, D., and He, H. Mindiffuser: Controlled image reconstruction from human brain activity with semantic and structural diffusion. In
+
+Proceedings of the ACM International Conference on Multimedia, pp. 5899-5908, 2023.  
+Mai, W. and Zhang, Z. Unibrain: Unify image reconstruction and captioning all in one diffusion model from human brain activity. arXiv preprint arXiv:2308.07428, 2023.  
+Mai, W., Zhang, J., Fang, P., and Zhang, Z. Brain-conditional multimodal synthesis: A survey and taxonomy. IEEE Transactions on Artificial Intelligence, 2024.  
+Naselaris, T., Kay, K. N., Nishimoto, S., and Gallant, J. L. Encoding and decoding in fmri. NeuroImage, 56(2):400-410, 2011.  
+Ozcelik, F. and VanRullen, R. Natural scene reconstruction from fmri signals using generative latent diffusion. Scientific Reports, 13(1):15666, 2023.  
+Ozcelik, F., Choksi, B., Mozafari, M., Reddy, L., and VanRullen, R. Reconstruction of perceived images from fmri patterns and semantic brain exploration using instance-conditioned gans. In International Joint Conference on Neural Networks, pp. 1-8. IEEE, 2022.  
+Perez, E., Strub, F., De Vries, H., Dumoulin, V., and Courville, A. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018.  
+Qian, C., Sun, X., Wang, Y., Zheng, X., Wang, Y., and Pan, G. Binless kernel machine: Modeling spike train transformation for cognitive neural prostheses. Neural Computation, 32(10):1863-1900, 2020.  
+Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pp. 8748-8763. PMLR, 2021.  
+Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022.  
+Rastegarnia, S., St-Laurent, M., DuPre, E., Pinsard, B., and Bellec, P. Brain decoding of the human connectome project tasks in a dense individual fmri dataset. NeuroImage, 283:120395, 2023.  
+Scotti, P., Banerjee, A., Goode, J., Shabalin, S., Nguyen, A., Dempster, A., Verlinde, N., Yundler, E., Weisberg, D., Norman, K., et al. Reconstructing the mind's eye: fmri-to-image with contrastive learning and diffusion priors. Advances in Neural Information Processing Systems, 36, 2024a.
+
+Scotti, P. S., Tripathy, M., Torrico, C., Kneeland, R., Chen, T., Narang, A., Santhirasegaran, C., Xu, J., Naselaris, T., Norman, K. A., and Abraham, T. M. Mindeye2: Shared-subject models enable fmri-to-image with 1 hour of data. In Proceedings of the International Conference on Machine Learning, volume 235 of PMLR, pp. 44038-44059, 2024b.  
+Seeliger, K., Güçlü, U., Ambrogioni, L., Güçlütürk, Y., and Van Gerven, M. Generative adversarial networks for reconstructing natural images from brain activity. NeuroImage, 181:775-785, 2018.  
+Shen, G., Zhao, D., He, X., Feng, L., Dong, Y., Wang, J., Zhang, Q., and Zeng, Y. Neuro-vision to language: Enhancing brain recording-based visual reconstruction and language interaction. Advances in Neural Information Processing Systems, 37:98083-98110, 2024.  
+Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., and Wojna, Z. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2818-2826, 2016.  
+Takagi, Y. and Nishimoto, S. High-resolution image reconstruction with latent diffusion models from human brain activity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14453-14463, 2023.  
+Tan, M. and Le, Q. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the International Conference on Machine Learning, pp. 6105-6114. PMLR, 2019.  
+Thual, A., Tran, Q. H., Zemskova, T., Courty, N., Flamary, R., Dehaene, S., and Thirion, B. Aligning individual brains with fused unbalanced gromov wasserstein. Advances in Neural Information Processing Systems, 35: 21792-21804, 2022.  
+Thual, A., Benchetrit, Y., Geilert, F., Rapin, J., Makarov, I., Banville, H., and King, J.-R. Aligning brain functions boosts the decoding of visual semantics in novel subjects. arXiv preprint arXiv:2312.06467, 2023.  
+Wang, S., Liu, S., Tan, Z., and Wang, X. Mindbridge: A cross-subject brain decoding framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11333-11342, 2024.  
+Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004.
+
+Zhou, Y., Liu, L., and Gou, C. Learning from observer gaze: Zero-shot attention prediction oriented by human-object interaction recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 28390-28400, 2024.
+
+# A. Explanation of Metrics
+
+Following prior work (Scotti et al., 2024a;b), we evaluate the image reconstruction results based on eight metrics, which are categorized into low-level and high-level groups. Low-level metrics, including Pixelwise Correlation (PixCorr), Structural Similarity Index (SSIM) (Wang et al., 2004), AlexNet(2) (Alex(2)), and AlexNet(5) (Alex(5)) (Krizhevsky et al., 2012), focus on textural and structural details. High-level metrics—Inception (Incep) (Szegedy et al., 2016), CLIP (Radford et al., 2021), EfficientNet-B (Eff) (Tan & Le, 2019), and SwAV-ResNet50 (SwAV) (Caron et al., 2020)—assess semantic fidelity. Alex(2), Alex(5), Incep, and CLIP metrics are derived by calculating Pearson correlation between the embeddings of the ground truth and reconstructed images, following the two-way identification framework of Ozcelik and VanRullen (Ozcelik & VanRullen, 2023). Eff and SwAV scores are based on the average distance between feature embeddings.
+
+In addition to the aforementioned metrics, we also evaluate the model using retrieval-based metrics to quantify the fine-grained image information in the fMRI embeddings, following the methodology in MindEye2 (Scotti et al., 2024b). Specifically, for image retrieval, each test fMRI scan is first transformed into its corresponding fMRI representation. We then compute the cosine similarity between this representation and the CLIP-derived image representations of 300 randomly selected images from the test set. Retrieval success is defined as the maximization of cosine similarity between the fMRI embedding and its ground truth CLIP embedding (top-1 retrieval, with random chance at 1/300). To mitigate variability from random batch sampling, the evaluation is repeated 30 times per test sample. The same procedure is applied for brain retrieval, with fMRI and image representations swapped.
+
+# B. Details on Model Parameters
+
+Table 4. Parameter counts of different modules.  
+
+<table><tr><td>Module</td><td>Parameter</td><td>Used during inference</td></tr><tr><td>BTM</td><td>122,888,192</td><td>✓</td></tr><tr><td>CNM</td><td>6,299,648</td><td>✗</td></tr><tr><td>FE</td><td>16,781,312</td><td>✗</td></tr><tr><td>MindEye2</td><td>2,227,290,748</td><td>✓</td></tr><tr><td>MindEye2.ridge_regression</td><td>64,405,504</td><td>✓</td></tr><tr><td>MindEye2.backbone</td><td>1,903,020,028</td><td>✓</td></tr><tr><td>MindEye2.diffusion_prior</td><td>259,865,216</td><td>✓</td></tr></table>
+
+Table 5. Trainable parameter share of different modules.  
+
+<table><tr><td>Module</td><td>Parameter %</td><td>Used during inference</td></tr><tr><td>MindEye2</td><td>100%</td><td>✓</td></tr><tr><td>MindAligner (Ours)</td><td>6.2%</td><td>-</td></tr><tr><td>MindAligner.BTM</td><td>5.2%</td><td>✓</td></tr><tr><td>MindAligner.CNM</td><td>0.3%</td><td>✗</td></tr><tr><td>MindAligner.FE</td><td>0.7%</td><td>✗</td></tr></table>
+
+We list the parameter counts of different modules in the pipeline. MindAligner comprises the Brain Transfer Matrix (BTM), Functional Embedder (FE), and Cross-Stimulus Neural Mapper (CNM). Tab. 4 shows the number of parameters for each module, while Tab. 5 displays the trainable parameter share for each module, helping us understand their relative contributions to the overall model. The results show that our model has a relatively small parameter count, accounting for only  $6\%$  of the parameter size of the visual decoding model. Moreover, the introduction of the FE and CNM modules during the training phase does not significantly increase the model's parameters, contributing to only  $1\%$  of the total parameter count. The BTM only accounts for  $5\%$  of the parameter size of the visual decoding model.
+
+![](images/8b97e8ad6207493431eafb554f8e6c953409f9923369e26d051b602078023f43.jpg)  
+Figure 8. More visualization of brain decoding results under different novel and known subjects.
+
+Table 6. Performance of MindAligner with different hidden sizes.  
+
+<table><tr><td rowspan="2">Hidden size</td><td colspan="4">Low-Level</td><td colspan="4">High-Level</td><td colspan="2">Retrieval</td></tr><tr><td>PixCorr↑</td><td>SSIM↑</td><td>Alex(2)↑</td><td>Alex(5)↑</td><td>Incep↑</td><td>CLIP↑</td><td>Eff↓</td><td>SwAV↓</td><td>Image↑</td><td>Brain↑</td></tr><tr><td>64</td><td>0.144</td><td>0.384</td><td>76.33%</td><td>85.34%</td><td>74.67%</td><td>73.99%</td><td>0.876</td><td>0.512</td><td>79.13%</td><td>64.00%</td></tr><tr><td>256</td><td>0.166</td><td>0.395</td><td>80.83%</td><td>87.81%</td><td>76.89%</td><td>76.47%</td><td>0.858</td><td>0.498</td><td>86.67%</td><td>77.83%</td></tr><tr><td>512</td><td>0.185</td><td>0.405</td><td>83.85%</td><td>90.95%</td><td>80.74%</td><td>78.55%</td><td>0.839</td><td>0.481</td><td>89.06%</td><td>82.97%</td></tr><tr><td>1024</td><td>0.204</td><td>0.415</td><td>87.01%</td><td>93.30%</td><td>83.51%</td><td>80.40%</td><td>0.811</td><td>0.463</td><td>90.30%</td><td>86.19%</td></tr><tr><td>2048</td><td>0.215</td><td>0.422</td><td>88.30%</td><td>93.30%</td><td>83.94%</td><td>82.75%</td><td>0.798</td><td>0.458</td><td>90.16%</td><td>85.96%</td></tr><tr><td>4096</td><td>0.218</td><td>0.425</td><td>88.36%</td><td>93.55%</td><td>84.17%</td><td>82.57%</td><td>0.794</td><td>0.455</td><td>90.09%</td><td>86.19%</td></tr></table>
+
+# C. More Results of Aligning to Different Subjects
+
+We visualized more results of fixing a new subject and aligning it with different known subjects in Fig. 8. MindAligner demonstrates robustness, as the generated images remain nearly identical when the novel subject is fixed. This is because MindAligner combines fMRI reconstruction between generated and real data under similar stimuli to ensure result fidelity, while also utilizing intrinsic correlations in visual semantics to guide the alignment of corresponding brain activities, enabling robust optimization.
+
+# D. Ablation Study on Hidden Size
+
+To investigate the potential for further reducing the model size, we adjusted the hidden size and evaluated the model's performance at different values. The experiments showed that when the hidden size is set to 1024, the model delivers comparable performance, while its size is reduced to one-quarter of the original. Compared to the  $d = 4096$  configuration, which only accounts for  $6\%$  of the framework, the  $d = 1024$  setting accounts for just  $2\%$ , further highlighting the efficiency
+
+Table 7. Detailed performance of our model compared with the baseline. Bold means our results outperform the baseline.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="4">Low-Level</td><td colspan="4">High-Level</td><td colspan="2">Retrieval</td></tr><tr><td>PixCorr↑</td><td>SSIM↑</td><td>Alex(2)↑</td><td>Alex(5)↑</td><td>Incep↑</td><td>CLIP↑</td><td>Eff↓</td><td>SwAV↓</td><td>Image↑</td><td>Brain↑</td></tr><tr><td>(subj1) MindEye2</td><td>0.235</td><td>0.428</td><td>88.02%</td><td>93.33%</td><td>83.56%</td><td>80.75%</td><td>0.798</td><td>0.459</td><td>93.96%</td><td>77.63%</td></tr><tr><td>(subj1) Ours</td><td>0.226</td><td>0.415</td><td>88.19%</td><td>93.26%</td><td>83.48%</td><td>81.76%</td><td>0.800</td><td>0.459</td><td>90.90%</td><td>86.88%</td></tr><tr><td>subj1→subj2</td><td>0.222</td><td>0.413</td><td>88.09%</td><td>93.28%</td><td>84.01%</td><td>81.82%</td><td>0.796</td><td>0.457</td><td>91.56%</td><td>87.49%</td></tr><tr><td>subj1→subj5</td><td>0.227</td><td>0.416</td><td>88.29%</td><td>93.36%</td><td>83.54%</td><td>80.94%</td><td>0.803</td><td>0.461</td><td>89.76%</td><td>85.78%</td></tr><tr><td>subj1→subj7</td><td>0.229</td><td>0.416</td><td>88.18%</td><td>93.13%</td><td>82.90%</td><td>82.52%</td><td>0.800</td><td>0.458</td><td>91.37%</td><td>87.36%</td></tr><tr><td>(subj2) MindEye2</td><td>0.200</td><td>0.433</td><td>85.00%</td><td>92.13%</td><td>81.86%</td><td>79.39%</td><td>0.807</td><td>0.467</td><td>90.53%</td><td>67.18%</td></tr><tr><td>(subj2) Ours</td><td>0.218</td><td>0.426</td><td>88.08%</td><td>93.33%</td><td>84.13%</td><td>82.47%</td><td>0.791</td><td>0.452</td><td>90.04%</td><td>85.61%</td></tr><tr><td>subj2→subj1</td><td>0.218</td><td>0.425</td><td>88.36%</td><td>93.55%</td><td>84.17%</td><td>82.57%</td><td>0.794</td><td>0.455</td><td>90.09%</td><td>86.19%</td></tr><tr><td>subj2→subj5</td><td>0.218</td><td>0.426</td><td>87.88%</td><td>93.13%</td><td>83.39%</td><td>82.05%</td><td>0.793</td><td>0.454</td><td>90.34%</td><td>85.67%</td></tr><tr><td>subj2→subj7</td><td>0.217</td><td>0.427</td><td>88.00%</td><td>93.32%</td><td>84.83%</td><td>82.78%</td><td>0.785</td><td>0.449</td><td>89.70%</td><td>84.98%</td></tr><tr><td>(subj5) MindEye2</td><td>0.175</td><td>0.405</td><td>83.11%</td><td>91.00%</td><td>84.33%</td><td>82.53%</td><td>0.781</td><td>0.444</td><td>66.94%</td><td>46.96%</td></tr><tr><td>(subj5) Ours</td><td>0.197</td><td>0.409</td><td>84.69%</td><td>91.61%</td><td>84.63%</td><td>82.76%</td><td>0.784</td><td>0.454</td><td>70.62%</td><td>65.95%</td></tr><tr><td>subj5→subj1</td><td>0.196</td><td>0.405</td><td>84.23%</td><td>91.28%</td><td>84.66%</td><td>82.93%</td><td>0.787</td><td>0.459</td><td>69.67%</td><td>65.14%</td></tr><tr><td>subj5→subj2</td><td>0.196</td><td>0.409</td><td>84.71%</td><td>91.88%</td><td>84.56%</td><td>82.88%</td><td>0.783</td><td>0.455</td><td>70.78%</td><td>66.38%</td></tr><tr><td>subj5→subj7</td><td>0.198</td><td>0.412</td><td>85.12%</td><td>91.67%</td><td>84.66%</td><td>82.47%</td><td>0.781</td><td>0.450</td><td>71.41%</td><td>66.32%</td></tr><tr><td>(subj7) MindEye2</td><td>0.170</td><td>0.408</td><td>80.70%</td><td>85.90%</td><td>74.90%</td><td>74.29%</td><td>0.854</td><td>0.504</td><td>64.44%</td><td>37.77%</td></tr><tr><td>(subj7) Ours</td><td>0.183</td><td>0.407</td><td>81.45%</td><td>88.31%</td><td>79.92%</td><td>77.82%</td><td>0.834</td><td>0.487</td><td>64.18%</td><td>62.58%</td></tr><tr><td>subj7→subj1</td><td>0.180</td><td>0.404</td><td>80.86%</td><td>87.47%</td><td>78.94%</td><td>77.05%</td><td>0.840</td><td>0.492</td><td>65.62%</td><td>63.69%</td></tr><tr><td>subj7→subj2</td><td>0.185</td><td>0.406</td><td>82.11%</td><td>89.01%</td><td>80.47%</td><td>77.53%</td><td>0.835</td><td>0.486</td><td>63.26%</td><td>61.06%</td></tr><tr><td>subj7→subj5</td><td>0.183</td><td>0.411</td><td>81.38%</td><td>88.46%</td><td>80.36%</td><td>78.87%</td><td>0.828</td><td>0.482</td><td>63.67%</td><td>62.98%</td></tr><tr><td>(1 h) MindEye2</td><td>0.195</td><td>0.419</td><td>84.2%</td><td>90.6%</td><td>81.2%</td><td>79.2%</td><td>0.810</td><td>0.468</td><td>79.0%</td><td>57.4%</td></tr><tr><td>(1 h) Ours</td><td>0.206</td><td>0.414</td><td>85.6%</td><td>91.6%</td><td>83.0%</td><td>81.2%</td><td>0.802</td><td>0.463</td><td>78.9%</td><td>75.3%</td></tr></table>
+
+performance advantages of our model.
+
+# E. More Detailed MindAligner Reconstruction Performance
+
+We provide more detailed MindAligner reconstruction evaluation results, as shown in Tab. 7. MindAligner surpasses the baseline in almost all metrics, even when applying the same novel subject to different known subjects. Notably, our method achieves a  $17.9\%$  improvement in brain retrieval performance. This improvement stems from addressing the limitations of implicit alignment strategies used in prior methods. Aligning multiple subjects with significant individual differences is inherently challenging and often leads to functional information loss during the alignment process. To overcome this, our approach employs an explicit alignment strategy, aligning one subject at a time, which effectively mitigates the conflicts arising from multi-subject alignment. By focusing on region-level cross-subject brain mapping, our model not only achieves superior visual performance but also captures more comprehensive brain region features for functional representation.
+
+# F. The Linear Hypothesis Underlying MindAligner
+
+Table 8. Comparison of different methods across various metrics.  
+
+<table><tr><td>Method</td><td>PixCorr↑</td><td>SSIM↑</td><td>Alex(2)↑</td><td>Alex(5)↑</td><td>Incep↑</td><td>CLIP↑</td><td>Eff↓</td><td>SwAV↓</td></tr><tr><td>FE (Transformer)</td><td>0.182</td><td>0.350</td><td>87.68%</td><td>93.55%</td><td>85.14%</td><td>81.65%</td><td>0.807</td><td>0.463</td></tr><tr><td>NM (Transformer)</td><td>0.169</td><td>0.339</td><td>85.40%</td><td>92.46%</td><td>84.23%</td><td>80.85%</td><td>0.819</td><td>0.482</td></tr><tr><td>Ours (Linear)</td><td>0.195</td><td>0.408</td><td>88.25%</td><td>93.51%</td><td>86.24%</td><td>82.72%</td><td>0.782</td><td>0.454</td></tr></table>
+
+The linear hypothesis in cross-subject brain difference modeling is a well-established principle in neuroscience. Haxby et al. (2011) demonstrated that inter-subject differences in visual representations can be eliminated via linear transformations, while Naselaris et al. (2011) showed that linear models account for over  $90\%$  of variance in primary sensory cortex decoding. Drawing on these established linear hypotheses, MindAligner employs linear structures to model brain variations across subjects and stimuli.
+
+To further substantiate the validity of our hypothesis, we perform ablation experiments on the subj2→subj1 transfer setting by replacing the original Functional Embedder (FE) and Cross-stimulus Neural Mapper (NM) components with nonlinear architectures, specifically employing Transformer layers. This allows us to directly assess the contribution of the linear design in each module and evaluate the impact of introducing nonlinearity on model performance and generalization.
+
+The superior performance of the linear layer further validates the linear hypothesis, demonstrating its effectiveness in modeling individual brain differences in limited data settings.

2025/MindAligner_ Explicit Brain Functional Alignment for Cross-Subject Visual Decoding from Limited fMRI Data/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f298633bd3c4ec6f56b1257a5e97dc17c1b18decc32352dddbf0cdced155b545
+size 1325836

2025/MindCustomer_ Multi-Context Image Generation Blended with Brain Signal/8c335697-32bf-4d12-8441-07e837ba6e05_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:636fc62374ec45c925801952c5111f471b455dd3777416a98ab4fcf7649777fb
+size 14799635

2025/MindCustomer_ Multi-Context Image Generation Blended with Brain Signal/full.md

@@ -0,0 +1,769 @@
+# MindCustomer: Multi-Context Image Generation Blended with Brain Signal
+
+Muzhou Yu  $^{*1}$  Shuyun Lin  $^{*2}$  Lei Ma  $^{34}$  Bo Lei  $^{3}$  Kaisheng Ma  $^{2}$
+
+![](images/e1e0257ba769edf1b6ad8398a50f0eb5764b965b17c5b43206d2cb8ced620148.jpg)  
+Overview of MindCustomer
+
+![](images/c1236ab4667f6a55f4eaf98101e8ad3d10999d370247848b3e9eefdd96178ab0.jpg)  
+Dual Contexts Generation
+
+![](images/0efa90101c2612bbbfc4d13ee362add82c13fb49253ef339bcbbb5197f6d26c0.jpg)
+
+![](images/e98260df72a4272d01eda8057e6118942e4042c16319b15b892d24d8175f70ea.jpg)  
+Triplet Contexts Generation
+
+![](images/a113c36d403730d61c83390ad456161046527fd17ba91b523c47102394e0f218.jpg)  
+Image Context
+
+![](images/58d5762419ff8256c038976e5d942c79909480f23956f16ecaddd4b9828b1dd6.jpg)
+
+![](images/c49379bdcd2c55c1b9eca40f8c6b427bd0a3bf259e232a874c2e00570fea1120.jpg)  
+New Human Subject Few-Shot Generation
+
+![](images/e14b92b9fb2fdd3d2f86ec041a1f899e7eac1e10ed364fbd5a075ff9e4291b60.jpg)  
+Figure 1. MindCustomer seamlessly integrates visual brain signals into multi-context image generation. Given the scarcity of real-world brain data, it efficiently performs real-time single-image generation without additional inputs (e.g. masks or multi-views of the same image context). Text context is optional, which provides more semantics or control over the image and brain contexts.
+
+# Abstract
+
+Advancements in generative models have promoted text- and image-based multi-context image generation. Brain signals, offering a direct representation of user intent, present new opportunities for image customization. However, it faces challenges in brain interpretation, cross-modal context fusion and retention. In this paper, we
+
+present MindCustomer to explore the blending of visual brain signals in multi-context image generation. We first design shared neural data augmentation for stable cross-subject brain embedding by introducing the Image-Brain Translator (IBT) to generate brain responses from visual images. Then, we propose an effective cross-modal information fusion pipeline that mask-freely adapts distinct semantics from image and brain contexts within a diffusion model. It resolves semantic conflicts for context preservation and enables harmonious context integration. During the fusion pipeline, we further utilize the IBT to transfer image context to the brain representation to mitigate the cross-modal disparity. MindCustomer enables cross-subject generation, delivering unified, high-quality, and natural image outputs. More
+
+over, it exhibits strong generalization for new subjects via few-shot learning, indicating the potential for practical application. As the first work for multi-context blending with brain signal, Mind-Customer lays a foundational exploration and inspiration for future brain-controlled generative technologies.
+
+# 1. Introduction
+
+With the advance of generative models (Chang et al., 2023; Ramesh et al., 2022a), text-to-image generation has garnered increasing attention. Users can achieve a desired visual effect or convey specific information by providing diverse text prompts. More recently, advancements in large-scale models, such as diffusion (Rombach et al., 2022; Xu et al., 2023), CLIP (Radford et al., 2021b), and large language models (Naveed et al., 2023; OpenAI, 2023) have further enhanced image generation (Kawar et al., 2023; Nichol et al., 2022). Based on this, numerous studies (Ding et al., 2024; Kumari et al., 2023; Ruiz et al., 2023; Gal et al., 2023c) have made significant strides in multi-context blended image generation from image and text prompts. It enables high-quality and customized creations with greater control, which has broad applications, including art creation, user-specific design, and virtual interactions.
+
+Simultaneously, we notice that research on brain signal interpretation leveraging decoding technology enhances the understanding of biological information (Gong et al., 2024; Gao et al., 2024). However, current works are mainly limited to the brain signal reconstruction (Scotti et al., 2024; Wang et al., 2024; Xia et al., 2024). As a significant semantic modality, brain signals straightforwardly reflect users' personalized thoughts and preferences. We believe that it has a broader extension to customized image creation. Compared to traditional prompts from text and image modalities, brain signals offer a direct representation of user intent, paving the way to produce more efficient and personalized generation. Based on the above observation, this paper pioneers the exploration of multi-context image generation blended with brain signals. It integrates cross-modal contexts from the image, text, and brain signal into a unified generative model to facilitate image customization.
+
+However, it lays several key challenges. (1) Cross-subject brain encoding: to achieve unified brain context extraction across different subjects, it is crucial to precisely capture semantic brain embeddings under the shared data scarcity. (2) Multi-modal integration: the generative model should integrate implicit representations from multiple modalities in a cohesive manner; otherwise, the embeddings from different modalities may confuse the model's latent space. (3) Multi-context preservation: when multiple contexts are fused in
+
+one generative model, it is important to ensure that the generated images retain the semantics originating from each context, without being overshadowed or altered.
+
+In this paper, we present MindCustomer to address this challenging task. We incorporate the visually evoked brain signals across different human subjects, rich in diverse semantic information from viewing natural images, as the brain contexts. MindCustomer leverages the diffusion model (Xu et al., 2023) as the base structure for prompt-guided image generation to synthesize high-quality images. More importantly, we identify three critical points: (1) We introduce the Image-Brain Translator (IBT) for cross-subject brain data augmentation, ensuring subsequent stable embeddings. Trained to generate pseudo-brain data from image inputs, IBT works alongside a semantic mapper to precisely capture semantic contexts in brain embeddings. (2) We propose an effective cross-modal information fusion pipeline for multi-context preservation during diffusion generation. In order to blend different contexts, we design diffusion fine-tuning based on image context to learn the image semantics, and then we freeze the fine-tuned diffusion model and lightly optimize the brain context with the target of the image context. This helps mitigate semantic conflicts between different contexts. (3) Besides, during the above process, we further utilize the IBT and brain embedding model to transfer image context to brain-embedding space to mitigate the representation disparity across modalities. Thus, the above-designed brain-blended cross-modal fusion enables final natural integration and semantic preservation across modalities, producing natural and high-quality blending results. MindCustomer performs real-time generation on a single image in light of the scarcity of neural data. Meanwhile, it efficiently achieves multi-context blending without additional masks or samples.
+
+As the first work, MindCustomer successfully implements multi-context blended with brain signal in image generation. It achieves high-quality image customization guided by cross-subject brain contexts. Furthermore, in real-world scenarios where data from new subjects is often limited, we apply MindCustomer with few-shot learning to showcase its generalization capability. Figure 1 illustrates the overview and performance of MindCustomer.
+
+In summary, our contributions are:
+
+- We explore the task of blending brain semantics in multi-context image generation for the first time, and present the effective MindCustomer to achieve it.  
+- We design the shared neural data augmentation via IBT to ensure stable cross-subject brain encoding. Furthermore, we present an effective pipeline to preserve and fuse contexts naturally during generation, alongside IBT to mitigate the cross-modal disparity.
+
+![](images/d1ba885e1d7aaa4592f4ddc253c964f9e370abf104b88885364ea3c687d57e3d.jpg)  
+Figure 2. Pipeline of MindCustomer. (1) IBT simulates shared brain data across human subjects to augment brain embedding and the semantic extractor is designed to refine the representation. (2) We transfer the image context to the brain embedding space and utilize it for the diffusion fine-tuning. (3) The brain context is optimized by using the fine-tuned diffusion, further ensuring cross-context adaptation. (4) We combine these multi-modal embeddings to a new one, and finally generate the coherent blending results.
+
+![](images/28c2d1f0560f9d0a4566f7da025bae8b9050cd6497cefc638b69fc8ead232c97.jpg)
+
+- MindCustomer achieves high-quality and seamless image generation guided by brain semantics. It can further generalize to new subjects, performing strong real-world application capability.
+
+# 2. Related Work
+
+Advancements in diffusion models (Ho et al., 2020; Dhariwal & Nichol, 2021; Song et al., 2021) have sparked interest in personalized image creation, allowing individuals to generate images based on their creative intent. Many works adopt fine-tuning diffusion as a core principle (Ruiz et al., 2023; Gal et al., 2023a). DreamBooth (Ruiz et al., 2023) fine-tunes the diffusion model to associate a unique identifier with a target subject, utilizing a prior preservation loss to maintain model generalization in few-shot tuning. Textual Inversion (Gal et al., 2023a) leverages a text embedding that is learnable for encoding a subject concept. Blip-Diffusion (Li et al., 2023a), building on BLIP-2's approach (Li et al., 2023b), pre-trains a multimodal encoder to generate visual representations aligned with text prompts, enabling a diffusion model to leverage these representations for learning subject-specific features and creating novel renditions. Recent research also explores merging multiple concepts for generation (Avrahami et al., 2023; Tewel et al., 2023a; Gu et al., 2023a; Liu et al., 2023b;c). Custom Diffusion (Kumari et al., 2023) implements this by employing a constrained optimization method with a closure solution.
+
+FreeCustom (Ding et al., 2024) designs a weighted mask strategy to fuse input concepts with more focus. Overall, these advancements in image creation facilitate user-driven image generation guided by text and image prompts.
+
+The latest developments in hardware have allowed for the recording and analysis of various brain signals in neuroscience labs (Sun et al., 2024; Taheri et al., 1994; Hari & Lounasmaa, 1989). Brain signals, rich in semantic information, are crucial for studying neural mechanisms (Daly & Wolpaw, 2008; Chaudhary et al., 2016; Weiskopf, 2012). Decoding visual brain signals has become a key research area, helping researchers understand human visual perception (Cox & Savoy, 2003; Horikawa & Kamitani, 2015; Schoenmakers et al., 2013; VanRullen & Reddy, 2018; Gu et al., 2023b; Quan et al., 2024). The development of generative models like diffusion has significantly advanced high-quality brain signal decoding (Wang et al., 2024; Xia et al., 2024; Scotti et al., 2023). A predominant approach (Scotti et al., 2024; Shen et al., 2024) involves aligning brain signal encodings with the CLIP space (Radford et al., 2021b), which serves as a condition input for generative models to perform the reconstruction. Additionally, Takagi et al. (Takagi & Nishimoto, 2023) reconstruct high-resolution images from brain activity while preserving rich semantic details, without the need for training or fine-tuning complex deep generative models. Mind-Vis (Chen et al., 2023) utilizes Masked Autoencoders (MAE) to enhance brain signal encoding, enabling more accurate reconstructions. Given
+
+the variability of brain signals across individuals, cross-subject decoding methods have been proposed (Wang et al., 2024; Xia et al., 2024) to integrate the encodings and reconstructions of multiple subjects within a unified model. Moreover, some works extend brain signals to more tasks (Xia et al., 2024; Shen et al., 2024; Chen et al., 2024a), such as visual grounding, visual-question-answer, and stylized reconstruction. As a concurrent work, MindPainter (Yu et al., 2025) uses the brain signal as a condition for masked image painting via self-supervised learning. Different from these works, this paper presents a mask-free, cross-subject, and multi-context (modalities in image, text, and brain) blending task for image generation. We design the shared neural data augmentation and cross-modal information fusing pipeline to effectively achieve high-quality image blending results. MindCustomer further broadens the way for brain-controlled image creation.
+
+# 3. Method
+
+# 3.1. Preliminaries
+
+NSD Data To better understand the brain-driven context, we first introduce the data related to visual brain signals. In this paper, we utilize the widely-used data, Natural Scenes Dataset (NSD) (Allen, 2022), as our brain contexts. This dataset includes high-resolution 7-Tesla fMRI scans from eight healthy adult participants, who were tasked with viewing various natural images sourced from the MS-COCO dataset (Lin et al., 2014). In accordance with standard practices (Scotti et al., 2024; Wang et al., 2024; Takagi & Nishimoto, 2023; Shen et al., 2024; Chen et al., 2024a), we utilize all subject-wise data from four subjects (Subj1, 2, 5, 7) as the training data. Each subject viewed 8859 individual images. And the remaining data of 982 images were viewed by all four subjects as the test data.
+
+Diffusion Model Diffusion models (Rombach et al., 2022; Xu et al., 2023) have shown remarkable efficacy in synthesizing high-quality images. Leveraging extensive text-image pair datasets during training, they are capable of generating highly accurate visual representations from textual input. In this study, we leverage versatile diffusion (VD) (Xu et al., 2023), which is a multi-modal diffusion with independent flows for image and text prompts. This inherently enables us to condition the model with both image and text modalities, and we mainly focus on integrating visual brain signals with image modality in VD to generate high-quality images. We initialize our model with the publicly pre-trained VD, and the subsequent fine-tuning process is built upon this prior.
+
+# 3.2. MindCustomer
+
+Given image contexts, brain signals from different human subjects and text prompts, our target is to perform mask-free blending to generate images. The generation results should
+
+naturally blend with the provided contexts, meanwhile preserving the relevant semantic characteristics. We propose MindCustomer to achieve this task. Considering the limited availability of brain data from different subjects, it is difficult to construct large, labeled datasets for training generative models. In this work, we perform efficient real-time generation on a single image, which holds greater practical significance for real-world applications. MindCustomer consists of four parts, which is illustrated in Figure 2.
+
+Brain Representation Pre-training To tackle the inconsistency for cross-subject brain embedding, inspired from (Scotti et al., 2024; Wang et al., 2024), we apply individual-to-shared encoding to encode fMRI voxels into the universal latent space of the pre-trained CLIP model. However, previous methods only allow the subject-wise data for training, where the limited data affects the encoding capability. We introduce Image-Brain Translator  $\mathcal{G}_{\eta}$  (IBT) to generate pseudo voxels to augment the training data without manual collecting.  $\eta$  refers to the parameters of the model  $\mathcal{G}$ . IBT is trained to simulate subject-specific brain signals from a given input image. During the brain representation pre-training stage, we concurrently feed fMRI data from multiple subjects (including true and generated fMRI) in response to the same stimulus into a shared encoding model. This approach effectively augments the fMRI training dataset, as the stimulus presented to each subject differs from those in the original dataset. As a result, we can generate shared brain representations across subjects, rather than treating each subject's data as isolated samples, thereby enhancing the stability of the embedding training data. Specifically, we utilize the pairs of fMRI voxels  $B$  and corresponding visual stimuli  $I \in \mathbb{R}^{N \times 256 \times 256}$  in a batch size of  $N$  to train the subject-wise IBT. We adopt the adaptive max pooling (Wang et al., 2024) to achieve the subject-invariant voxel sizes. Thus, we construct the fixed size of ground truth voxel  $B'$  from  $(13000 \sim 16000)$  to 8192. The training function of IBT  $\mathcal{G}_{\eta}$  can be formulated as:
+
+$$
+E _ {I} = \operatorname {C L I P} _ {\text {I m a g e}} (I),
+$$
+
+$$
+\mathcal {L} _ {I B T} = L _ {\mathrm {M S E}} \left(\mathcal {G} \left(E _ {I}, \eta\right), B ^ {\prime}\right), \tag {1}
+$$
+
+We feed the  $I$  into the pre-trained CLIP ViT/L-14 (Radford et al., 2021b) to obtain the latent feature  $E_{I}$  of size  $N \times 1024$ .  $L_{MSE}$  is the MSE loss function. Note that the CLIP model remains frozen.
+
+Next, we elaborate on the pre-training of cross-subject brain embedder. We denote different subjects as  $a$ ,  $b$ ,  $c$ , and  $d$ . For a pair of  $(I_a, B_a)$  from subject  $a$ , we first obtain the text caption  $T_a$  from image  $I_a$ , and feed both  $I_a$  and  $T_a$  into CLIP model to get the corresponding image and text CLIP embedding  $E_{Ia}$  and  $E_{Ta}$ . Then we fix the size of  $B_a$  to  $B_a' \in \mathbb{R}^{8192}$ , and apply the IBT to generate  $B_b' \in \mathbb{R}^{8192}$ ,  $B_c' \in \mathbb{R}^{8192}$  and  $B_d' \in \mathbb{R}^{8192}$  from input  $I_a$ . We employ the shallow subject-wise em
+
+![](images/3d57f2bd94f6c21d6963359f1a7aa4c8bd71dc2afd566358b46102cd0ced66e9.jpg)  
+Figure 3. Multi-context generation results across different subjects.
+
+bedders  $\mathcal{E}_f$  and deep shared embeddings  $\mathcal{E}_{\int}$  to output embeddings of voxels,  $(e_{Ia} \in \mathbb{R}^{8192}, e_{Ta} \in \mathbb{R}^{8192})$ ,  $(e_{Ib}, e_{Tb})$ ,  $(e_{Ic}, e_{Tc})$ , and  $(e_{Id}, e_{Td})$ , which will be supervised by CLIP-encoded text and image. Further, we utilize the SoftCLIP loss and MSE loss to optimize the error between the predicted embeddings and corresponding CLIP embeddings  $\mathcal{L}_{\mathrm{emb}} = \mathcal{L}_{\mathrm{SoftCLIP}} + \mathcal{L}_{\mathrm{MSE}}$ , where
+
+$$
+\mathcal {L} _ {\text {S o f t C L I P}} = - \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {N} \left[ \log \frac {\exp \left(E _ {i} \cdot E _ {j} / \tau\right)}{\sum_ {n = 1} ^ {N} \exp \left(E _ {i} \cdot E _ {n} / \tau\right)} + \log \frac {\exp \left(e _ {i} \cdot E _ {j} / \tau\right)}{\sum_ {n = 1} ^ {N} \exp \left(e _ {i} \cdot E _ {n} / \tau\right)} \right] \tag {2}
+$$
+
+$$
+\mathcal {L} _ {\mathrm {M S E}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(e _ {i} - E _ {i}\right) ^ {2}, \tag {3}
+$$
+
+With the pseudo voxel augmentation and pre-training process, we can embed the cross-subject voxels into latent CLIP embedding, which benefits accurate semantic extraction for diffusion-based generation. Furthermore, to extract more semantically enhanced features, we introduce the mapper from ClipCap (Mokady et al., 2021) to guide our embedding model to explore the text semantics from brain signals. Since the mapper is formerly trained for image caption, it effectively transfers the vision features to language features, we utilize it as the supervisor for enhanced brain embeddings. We design the training of semantic extractor  $\mathcal{M}_{\sigma}$  for the embeddings of voxels. The output features of the extractor  $\mathcal{M}_{\sigma}$  are also aligned with the loss of SoftCLIP and MSE.
+
+![](images/4ce24e9780e1db49bf8132c2e982749e3db01c215d6472edd90ca925b9e9fd3e.jpg)  
+Figure 4. We utilize the same image context (I) and visual stimuli (B) for different subjects to blend. It demonstrates the robustness of MindCustomer and individual variability in generation.
+
+$$
+\mathcal {L} _ {\mathrm {M S E}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(s _ {i} - S _ {i}\right) ^ {2}, \tag {5}
+$$
+
+$S$  is the feature of the ground truth mapper, and  $s$  is the predicted feature of the extractor  $\mathcal{M}_{\sigma}$ . Note that the IBTs, shallow embedders, deep embedders, and semantic extractor are all MLP-based models. Especially, the IBT has only 3 layers of MLPs, which are light for efficient training.
+
+Diffusion Fine-tuning with Embedding Optimization We observed that when directly fusing brain modality with image modality, the semantic gap between the modalities often results in inferior generation outcomes. Although alignment within the CLIP latent space has been attempted, the implicit and complex representations of brain signals still exhibit semantic discrepancies with the image space. Therefore, we propose to apply the IBT  $\mathcal{G}_{\eta}$  to transform the image context  $I$  into brain voxel, aiming to eliminate inter-modal inconsistencies and reduce the semantic bias. We feed the transferred image context  $B_{p}$  into brain embedder  $\mathcal{E}$  to obtain the embeddings  $e_p$  and fine-tune the versatile diffusion to reconstruct the image. It not only allows the diffusion model to maintain the image context represented as the transferred brain signal but also benefits the later natural fusion of brain context. The fine-tuning of the diffusion model with the parameters  $\theta$  can be formulated as:
+
+$$
+\mathcal {L} _ {t} ^ {\mathrm {i m g}} (\theta) = \mathbb {E} _ {x, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| \epsilon - \epsilon_ {\theta} \left(x _ {t}, t, e _ {p}\right) \| _ {2} ^ {2} \right], \tag {6}
+$$
+
+$$
+\mathcal {L} _ {\text {S o f t C L I P}} = - \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {N} \left[ \log \frac {\exp \left(S _ {i} \cdot S _ {j} / \tau\right)}{\sum_ {n = 1} ^ {N} \exp \left(S _ {i} \cdot S _ {n} / \tau\right)} + \log \frac {\exp \left(s _ {i} \cdot S _ {j} / \tau\right)}{\sum_ {n = 1} ^ {N} \exp \left(s _ {i} \cdot S _ {n} / \tau\right)} \right] \tag {4}
+$$
+
+where  $t = 1,\ldots ,T$  and  $x_{t}$  is Gaussian corrupted image of input  $I$ .  $\epsilon_{\theta}(x_t,t)$  is a UNets-based denoising function.
+
+![](images/7939c6df6dfa2a0dc090b7d0c8562a8eca0393f7c3c74965b38d5f8e55d978d0.jpg)  
+Figure 5. Visualization of the generation process through gradual interpolation between image and brain contexts.
+
+![](images/fd66d159124d539a6fb983b6bbb72e8fd9cfb40fed2cb92b6b3f63fc8ad9e60b.jpg)  
+Img / Brain
+
+![](images/c2c8b195e50c40054767a6dce07a574451115d95ee9d1278ddf634139417a94c.jpg)  
+"in the snow"
+
+![](images/68e0270220cb2c65c9412d15a0a7e29e1380388a4724510629a8f758b90c8945.jpg)  
+"with beautiful sunset"
+
+![](images/7118b75a1e31dfd4552001afc2b5f84abd4f96dfcb39cd2e35f27a6f62e42324.jpg)  
+"acldren's drawing
+
+![](images/31db8843a6ac48af7c69d26f74ab17957b1c322f5379047eff01f2a53d81e7d9.jpg)  
+Img / Brain  
+Figure 6. Multi-context blending results with only different text contexts. MindCustomer robustly creates multi-context images that are content-consistent and naturally integrated.
+
+![](images/10a514abeaa9fcf9de7fe2608f57cfe2e9c5d85c31256c7aaa618bba01d62a80.jpg)  
+"play"
+
+![](images/e3f4e384c02875bf3023703b2e297ec6fbffb47555df09713305692d62ecd111.jpg)  
+"earth in the screen"
+
+![](images/ad7333b8c996bd81b7236fb4bab4792d44fe72cbb662947f189308538efc43c5.jpg)  
+"christmas atmosphere"
+
+![](images/1ccca37b402f86261ed929d1973b997ac86f9d1bbb3f8c21ae2534520b0f008b.jpg)  
+Img / Brain
+
+![](images/c501d185bc9f5d7f00b50558d136ac6fa4e7d562f292f3fcd8e6b2c40b9c3e16.jpg)  
+"cyberpunk style"
+
+![](images/b7d6bd04f3cf62792615ab3de64c008e8eb40a19e9f4fc7c6969339cc7845a89.jpg)  
+"snowy mountain with grand nebula"
+
+![](images/8582a194d6473fb79317c09e1b4bad8c38a139207887d244e5eb39d178ddb455.jpg)  
+traveling with beautiful coastline
+
+![](images/fdf1a65b924e3da5a78bd901b9a440a88bff41c24d34acb8f1fd82e01f85b3f6.jpg)  
+Img / Brain
+
+![](images/a036cd1acd7d40a28906fd453a0786e44aad2cf9f7c290f725ca2bab454fdba9.jpg)  
+"bus in the background"
+
+![](images/660da6f3ee5ef473fc91d1fc7a83dfc99cb55c7a4e6e27e3df60ec01638d90b4.jpg)  
+"Victory Gate behind"
+
+![](images/f88de9dcaef72a53f0a96728288dc385f10a02ddd646c8dcda7bee6d866f7a8d.jpg)  
+"in HongKong"
+
+Next, we empirically observe that when different contexts are fused often results in semantic conflict. Partial context may lost or disturbed. To better preserve the multi-semantics and enhance seamless fusion, we lightly optimize the brain context embedding  $e_b$  for approximating the semantics of the image context. The optimization only modifies the brain context embedding while we freeze the parameters of the diffusion model. We run this optimization for a few steps and this process can be formulated as:
+
+$$
+\mathcal {L} _ {t} ^ {\text {b r a i n}} \left(e _ {b}\right) = \mathbb {E} _ {x, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| \epsilon - \epsilon_ {\theta} \left(x _ {t}, t, e _ {b}\right) \| _ {2} ^ {2} \right], \tag {7}
+$$
+
+Consequently, we can adapt the semantics between the image and brain contexts in latent space that enables the diffusion model to better understand the fusion of both embeddings to generate natural images.
+
+Embeddings Integration Finally, we apply the fine-tuned
+
+diffusion for blended image outcomes. As for the generation only guided by dual contexts, we directly concatenate the embeddings of  $e_p$  and  $e_b$  for the VD to obtain the results. As for triplet contexts, the concatenated dimension of embeddings will exceed the model capacity. Thus, we utilize linear interpolation of  $e_p$  and optimized  $e_b$  with a hyperparameter  $\alpha \in [0,1]$  to build the combined embedding  $e_c$ :
+
+$$
+e _ {c} = (1 - \alpha) \cdot e _ {p} + \alpha \cdot e _ {b}. \tag {8}
+$$
+
+We employ the combined embedding  $e_{c}$  to the image flow of the fine-tuned versatile diffusion, and meanwhile feed the text context embedding to the text flow, to jointly generate the final multi-context blended image. Note that the  $\alpha$  can be smoothly adjusted for asymptotical generation results balancing the image and brain contexts.
+
+New-subject Few-shot Generation Collecting brain signals in real-world scenarios is time-consuming and costly, often resulting in limited data for new subjects. To address
+
+![](images/bd547a10864f2f44cf70053c631bc6f5aabf83c1ae52258d6bf2e8c74bf1afe1.jpg)  
+Figure 7. Baseline comparison. Baseline-1: We replace the brain context with its visual stimuli to construct a multi-image blending. Baseline-2: We reconstruct the brain context to the image and merge all the image contexts for blending. Benefiting from the designed cross-modal fusion pipeline, MindCustomer naturally preserves and fuses different contexts to generate more visually pleasing results.
+
+Table 1. Quantitative comparisons. We compare our method with the baselines on context similarity and generation quality.  
+
+<table><tr><td></td><td>Baseline-1</td><td>Baseline-2</td><td>Ours</td></tr><tr><td>CLIP-I↑</td><td>0.559</td><td>0.526</td><td>0.563</td></tr><tr><td>DINOv2↑</td><td>0.630</td><td>0.612</td><td>0.654</td></tr><tr><td>CLIP-IQA↑</td><td>0.734</td><td>0.830</td><td>0.892</td></tr></table>
+
+this, MindCustomer constructs a cross-subject model that integrates brain signals from different subjects into a unified framework, enabling better generalization to new subjects. Specifically, we first use limited new-subject data to train a new IBT and fine-tune the brain embedding network. Leveraging the Reset-Tuning operation (Wang et al., 2024), we train only the shallow embedder for the new subject while freezing the deep shared embedder, without IBT augmentation. The remaining generation steps are the same as the above processes. Empirical results show that using less than  $20\%$  of the full dataset is sufficient for few-shot learning to generate natural outputs.
+
+# 4. Experiments
+
+Evaluation Metrics To assess the semantic preservation of the generated image, we use the common metrics on image similarity, DINOv2 (Oquab et al., 2024) and CLIP-I (Radford et al., 2021c). We further apply CLIP-IQA (Wang et al., 2023) to test the generation quality, a measure evaluating both the visual appeal and the conceptual perception of an image. Moreover, we adopt metrics of visual fMRI reconstruction to assess brain decoding performance. Following previous works (Wang et al., 2024; Xia et al., 2024), PixCorr, SSIM (Wang et al., 2004), AlexNet(2), and AlexNet(5)
+
+(Krizhevsky et al., 2012) are low-level property evaluation, and metrics of Inception (Szegedy et al., 2016) and CLIP (Radford et al., 2021a) are high-level property evaluation. The implementation is described in Appendix-B.
+
+# 4.1. Qualitative Results
+
+We use MindCustomer for multi-context image generation by integrating semantic information from brain signals. Image contexts from the NSD COCO dataset are paired with brain contexts and optionally enhanced with concise textual prompts to control spatial, environmental, and stylistic attributes. We present the customization results across all the subjects of NSD in Figure 3, showcasing the generation of high-quality and contextually faithful images. More results are presented in Appendix-D.1.
+
+Since the visual stimuli in test data are shared across all four subjects, we conduct experiments under identical conditions using the same contexts for each subject. This setup allows us to evaluate the model's robustness and showcase the individual variability among different subjects. The results are shown in Figure 4.
+
+In the embedding interpolation of MindCustomer, we apply a hyperparameter  $\alpha$  to control the ratio of transferred image embedding and brain instruction embedding. In Figure 5, we visualize the generation process via smoothly changing  $\alpha$ . It can be observed that when the alpha value is around 0.5, our method achieves a natural fusion of two embeddings while preserving their respective feature information. Empirical experiments indicate that  $\alpha$  range between 0.4 and 0.7 generally yields optimal blending results.
+
+Moreover, we present more visual results in Figure 6 with only different text contexts. As we can see, MindCus
+
+![](images/b341429e92b1925209c1bae44b4098a7a967a1f6861ab9d3b2403d2c414d3022.jpg)  
+Figure 8. Few-shot generation to the new human subject. We separately utilize  $10\%$ ,  $20\%$ , and  $40\%$  of the subject's complete dataset to perform generation fine-tuning. MindCustomer generalizes well to the new subject, indicating its real-world applicability.
+
+tomer robustly creates multi-context images that are content-consistent and naturally integrated.
+
+# 4.2. Comparisons and Analysis
+
+Baseline Unlike traditional methods (Ruiz et al., 2023; Ding et al., 2024; Kumari et al., 2023), relying on masks or multiple images of the same object, MindCustomer enables mask-free, single-image generation, which makes direct comparisons challenging. To address this, we design two baselines using versatile diffusion (VD) (Xu et al., 2023), capable of multi-context blending. Baseline-1: We replace the brain context with its image-based visual stimuli and input all the contexts for VD to generate results. Baseline-2: We use an SOTA fMRI-to-image model (MindEye (Scotti et al., 2023)) to reconstruct the brain context as another image context, and merge all the contexts for VD blending. Note that as baselines, VD requires no extra fine-tuning or optimization and relies on the original large text-image pre-training capability for direct inference.
+
+Here we present both qualitative and quantitative comparisons of generated images in Figure 7 and Table 1. Due to the brain signal interpretation error caused by fMRI reconstruction and some semantic overlap between distinct image contexts resulting from the blending diffusion model, these baselines show some semantic deviation and loss. While
+
+Table 2. User study on multi-context image generation. Users are asked to score the generated images from 1 to 3 (3 is the best).  
+
+<table><tr><td>Method</td><td>Quality↑</td><td>Consistency↑</td></tr><tr><td>Baseline</td><td>1.868</td><td>1.675</td></tr><tr><td>Ours</td><td>2.618</td><td>1.868</td></tr></table>
+
+Table 3. Quantitative results on Subj1 few-shot generation.  
+
+<table><tr><td></td><td>10%</td><td>20%</td><td>40%</td></tr><tr><td>CLIP-I↑</td><td>0.523</td><td>0.541</td><td>0.561</td></tr><tr><td>DINOv2↑</td><td>0.642</td><td>0.650</td><td>0.652</td></tr><tr><td>CLIP-IQA↑</td><td>0.944</td><td>0.947</td><td>0.952</td></tr><tr><td>Average ↑</td><td>0.703</td><td>0.713</td><td>0.722</td></tr></table>
+
+in MindCustomer, we designed VD fine-tuning based on image context to learn the image semantics, and then we freeze the fine-tuned VD model, and lightly optimize brain contexts with the target of the image context. This helps mitigate semantic conflicts between different contexts. During the above process, we further utilize the IBT to transfer image context to brain-embedding space to reduce the gap between modalities. Thus, the designed brain-blended cross-modal fusion pipeline allows the blending to better preserve each modality's semantics and produce more accurate, high-quality results. The quantitative metrics also demonstrate that MindCustomer significantly outperforms the baselines in terms of context similarity and image generation quality.
+
+User Study We also conduct the human evaluation study, surveying our method and baseline. As the overall performance of Baseline-1 surpasses that of Baseline-2, we choose Baseline-1 as the baseline for this analysis. In the study, each generated image is scored according to the generation quality and the alignment to the contexts. In total, we collect 220 answers, whose results are summarized in Table 2. Participants exhibit a notable preference for our method in both perspectives. More details of the study protocols are described in Appendix-C.
+
+# 4.3. Few-shot Generation on New Subject
+
+In real-world scenarios, collecting brain signals from human subjects is highly costly, leading to data scarcity. Therefore, the ability of a designed algorithm to generalize to new individuals is crucial. To demonstrate MindCustomer's practical generalization capability, we pre-trained the model from Subj2, 5, and 7 and performed few-shot learning on Subj1. We fine-tuned our model using  $10\%$ ,  $20\%$ , and  $40\%$  of the total data from Subj1, respectively.
+
+As shown in Figure 8, MindCustomer achieved natural and competitive generation results even with limited data for fine-tuning. As the data point increased, the model consistently generated higher-quality images. The quantitative
+
+![](images/de30485ac7247bd20fa3284329ae560f495d744cbd3cdc439d2fae6b67891710.jpg)  
+Image
+
+![](images/c980c811b2d793f88c2a5331d11199b36596d88cf342dabcd909a5c2c12914a4.jpg)  
+Brain
+
+![](images/6704428401960d32c2732f7cae8a3cbc36dcd5f06480bfe1c53009ec65bbf660.jpg)  
+"Cyberpunk Style"
+
+![](images/1c5dd61b54d94b4fccf46766c298e01020882c5467091b46d9e97ccb7dfa8613.jpg)  
+(1)
+
+![](images/c6295c36d129efd115d6a0854b744ddf3d7b59eeeb6a4a78225ddd46d2c0f658.jpg)  
+Figure 9. Qualitative ablation. (1) wo [IBT & fine-tuning & optimization]: Direct generation by concatenating CLIP-encoded image and Brain embedder-encoded contexts leads to semantic inconsistency and context overlap. (2) wo IBT: Adding fine-tuning and optimization fails to resolve these issues. (3) wo [fine-tuning & optimization]: Using IBT for alignment before concatenation allows partial explicit representation of semantic information. (4) wo optimization: Adding fine-tuning significantly improves results, enabling better representation of different semantic contexts. (5) wo ClipCap: Without the introduced ClipCap for semantic enhancement in Brain Representation Pretraining may result in low details of context representation in blending results. MindCustomer: Incorporating all techniques, we can produce higher-quality image generation with detailed context fidelity.
+
+![](images/ea818478ab20746daf3cefceb7e22920061e624bc89e6da40e3a46c6aee4f79f.jpg)  
+(2)
+
+![](images/9c08cbad4855e2a21efbe1ce0fed3fe677c00bccd805ff4cbbe25c8ff9b97ea1.jpg)
+
+![](images/d2c7da88bf9363b06532a55e8346ad3a4493da6d0a13a212e4b834f7bf2ff5d2.jpg)  
+(3)
+
+![](images/9f9e90366d334b81a7f14457ff83c4a455e113c8b783dee76c2a87f317b3f0fd.jpg)
+
+![](images/514e3896d2973f76a2546aad62c9e0d9748270916176529e97274abc40b56d54.jpg)  
+(4)
+
+![](images/0b8c3d37d8b02bedcf6cb02e39dedced3ff25e7cd111121cb0e11da37dcb0a58.jpg)
+
+![](images/c31161c4d6ad5265b6d01c3455d30bfd4ecb0bcc3b35d2a54230fec081a87ad5.jpg)  
+(5)
+
+![](images/fb0d7f0c7aa7729ab264c0f0b8e88a31101bc8d8b52c928fc58d9284f07f8d5f.jpg)
+
+![](images/acc59e88c5ea903d02fb0824eae6715018c5eddbe3caf359db4cecbd9a57afeb.jpg)  
+MindCustomer
+
+![](images/7c9a10c68d4a6f8e77605116a633b6893902955776fbddfcf88c1b508f89439b.jpg)
+
+Table 4. Quantitative ablation. We compare MindCustomer with the ablations on context similarity and generation quality. (1) wo [IBT & fine-tuning & optimization]; (2) wo IBT; (3) wo [finetuning & optimization]; (4) wo optimization; (5) wo ClipCap.  
+
+<table><tr><td></td><td>(1)</td><td>(2)</td><td>(3)</td><td>(4)</td><td>(5)</td><td>MindCustomer</td></tr><tr><td>CLIP-I↑</td><td>0.518</td><td>0.515</td><td>0.529</td><td>0.548</td><td>0.560</td><td>0.563</td></tr><tr><td>DINOv2↑</td><td>0.594</td><td>0.595</td><td>0.614</td><td>0.635</td><td>0.637</td><td>0.654</td></tr><tr><td>CLIP-IQA↑</td><td>0.842</td><td>0.838</td><td>0.893</td><td>0.893</td><td>0.875</td><td>0.892</td></tr></table>
+
+results of few-shot learning under different data volumes are presented in Table 3, indicating the promising generalization ability of MindCustomer. More few-shot generation results are illustrated in Appendix-D.1.
+
+# 4.4. Ablation
+
+We conduct experiments to validate the components of MindCustomer. First, we design IBT to achieve cross-modal alignment by mapping semantic representations from different modalities into a shared implicit space, forming the basis for fusion. Second, we fine-tune the diffusion model on image contexts to preserve semantics and optimize brain contexts for high-level semantic adaptation, generating coherent images. Figure 9 illustrates the key components of our design, and Table 4 provides quantitative ablation results. Besides, we visualize the brain reconstruction ability of MindCustomer in Figure 10. Compared with the SOTA method (Xia et al., 2024), MindCustomer also demonstrates superiority and robustness across diverse subjects. The quantitative results on brain reconstruction and more detailed ablation experiments are conducted in Appendix-D.
+
+![](images/dc813c00cd6abcf3aed280a3edc4b8df414b732ad0aedf640ddf776db8ca57b3.jpg)  
+Visual Stimuli UMBRAE MindCustomer Visual Stimuli UMBRAE MindCustomer
+
+![](images/736b497120d3a13cd2135ba98d492170b76c185b754a0dd9a4a9e704cc11996b.jpg)
+
+![](images/019a9b5d604d1f08aae2029a683d4d7de863d5b5e144d8671a7db6a01179c27e.jpg)
+
+![](images/faf6f49b2f0f3bdb0daaf80a0fb56d0daedddd02d0110f6dfd4243c4033f15e6.jpg)
+
+![](images/280a13461f0075cf2aa6281edf9c380c6b4b5cd45cb143dae34884480ff658d1.jpg)
+
+![](images/ea2b0b2b944ff4336b59ccc240a16bf91c4ee75a58993f27727b6c3dc251a345.jpg)  
+Visual Stimuli  
+Figure 10. Above: Comparison of brain decoding on Subj1. Below: Brain decoding across subjects by MindCustomer.
+
+![](images/52cb9564925bd44a67e03241240cd68b9d375125098f514e6a20bde9dd8469d0.jpg)  
+Subj1  
+Subj2
+
+![](images/3db0e3b10e81dbfed1ca4719e4c17f5903a9dd3b1664bc22eeb79586f050bb3f.jpg)
+
+![](images/5c3917519d3a5336ab4dac4660650555c311d165c3f43fe7d967aec9fcb4e32c.jpg)  
+Subj5
+
+![](images/212d6e517de28549d0860e7f68d7231b65c2aa32096a86f6defeae50aded4f33.jpg)  
+Subj7
+
+# 5. Conclusion
+
+In this paper, we propose a novel brain-blended image customization method, MindCustomer. We introduce the IBT to enhance the data augmentation for cross-subject brain embedding. Further, the proposed fusion information pipeline, along with IBT, effectively adapts distinct semantics across multiple contexts. Thus, MindCustomer can generate context-preserved and naturally coherent blending results. Besides, it can also generalize to new human subjects via few-shot learning, demonstrating the potential for real-world applications. In all, we hope MindCustomer serves as an exploration of brain context in the image customization task and gives inspiration for follow-up research, promoting the development of brain-computer interaction.
+
+# Acknowledgements
+
+This work was supported by the Beijing Natural Science Foundation (Grant No.JQ24023 and No.F251020), Beijing Municipal Science & Technology Commission Project (No.Z231100006623010), National Science and Technology Major Project (2022ZD01163013 to B.L.), and Beijing Municipal Science and Technology Project (No.Z241100004224028).
+
+# Impact Statement
+
+With the development of generative models and large language models, AI-driven personalized generation is garnering increasing attention and application. At the same time, advancements in neural data decoding are empowering researchers to delve deeper into brain-machine interfaces (BMIs) and human-computer interaction (HCI). The integration of BMIs and generative models is poised to make personalized brain-controlled generation a significant trend in the future of generative artificial intelligence. Mind-Customer represents the first exploration of multi-context image generation guided by brain semantics. We aim for it to serve as foundational research, offering potential insights and inspiration for future research in related fields as well as practical applications. We believe personalized brain-controlled generation will lead to groundbreaking developments in areas such as artistic creation, game design, and advertising industries.
+
+# References
+
+Abdal, R., Zhu, P., Mitra, N. J., and Wonka, P. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Trans. Graph., 40(3):21:1-21:21, 2021.  
+Alaluf, Y., Richardson, E., Metzer, G., and Cohen-Or, D. A neural space-time representation for text-to-image personalization. ACM Trans. Graph., 42(6):243:1-243:10, 2023.  
+Allen, E.J., S.-Y. G. W. Y. e. a. A massive 7t fmri dataset to bridge cognitive neuroscience and artificial intelligence. pp. 116-126, 2022.  
+Avrahami, O., Lischinski, D., and Fried, O. Blended diffusion for text-driven editing of natural images. In CVPR, pp. 18187-18197. IEEE, 2022.  
+Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., and Lischinski, D. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia, pp. 96:1-96:12. ACM, 2023.  
+Brooks, T., Holynski, A., and Efros, A. A. Instructpix2pix:
+
+Learning to follow image editing instructions. In CVPR, pp. 18392-18402. IEEE, 2023.  
+Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., and Zheng, Y. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, pp. 22503-22513. IEEE, 2023.  
+Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M., Murphy, K. P., Freeman, W. T., Rubinstein, M., Li, Y., and Krishnan, D. Muse: Text-to-image generation via masked generative transformers. In ICML, volume 202 of Proceedings of Machine Learning Research, pp. 4055-4075. PMLR, 2023.  
+Chaudhary, U., Birbaumer, N., and Ramos-Murgualday, A. Brain-computer interfaces for communication and rehabilitation. Nature Reviews Neurology, 12(9):513-525, 2016.  
+Chee, J. Pearson's product-moment correlation: Sample analysis, 05 2015.  
+Chen, J., Qi, Y., Wang, Y., and Pan, G. Mind artist: Creating artistic snapshots with human thought. In CVPR, pp. 27197-27207. IEEE, 2024a.  
+Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., and Zhao, H. Anydoor: Zero-shot object-level image customization. In CVPR, pp. 6593-6602. IEEE, 2024b.  
+Chen, Z., Qing, J., Xiang, T., Yue, W. L., and Zhou, J. H. Seeing beyond the brain: Conditional diffusion model with sparse masked modeling for vision decoding. In CVPR, pp. 22710-22720. IEEE, 2023.  
+Cox, D. D. and Savoy, R. L. Functional magnetic resonance imaging (fmri) "brain reading": detecting and classifying distributed patterns of fmri activity in human visual cortex. NeuroImage, 19(2):261-270, 2003.  
+Daly, J. J. and Wolpaw, J. R. Brain-computer interfaces in neurological rehabilitation. The Lancet Neurology, 7(11): 1032-1043, 2008.  
+Dhariwal, P. and Nichol, A. Q. Diffusion models beat gans on image synthesis. In NeurIPS, pp. 8780-8794, 2021.  
+Ding, G., Zhao, C., Wang, W., Yang, Z., Liu, Z., Chen, H., and Shen, C. Freecustom: Tuning-free customized image generation for multi-concept composition. In CVPR, pp. 9089-9098. IEEE, 2024.  
+Evans, J. D. Straightforward statistics for the behavioral sciences. 1995. URL https://api-semanticscholar.org/CorpusID:121954573.
+
+Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A. H., Chechik, G., and Cohen-Or, D. An image is worth one word: Personalizing text-to-image generation using textual inversion. In *ICLR*. OpenReview.net, 2023a.  
+Gal, R., Arar, M., Atzmon, Y., Bermano, A. H., Chechik, G., and Cohen-Or, D. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Trans. Graph., 42(4):150:1-150:13, 2023b.  
+Gal, R., Arar, M., Atzmon, Y., Bermano, A. H., Chechik, G., and Cohen-Or, D. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Trans. Graph., 42(4):150:1-150:13, 2023c.  
+Gao, J., Fu, Y., Wang, Y., Qian, X., Feng, J., and Fu, Y. fmri-3d: A comprehensive dataset for enhancing fmri-based 3d reconstruction. CoRR, abs/2409.11315, 2024.  
+Gong, Z., Bao, G., Zhang, Q., Wan, Z., Miao, D., Wang, S., Zhu, L., Wang, C., Xu, R., Hu, L., Liu, K., and Zhang, Y. Neuroclips: Towards high-fidelity and smooth fmri-to-video reconstruction. CoRR, abs/2410.19452, 2024.  
+Gu, Y., Wang, X., Wu, J. Z., Shi, Y., Chen, Y., Fan, Z., Xiao, W., Zhao, R., Chang, S., Wu, W., Ge, Y., Shan, Y., and Shou, M. Z. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. In NeurIPS, 2023a.  
+Gu, Z., Jamison, K., Kuceyeski, A., and Sabuncu, M. R. Decoding natural image stimuli from fmri data with a surface-based convolutional network. In MIDL, volume 227 of Proceedings of Machine Learning Research, pp. 107-118. PMLR, 2023b.  
+Hari, R. and Lounasmaa, O. V. Recording and interpretation of cerebral magnetic fields. Science, 244(4903):432-436, 1989.  
+Härkönen, E., Hertzmann, A., Lehtinen, J., and Paris, S. Ganspace: Discovering interpretable GAN controls. In NeurIPS, 2020.  
+Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., and Cohen-Or, D. Prompt-to-prompt image editing with cross-attention control. In *ICLR*. OpenReview.net, 2023a.  
+Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., and Cohen-Or, D. Prompt-to-prompt image editing with cross-attention control. In ICLR. OpenReview.net, 2023b.  
+Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. In NeurIPS, 2020.
+
+Horikawa, T. and Kamitani, Y. Generic decoding of seen and imagined objects using hierarchical visual features. CoRR, abs/1510.06479, 2015.  
+Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., and Irani, M. Imagic: Text-based real image editing with diffusion models. In CVPR, pp. 6007-6017. IEEE, 2023.  
+Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet classification with deep convolutional neural networks. In NIPS, pp. 1106-1114, 2012.  
+Kumari, N., Zhang, B., Zhang, R., Shechtman, E., and Zhu, J. Multi-concept customization of text-to-image diffusion. In CVPR, pp. 1931-1941. IEEE, 2023.  
+Lang, O., Gandelsman, Y., Yarom, M., Wald, Y., Elidan, G., Hassidim, A., Freeman, W. T., Isola, P., Globerson, A., Irani, M., and Mosseri, I. Explaining in style: Training a GAN to explain a classifier in stylespace. In ICCV, pp. 673-682. IEEE, 2021.  
+Li, D., Li, J., and Hoi, S. C. H. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS, 2023a.  
+Li, J., Li, D., Savarese, S., and Hoi, S. C. H. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, volume 202 of Proceedings of Machine Learning Research, pp. 19730–19742. PMLR, 2023b.  
+Lin, T., Maire, M., Belongie, S. J., Hays, J., Perona, P., Ramanan, D., Dollar, P., and Zitnick, C. L. Microsoft COCO: common objects in context. In ECCV (5), volume 8693 of Lecture Notes in Computer Science, pp. 740-755. Springer, 2014.  
+Liu, X., Park, D. H., Azadi, S., Zhang, G., Chopikyan, A., Hu, Y., Shi, H., Rohrbach, A., and Darrell, T. More control for free! image synthesis with semantic diffusion guidance. In WACV, pp. 289-299. IEEE, 2023a.  
+Liu, Z., Feng, R., Zhu, K., Zhang, Y., Zheng, K., Liu, Y., Zhao, D., Zhou, J., and Cao, Y. Cones: Concept neurons in diffusion models for customized generation. In ICML, volume 202 of Proceedings of Machine Learning Research, pp. 21548-21566. PMLR, 2023b.  
+Liu, Z., Zhang, Y., Shen, Y., Zheng, K., Zhu, K., Feng, R., Liu, Y., Zhao, D., Zhou, J., and Cao, Y. Cones 2: Customizable image synthesis with multiple subjects. CoRR, abs/2305.19327, 2023c.  
+Mokady, R., Hertz, A., and Bermano, A. H. Clipcap: CLIP prefix for image captioning. CoRR, abs/2111.09734, 2021.
+
+Naveed, H., Khan, A. U., Qiu, S., Saqib, M., Anwar, S., Usman, M., Barnes, N., and Mian, A. A comprehensive overview of large language models. CoRR, abs/2307.06435, 2023.  
+Nichol, A. Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., and Chen, M. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In ICML, volume 162 of Proceedings of Machine Learning Research, pp. 16784-16804. PMLR, 2022.  
+OpenAI. GPT-4 technical report. CoRR, abs/2303.08774, 2023.  
+Oquab, M., Darcet, T., Moutakanni, T., Vo, H. V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P., Li, S., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., and Bojanowski, P. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024, 2024.  
+Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., and Lischinski, D. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, pp. 2065–2074. IEEE, 2021.  
+Quan, R., Wang, W., Tian, Z., Ma, F., and Yang, Y. Psychometry: An omnifit model for image reconstruction from human brain activity. CoRR, abs/2403.20022, 2024.  
+Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning transferable visual models from natural language supervision. In ICML, volume 139 of Proceedings of Machine Learning Research, pp. 8748-8763. PMLR, 2021a.  
+Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning transferable visual models from natural language supervision. In ICML, volume 139 of Proceedings of Machine Learning Research, pp. 8748-8763. PMLR, 2021b.  
+Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning transferable visual models from natural language supervision. In ICML, volume 139 of Proceedings of Machine Learning Research, pp. 8748-8763. PMLR, 2021c.  
+Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022a.
+
+Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022b.  
+Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In CVPR, pp. 10674-10685. IEEE, 2022.  
+Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., and Aberman, K. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pp. 22500-22510. IEEE, 2023.  
+Schoenmakers, S., Barth, M., Heskes, T., and van Gerven, M. Linear reconstruction of perceived images from human brain activity. NeuroImage, 83:951-961, 2013.  
+Scotti, P., Banerjee, A., Goode, J., Shabalin, S., Nguyen, A., cohen, e., Dempster, A., Verlinde, N., Yundler, E., Weisberg, D., Norman, K., and Abraham, T. Reconstructing the mind's eye: fmri-to-image with contrastive learning and diffusion priors. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2023.  
+Scotti, P. S., Tripathy, M., Torrico, C., Kneeland, R., Chen, T., Narang, A., Santhirasegaran, C., Xu, J., Naselaris, T., Norman, K. A., and Abraham, T. M. Mindeye2: Shared-subject models enable fmri-to-image with 1 hour of data. In ICML. OpenReview.net, 2024.  
+Shen, G., Zhao, D., He, X., Feng, L., Dong, Y., Wang, J., Zhang, Q., and Zeng, Y. Neuro-vision to language: Image reconstruction and language enabled interaction via brain recordings. CoRR, abs/2404.19438, 2024.  
+Shen, Y. and Zhou, B. Closed-form factorization of latent semantics in gans. In CVPR, pp. 1532-1540. Computer Vision Foundation / IEEE, 2021.  
+Shen, Y., Gu, J., Tang, X., and Zhou, B. Interpreting the latent space of gans for semantic face editing. In CVPR, pp. 9240-9249. Computer Vision Foundation / IEEE, 2020.  
+Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. In *ICLR*. OpenReview.net, 2021.  
+Sun, Y., Chen, X., Liu, B., Liang, L., Wang, Y., Gao, S., and Gao, X. Signal acquisition of brain-computer interfaces: A medical-engineering crossover perspective review. Fundamental Research, 2024. ISSN 2667-3258. doi: https://doi.org/10.1016/j.fmre.2024.04.011. URL https://www.sciencedirect.com/science/article/pii/S2667325824001559.
+
+Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., and Wojna, Z. Rethinking the inception architecture for computer vision. In CVPR, pp. 2818-2826. IEEE Computer Society, 2016.  
+Taheri, B. A., Knight, R. T., and Smith, R. L. A dry electrode for eeg recording. Electroencephalography and clinical neurophysiology, 90(5):376-383, 1994.  
+Takagi, Y. and Nishimoto, S. High-resolution image reconstruction with latent diffusion models from human brain activity. In CVPR, pp. 14453-14463. IEEE, 2023.  
+Tewel, Y., Gal, R., Chechik, G., and Atzmon, Y. Key-locked rank one editing for text-to-image personalization. In SIGGRAPH (Conference Paper Track), pp. 12:1-12:11. ACM, 2023a.  
+Tewel, Y., Gal, R., Chechik, G., and Atzmon, Y. Key-locked rank one editing for text-to-image personalization. In SIGGRAPH (Conference Paper Track), pp. 12:1-12:11. ACM, 2023b.  
+VanRullen, R. and Reddy, L. Reconstructing faces from fmri patterns using deep generative neural networks. CoRR, abs/1810.03856, 2018.  
+Wang, J., Chan, K. C. K., and Loy, C. C. Exploring CLIP for assessing the look and feel of images. In AAAI, pp. 2555-2563. AAAI Press, 2023.  
+Wang, S., Liu, S., Tan, Z., and Wang, X. Mindbridge: A cross-subject brain decoding framework. In CVPR, pp. 11333-11342. IEEE, 2024.  
+Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004.  
+Wasserman, N., Rotstein, N., Ganz, R., and Kimmel, R. Paint by inpaint: Learning to add image objects by removing them first. CoRR, abs/2404.18212, 2024.  
+Weiskopf, N. Real-time fmri and its application to neurofeedback. Neuroimage, 62(2):682-692, 2012.  
+Xia, W., de Charette, R., Öztireli, A. C., and Xue, J. UM-BRAE: unified multimodal brain decoding. In ECCV (7), volume 15065 of Lecture Notes in Computer Science, pp. 242-259. Springer, 2024.  
+Xu, X., Wang, Z., Zhang, E. J., Wang, K., and Shi, H. Versatile diffusion: Text, images and variations all in one diffusion model. In ICCV, pp. 7720-7731. IEEE, 2023.  
+Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., and Wen, F. Paint by example: Exemplar-based image editing with diffusion models. In CVPR, pp. 18381-18391. IEEE, 2023.
+
+Yu, M., Lin, S., Yan, H., and Ma, K. Mindpainter: Efficient brain-conditioned painting of natural images via cross-modal self-supervised learning. In AAAI, pp. 14468-14476. AAAI Press, 2025.
+
+# Appendix
+
+*Note that all the appendix figures are at the end of the text.
+
+# A. More Related Works
+
+Apart from the image customization, blending and composition introduced in the main paper, here we describe other related works on editing task. Image editing applies more precise control to the source image, allowing users to slightly modify image contents according to external instructions. Many works perform image manipulation based on GAN models (Abdal et al., 2021; Harkonen et al., 2020; Lang et al., 2021; Patashnik et al., 2021; Shen et al., 2020; Shen & Zhou, 2021) to edit images with contents and styles. Recently, the success of the diffusion model (Rombach et al., 2022; Xu et al., 2023) and large-scale pre-trained CLIP model (Radford et al., 2021b) also improves the development of editing (Avrahami et al., 2022; Hertz et al., 2023b; Ramesh et al., 2022b). Several methods integrate text conditions into the denoising process of latent diffusion to alter the resulting images (Brooks et al., 2023; Kawar et al., 2023). Liu et al. (Liu et al., 2023a) trained diffusion models to generate images that approximate the text condition, enabling edits to the source images. Hertz et al. (Hertz et al., 2023a) enhanced cross-attention mechanisms to more accurately fuse the source image with the condition, allowing for more precise modifications. Concurrently, research on image inpainting (Yang et al., 2023; Wasserman et al., 2024) has focused on utilizing manually added masks to incorporate text conditions, facilitating targeted image editing. In all, image editing focuses on more specific modifications to the reference image while customization and blending mainly deal with context integration.
+
+# B. Implementation
+
+We utilize the training set of NSD data for the Brain Representation Pre-training, and use the test set as the brain context for generation. We train the subject-wise IBT with a learning rate of  $5e^{-5}$  for 200 epochs and then adopt AdamW schedule for our Brain Embedder with a learning rate of  $3e^{-3}$  for 600 epochs. MindCustomer is implemented with the publicly available versatile diffusion. The model supports multi-modal generation with multiple flows in the latent space (of size  $4 \times 64 \times 64$ ), accompanied by VAE encoders and decoders for image and text. We fine-tune the VD for 200 epochs with a learning rate of  $5e^{-8}$  utilizing Adam schedule. Then we optimize the brain embedding for 100 epochs with a learning rate of  $1e^{-5}$ . We present the intuitively detailed setting in Table 5. The time of one image generation is about 6 minutes on a single Tesla A100 GPU. The efficiency enables us to perform single-image real-time creation for MindCustomer.
+
+Table 5. Training parameters of MindCustomer  
+
+<table><tr><td></td><td>Optimizer</td><td>Learning rate</td><td>Weight decay</td><td>Training epochs</td><td>Batch size</td><td>LR scheduler</td></tr><tr><td>IBT</td><td>AdamW</td><td>5e-5</td><td>1e-2</td><td>200</td><td>50</td><td>OneCycleLR</td></tr><tr><td>Brain Representation</td><td>AdamW</td><td>3e-3</td><td>1e-2</td><td>600</td><td>50</td><td>OneCycleLR</td></tr><tr><td>VD Finetune</td><td>Adam</td><td>5e-8</td><td>0</td><td>200</td><td>1</td><td>None</td></tr><tr><td>Embedding Opt</td><td>Adam</td><td>1e-5</td><td>0</td><td>100</td><td>1</td><td>None</td></tr></table>
+
+# C. Details of User Study
+
+We randomly selected 110 pairs of comparison images, including ours and the baseline, and divided them into 11 groups. A total of 22 participants were involved, and to minimize bias caused by individual preferences, we randomly assigned every two participants to the same group, ensuring that each image received ratings from two different individuals. By comparing ours with the baseline, participants are required to score each generated image based on the quality and consistency from 1 to 3, with 3 being better. Ultimately, we collected 220 answers for the 110 randomly selected image pairs from the 22 participants. The results broadly reflect the participants' consistent preference for our approach.
+
+Table 6. Pearson correlation coefficient and significance analysis on brain simulation of IBT with the ground truth. The range of Pearson and significance coefficient  $\alpha$  is [-1, 1] and [0, 1].  
+
+<table><tr><td></td><td>Subj1</td><td>Subj2</td><td>Subj5</td><td>Subj7</td><td>Avg</td></tr><tr><td>Pearson</td><td>0.389</td><td>0.423</td><td>0.502</td><td>0.389</td><td>0.426</td></tr><tr><td>α</td><td>1 - 5.73 × 10-8</td><td>1 - 1.170 × 10-12</td><td>1 - 7.370 × 10-19</td><td>1 - 9.250 × 10-4</td><td>0.999</td></tr></table>
+
+# D. More Experiments
+
+# D.1. Additional Results on Multi-Context Generation
+
+In Figure 11, we provide more results of multi-context generation across different subjects. And in Figure 12, we illustrate more few-shot generation results on Subj1.
+
+# D.2. Different Seeds for Generation
+
+To demonstrate the robustness of MindCustomer, we present generation results across different random seeds in Figure 13. Our method is capable of producing images of consistently high quality, with minor variations. We also recommend that users experiment with multiple random seeds to obtain results that best suit their preferences.
+
+# D.3. Discussion on Embedding Integration
+
+In this paper, we employ two different methods of embedding fusion: interpolation and concatenation. In the triplet contexts generation, we apply linear interpolation to the image and brain contexts to prevent exceeding the dimensional capacity of the VD model. In the case of dual contexts, we directly concatenate the two. To demonstrate the impact of these two fusion methods on generation quality, we input identical image and brain contexts with both fusion strategies in the absence of text context. In Figure 14, both methods successfully merge the information in a natural manner. However, we observe that, compared to concatenation, linear interpolation slightly alters the original semantic representations of the two modalities in the latent space, resulting in generated images with somewhat reduced fidelity and details. We speculate that this issue may be addressed through more sophisticated mechanisms, such as cross-attention. Additionally, optimizing this process with the currently limited brain data presents a significant challenge. Nevertheless, our method shows that even with simple fusion techniques, satisfactory results can be achieved. As research on neural data continues to evolve, we anticipate the development of more effective mechanisms to generate even higher-quality images.
+
+# D.4. Text Context for Enhanced Generation
+
+In Figure 15, we compare the impact of including versus excluding the text context on generation results using the same image and brain contexts. It is evident that the text context provides additional semantic information or more precise interaction control between the image and brain contexts. This empowers MindCustomer to better meet user needs, enabling more accurate, richer, and higher-quality image generation.
+
+# D.5. Effectiveness of Image-Brain-Translator
+
+We conduct experiments to demonstrate the effectiveness of IBT in simulating pseudo-brain signals. First, we simulate brain signals for visual stimulus images from the test set and quantitatively assess them against ground truths (GT). Following standard practices in neuroscience for brain signal simulation, we measure the similarity between the simulated and GT brain signals using the Pearson correlation coefficient. In Table 6, the average Pearson correlation coefficient across the four subjects is 0.426 (ranging from -1 to 1). According to relevant literature (Chee, 2015; Evans, 1995), it indicates a moderate positive correlation, meanwhile with a high confidence level of  $99\%$ . The results show that IBT can approximate the brain signal waves corresponding to the input images, providing a foundational prior for natural cross-modal fusion. Additionally, we visualize the brain signal simulations for four human subjects in Figure 16, further validating the method's effectiveness. Moreover, we transfer the brain signals derived from the images and directly fed them in the diffusion model to explicitly showcase the semantic accuracy of the simulated brain signals in Figure 17.
+
+Table 7. Ablations on brain decoding. We compare our method with the open-sourced SOTA methods. For a fair comparison, all methods are trained on four subjects from NSD with only one generative model.  
+
+<table><tr><td>Method</td><td>PixCorr↑</td><td>SSIM↑</td><td>AlexNet(2)↑</td><td>AlexNet(5)↑</td><td>Inception↑</td><td>CLIP↑</td></tr><tr><td>MindBridge (Wang et al., 2024)</td><td>.151</td><td>.263</td><td>87.7%</td><td>95.5%</td><td>92.4%</td><td>94.7%</td></tr><tr><td>UMBRAE (Xia et al., 2024)</td><td>.283</td><td>.328</td><td>93.9%</td><td>96.7%</td><td>91.7%</td><td>93.5%</td></tr><tr><td>Ours</td><td>.157</td><td>.289</td><td>88.3%</td><td>95.5%</td><td>92.1%</td><td>94.5%</td></tr><tr><td>wo IBT</td><td>.152</td><td>.272</td><td>87.9%</td><td>95.2%</td><td>91.5%</td><td>94.1%</td></tr></table>
+
+# D.6. Additional Results on Brain Decoding
+
+Table 7 shows the reconstruction performance, which highlights the impact of our Brain Representation Pre-training, where reconstructing brain signals with IBT improves cross-subject embeddings and achieves competitive results with SOTA methods (Wang et al., 2024; Xia et al., 2024). In Figure 18 & 19 & 20, we conduct more brain reconstruction comparisons with another SOTA MindBridge (Wang et al., 2024) and illustrations to demonstrate the effectiveness of our Brain Representation Pre-training. Moreover, we further illustrate the few-shot reconstruction capability of MindCustomer, as shown in Figure 20, where our method can also precisely reconstruct the visual stimuli under limited data.
+
+# E. Discussion and Limitation
+
+MindCustomer utilizes visual brain signals as one of the key information sources in multi-context image generation. Here we discuss our approach with traditional methods (Alaluf et al., 2023; Cao et al., 2023; Chen et al., 2024b; Gal et al., 2023b; Tewel et al., 2023b) that rely solely on image and text modalities in three aspects. (1) Data: Brain signal acquisition directly and implicitly reflects the user's intentions, without the need for additional modality conversion. In contrast, traditional methods often require users to explicitly convey their ideas through images or text, which may introduce deviations and complexities. (2) Technology: Traditional methods can directly embed prompts with the assistance of large-scale pre-trained models, such as CLIP and diffusion models. In contrast, brain-controlled image customization presents more challenges, requiring precise embedding of brain signals and bridging the implicit gaps between modalities to achieve naturally blended results. (3) Result: Since text and images can provide fine-grained semantic information, image generation (e.g. customization, editing, manipulation) based on these modalities can achieve more detailed alignment. However, current brain signals primarily record coarse information due to noise or disturbance, making them more suitable for global context blending but potentially insufficient for handling detailed features or specific instructions.
+
+Besides, prior works that focus on better decoding brain signals themselves (Scotti et al., 2023; 2024; Wang et al., 2024; Xia et al., 2024) and utilize neural data for image stylization (Chen et al., 2024a) or painting (Yu et al., 2025) have made tremendous progress in brain analysis and applications. Different from these works, MindCustomer further researches how to fuse brain signals with other contexts in image and text modalities to mask-freely create naturally blended images. Our experimental results have proven that MindCustomer is capable of effectively extracting brain signal semantics, and more importantly, it has demonstrated excellent performance in the personalized blending task. Therefore, we believe that previous works and MindCustomer can mutually inspire and complement each other, sparking more interesting and meaningful research and applications in the field of brain-computer interaction.
+
+The limitation of this work primarily lies in current brain signal acquisition and analysis, which requires further development. This includes improving the quality (e.g. Signal to Noise Ratio) and diversity (e.g. type, content) of neural data, and expanding data samples, as these factors currently constrain our precision and richness of personalized generation. Further, we believe that the collection of brain imagery data specifically for brain-controlled generation tasks may inject new momentum into the development of this field. In the future, we will also explore and acquire these data. As brain-machine interface technology continues to evolve, brain-controlled creative research will become increasingly sophisticated. We hope MindCustomer provides a critical foundation for advancing this emerging area of research.
+
+![](images/187d8fc3cccb117fe7833862e9ec2827489c4e1a547061ee5b3638d9a428ff1e.jpg)  
+Image
+
+![](images/08a2da359bbad235a29d6d1360a03c284e29a2fa2523d140325669bfb3d401f3.jpg)  
+Brain
+
+![](images/16ab51101b4afd4819d2e051617aea07999d6dbf12629512a6b900b5b652a92c.jpg)  
+MindCustomer
+
+![](images/8822827a0b20f5510eeb8de4552966fdd94aeff2962c75a55e3e2f071565041f.jpg)
+
+![](images/21f4855c23b8728453a62185f49a5e6cf2b7dc5485520c458aedd79bc0f1e396.jpg)
+
+![](images/1f938392f2124755d0e7d7e904d4c0a1a53579fc62a575eead8e6d5413c4bc7d.jpg)
+
+![](images/fa497d4b787dcdc9ad0cd47681eb6ee4a83d9254e02ef7d9f488a10e9c0da723.jpg)
+
+![](images/e2b470dde76cfa5442e87a5f09eebabb9d1c00245b0c191216627d6321364671.jpg)
+
+![](images/718ba65ef293ffa8696b2e9dfe2d3563d4c3e6c10f1e2a040af9e62eea8b25c7.jpg)
+
+![](images/1318386407450cc8370c6fc2fb8f56b4ce91d865b3af067f0d0df089f726e8c6.jpg)  
+"with sunglasses"
+
+![](images/4b636f4a3ef542cbf6acb4cdc5e7244e635f23b16b63b338cf8121c316fdbe0a.jpg)
+
+![](images/e1e391a9fad58d60dc2bc4a677b4ca88542aee41082d382e6e260f470bcd581d.jpg)
+
+![](images/e312a0bbfcee93a354e0c47a7e6bb4a2d8b2d36baed891192ff3ad33ac497e60.jpg)  
+"outside"
+
+![](images/15ba59569545d63aea75ffc8fec248adbf866519824fe1427493e349dfdc0d6a.jpg)
+
+![](images/1efa754f16296eafdf9ce34b6c43f7dd522846c3d949cf98b2f61b01c14a2739.jpg)
+
+![](images/3ed6b14e4e5241951dad3e7168183a35336b1f01800f9d698bd03e02169bad6e.jpg)  
+"turn nose appearance"
+
+![](images/a4cb23ec0b15b23e00afc6389bd3b580daa1503717866f1f28cbf93a6b19ba43.jpg)  
+Figure 11. More brain-blended multi-context generation results.
+
+![](images/8c643c3739391bd791f98a4870a126ab2aa599bcdbc09c6f05a619be4e3d920e.jpg)
+
+![](images/cebfbfe6ee14fda6910ca36f5acf5fdf2805ea4338de82406996b16d6069f77c.jpg)  
+Image
+
+![](images/ffaab469d9fbfd887fe2c0c45f941f08d53b1c4cc04cb87f720f06e5631d9fbc.jpg)  
+Brain
+
+![](images/9da65ea839af934f24059bf67fe179077d0c2d4daaaf11c1d9b8e9221ab300bb.jpg)  
+MindCustomer
+
+![](images/65d6e07cd9c5a4ed32883c770dd9739397ba51472d1d397256a7e3c059164bba.jpg)
+
+![](images/7be5a05d7aa225d4ac303581a0908763b73c415fdff43d11cc0087425f331318.jpg)
+
+![](images/081ece2f7126ae0f2f7e76337dcafda53bfc2cd2e393e04ea7617a744e64d817.jpg)
+
+![](images/f96b8e8825ae48987bc2579157795757d95166354ad95f20cd33709592a6f886.jpg)
+
+![](images/d18a878db37588f39861ebf2feadda4ffb02b28870b2058735ac984b611c86e5.jpg)
+
+![](images/1582a775e65ac237b6c3a0430d094d6b23a1ea36d23ae922dc660f54db31af5e.jpg)
+
+![](images/a5f992b613d5ea22f9cbb83645af23f65c308debd83e80dd11f36d978c4f576f.jpg)  
+"riding"
+
+![](images/56730bad7dcae8d49083d6b40e4979224bc47adf9582437c7bc798f24ac2dfa7.jpg)
+
+![](images/d8ab4a36e5c085dfa8783c02998e61c4b1b85a950c8e9363cf934e9edf4e8706.jpg)
+
+![](images/5860ce33f54c380b2c4be92c32147ef0db5cedafb32ca76a5b8825abc6ab059d.jpg)  
+"made of"
+
+![](images/0bf8640f8592d19d9acc5644f662f0fc07bdf29113890fd5bd255f12ef344cb8.jpg)
+
+![](images/5c16955d5d2c4bd363aa742d2172cbf973d20ac0bcbcf5fcacb32a6e5127eb6c.jpg)
+
+![](images/27e8d42de4bf8ceef2b90a1d17bfb20de7310507dcd0aa0a1baa6448f0a6ce51.jpg)  
+"Eiffel Tower in the behind"
+
+![](images/4119a8db4704ec358289e59b0dc640d3ec3872f80dcd886a885c15a31f50ef34.jpg)
+
+![](images/4672d2f530ac49a1c9a88bce93882fd5ed86e159502e6eaa4edb643f932be9ae.jpg)
+
+![](images/b6c5c3178fa5d51559a04047781590941bf58add4df41a52608fda006edf9a6b.jpg)  
+Figure 12. More few-shot generation results on Subj1. We only use  $10\%$ ,  $20\%$ , and  $40\%$  of the subject's complete training data for few-shot learning.
+
+![](images/45d5d784b4c2e074537fa5e3434dea61c77dfe8eb8e019f4aba69c51b6af5e1e.jpg)  
+Figure 13. Different seeds for generation.
+
+![](images/87adec65b28f8e45199cd8e485cb642a37aeaaafb90592f8c2939e57868ee78f.jpg)  
+Figure 14. Comparison on the methods of brain and image embedding integration: interpolation and concatenation.
+
+![](images/7f32fdfd931694f0eef617960f53685cdb871753556d46599749443b5d234c41.jpg)  
+Image
+
+![](images/271260ff8f719118113feecefd109610b3756ed76406c8ef0bca97b4255b3078.jpg)  
+Brain
+
+![](images/882021eca62a4c97f0f26d7c684314df07c05e2f5a1b3cb1e4037a1e7469cbb4.jpg)  
+Dual Context Gen
+
+"play"
+
+![](images/bd65fc6495be39d56a609965a264bdd06d6be8048d2609a2d3ae2b90861386e9.jpg)  
+Triplet Context Gen
+
+"made of"
+
+![](images/4c5ffdc5577233549e8f1609475d9eb49f8953fc5fc206656fe96543e2a42a01.jpg)
+
+![](images/648cd5a354024e18235f65dd6ab03dc8baa736ab30bfa4184b89c46bfd1f096a.jpg)
+
+![](images/4c4fb1cb2ac4096ae56ee831010d5d0804b96b46f4b12ed82e257e5f1dac0240.jpg)
+
+![](images/8368b950ce242164be76c146408c8b8f6de1544ccafeee8425ce74f25234e9c5.jpg)
+
+"in the snow"
+
+![](images/b09f0bcc72994dc58143769a311f9921b5507781f89158b2bf4dbe7282244a3e.jpg)
+
+![](images/ce12152b7962ad2cdd5fe77afc5404af00ae0d793f5236d11d726cc54c70ceb0.jpg)
+
+![](images/98512772e475d9e53eef85e746c1912d31c7262cf1d791eb11503ef30fff635e.jpg)
+
+![](images/11fafce979437b8144a1be357ea01e4b0da20d11d7f07cd3400c208ea28579dd.jpg)
+
+"bus in the background"
+
+![](images/6ee4c7cbe85ac66c0fae39aae6a2f73151063dfc3e1b0d9e04ca53960ec1842e.jpg)
+
+![](images/b6964962d24c9a9e96aecc89c802b537903ec6de6b600795e66147f49c22adcd.jpg)  
+Figure 15. Generation results w / wo text contexts. It demonstrates that the text context provides additional semantic information or more precise control over the image and brain contexts.
+
+![](images/8bc97e32f858abdf5daf067f54918f3a6e21dc7ee00943baabbf69fe1e93aafe.jpg)
+
+![](images/8c4b14d7b006ef3efe0587b202c6d208924c5e3fbd62535532a7d1ebc5c9f849.jpg)
+
+![](images/ddd5f45f50092905a3fefbbe385d8c9cecba401416ec432d9e43188799cd4b05.jpg)  
+Figure 16. Visualization of the predicted brain voxel by IBT with its ground truth (GT). It illustrates voxels in brain signals with their corresponding feature values. Here we randomly choose three examples for each subject, and report the difference error between the prediction and the ground truth. We simply compute the absolute difference in value for each voxel, and then take the average of the absolute differences across all voxels, symbolized as Avg Diff. Note that the range of Avg Diff is [0, 2], lower is better. It is observed that the subject-wise IBT can effectively simulate brain signals with low differences to GT. It facilitates the transferring accuracy of image features to brain modality in our method.
+
+![](images/a18c6a0058aa4243583d7b314aaa95665d15177f9b59bd3991ff0fb31a92cc5a.jpg)  
+Visual Stimuli
+
+![](images/762652cb2fd6568e576d6ef70e9194ef59b7301e72972921dcdcd660ea5dfe33.jpg)  
+IBT Subj1  
+Avg Diff: 0.2
+
+![](images/5f87bcaee94e64b95a9fd2ddd7ce6597d4b8d05eef4c9f0e04c4b1def71fe584.jpg)  
+IBT Subj2  
+Avg Diff: 0.3
+
+![](images/49f41e250f2f20522a7703850cc5d5642dbd8814712fd21f6fd73434ca6424ee.jpg)  
+IBT Subj5  
+Avg Diff: 0.15
+
+![](images/0e325a2dbb71c89721cd19559fe77975215ee2b967a5d19ddc41a43d1776f647.jpg)  
+IBT Subj7  
+Avg Diff: 0.1
+
+![](images/01bfd431a6ddca18bfc52482cbe976e077860b83efe21605e4d89be013724c97.jpg)
+
+![](images/eb7956bcbcd9ec80f3e1c0366260f5a73854b2cb1ab930c21f8cb0a6195fb6da.jpg)  
+Avg Diff: 0.15
+
+![](images/8ab008871e5defa3c3ff81dee08164aea3118c670e1f96fc3a3b2440d23ea5be.jpg)  
+Avg Diff: 0.2
+
+![](images/07812bb2406752e4f37edb5d2a00adc391bba63ba0bc6b8374ef088acd580bbf.jpg)  
+Avg Diff: 0.2
+
+![](images/dbf888e1d830dd2c817de215e8856b423db9aa7d0cb360e026ef7398a322278c.jpg)  
+Avg Diff: 0.25
+
+![](images/4bb878badb756fe181c5882d06ee1467acf7cde041e230295b5111a279036be7.jpg)
+
+![](images/60948af4d04623911446b82ff3bee7518352848b5852fc7d45faba0981f5cbf5.jpg)  
+Avg Diff: 0.15
+
+![](images/bbc78df12630ec1795b3302832c57c606fbb2b32b6442d0c02b4e1f892bbe3fa.jpg)  
+Avg Diff: 0.15
+
+![](images/3eb0845af43d42821481121d67032fa21ca2be041306547686fb2940a8324db9.jpg)  
+Avg Diff: 0.3
+
+![](images/de70d735ad82c45e38fe3b5011e7eeaa71c34eedf28b9bbff0e2d2f64c209474.jpg)  
+Avg Diff: 0.2
+
+![](images/df9832612efb8607b8d6c4a78fa9fa7f9f00829b8aba16a8ad457b23248a7692.jpg)
+
+![](images/ae1dcd3d0dbb2cdd437d767f5d7df4ea6fdfe5520fa85938b700715d3e14085c.jpg)  
+Avg Diff: 0.3
+
+![](images/a2377cba951fe49899f687a1aefa9d0f4155ec2a18d71c7a6ccbfbd0f7ef2dbf.jpg)  
+Avg Diff: 0.2
+
+![](images/53013dc5d897e8ffede7ae9a95c7a9bab1a1a7dd6311e8d53396821b388ed13d.jpg)  
+Avg Diff: 0.1
+
+![](images/5e83eb0feb0ce2f58a644ec08a97705c8de6fafcdd04f2086af47126f6173be2.jpg)  
+Avg Diff: 0.2
+
+![](images/6d053ffd2154b67ea0d57fc14787019611a5cdaa025b3fdda20e6b7b9c04fdc5.jpg)  
+Figure 17. We utilize the predicted brain voxels of IBT for direct reconstruction to explicitly demonstrate the effectiveness of accurate semantics. Same with Figure 16, here we also report the Avg Diff of brain voxels between the prediction and the ground truth.
+
+![](images/862da0caecee5131977a3cabdc0c4a5d3bb1c2d0faa71d6800d4db596aed26b0.jpg)  
+Avg Diff: 0.3
+
+![](images/cba3020544468afe41a9b132de36c8c1e995090172c65842e9fabe9c6bd592f4.jpg)  
+Avg Diff: 0.2
+
+![](images/4dd1364158a2315351b78b84597b2cc646832fadc1baf8c45c51b0e5f2423d37.jpg)  
+Avg Diff: 0.2
+
+![](images/1ab8a33858f79bbd2806b40a325005bf67b624ce2a63aaca99a3d1fda171795d.jpg)  
+Avg Diff: 0.3
+
+![](images/8edb2197c5908d3a19ee35daed247376839a1c3528a424e3b865344d86e25061.jpg)  
+Figure 18. Comparison on the cross-subject brain reconstruction. We compare our method with the SOTA MindBridge (Wang et al., 2024) to show its competitive performance of MindCustomer on the brain decoding task.
+
+![](images/96d62b7eab4d4b9f467edce1ba5da47f05f226a385ba76daff914b3cccf78d04.jpg)  
+Figure 19. More cross-subject brain reconstruction results of MindCustomer.
+
+![](images/41896feedb476909224bdf0f889c37be745c639d3fe386ed48d85cd97119e6d3.jpg)  
+Visual Stimuli
+
+![](images/5fa9f0bf9934be3d419501330a500a73bf295f65b4f05881e8223f5d8744337a.jpg)  
+10%
+
+![](images/9b819c9e7b44fe832ef7b1c169ebd50ed75f2abcf105aa751163b41cde5c124f.jpg)  
+20%
+
+![](images/48ee8a8c3f2cac09587f53d6574fb0e1bb4d784fd8874e0f807910e9844d770a.jpg)  
+40%
+
+![](images/dc0bf3fa88b273105c8f0375afb9c26dee440ee08cd039f51f7b97cd93df7fc6.jpg)  
+100%
+
+![](images/e1225c26826cf26d5e12c91a5a8b1a8413baeaf99cb1da8392cf59a589c8bbd6.jpg)
+
+![](images/64ce0bbda679568f67516630ceebdd90d65d60e4d9861a46da5c93cede0f533f.jpg)
+
+![](images/ebff872ff6f2801a493d924572168c28e39cc45cebeb2adf444160a2053f2585.jpg)
+
+![](images/5f7d79f01e418f2459f79f22149bb32af82c12895f179f48ec7b6775b3d69edb.jpg)
+
+![](images/56cce85c71a9b98e2b3e968541c91f2358843c323bf558632739054ad2ed752e.jpg)
+
+![](images/b2c44688f235cdd657d462eaba27f711259638a117b3f0c1df9ddb9e06d50b55.jpg)
+
+![](images/8b8b3c2162c1906d53c8734eb021427cc2ab0e191119bb374a52f3ef6033ba23.jpg)
+
+![](images/bf0954f7b48ba2f4c564221b247112d4d90ed5dbd3896e20e414b6838172f858.jpg)
+
+![](images/8b1e010b1e85e7758be0261f79c500b36a9f1c30c3a44969bb06f8a1d3aadeb8.jpg)
+
+![](images/4ce2ef69a624282232cd61c16e551611eb92c3f610e348c8e6e341310e9e8901.jpg)
+
+![](images/bc507e2355b03fa0e04fb5761ea419caee551742f09c4875d54e7cc3a00371b1.jpg)  
+Figure 20. Few-shot brain reconstruction results on Subj1. Compared to full training (100%), our method can also roughly decode the brain signals with limited data. No more than  $40\%$ , MindCustomer is already capable of precise reconstruction.
+
+![](images/d7182320a93e89e4ef83aa6646e19ff2e8bad69f50768045fb742e0251da23d2.jpg)
+
+![](images/2140326837fd883392dce775f8ab4e7ba3f49d899333b6c19062f2d6569d9863.jpg)
+
+![](images/f6ac494aa6c30e590a7985f1b01fd6dfcf166c3df2f411f29d356cbd9c7f3b45.jpg)
+
+![](images/31f6d009d8e0177c121869e9ec8fca1f86605f5e2818ddc202af7e9695e621a8.jpg)

2025/MindCustomer_ Multi-Context Image Generation Blended with Brain Signal/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71ed5d151b8a26e2de9211397a55ab2cfef506b69bb2ed175b9906ff0ab6f2a4
+size 3631197

2025/MindLLM_ A Subject-Agnostic and Versatile Model for fMRI-to-text Decoding/33b83fb9-ebc9-4760-81ee-af53162ee500_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd8d7d47ae381e34d681d239e0b735bc3962564a6397c7b9b20b3ca7ae9fc3c1
+size 12649073

2025/MindLLM_ A Subject-Agnostic and Versatile Model for fMRI-to-text Decoding/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0b52b1051d594236fcd80146b4975c3a44e0300519abb329baacd8580c05c6c
+size 1856547

2025/Minerva_ A Programmable Memory Test Benchmark for Language Models/08a6cf6a-c318-43be-96ff-f6dc835f785d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53a047e545891d6e3ca2bf3a6d628e4012697586cc05cb0c2196fbe6ba9dfb3c
+size 867084

2025/Minerva_ A Programmable Memory Test Benchmark for Language Models/full.md

@@ -0,0 +1,522 @@
+# Minerva: A Programmable Memory Test Benchmark for Language Models
+
+Menglin Xia  $^{*1}$  Victor Ruhle  $^{1}$  Saravan Rajmohan  $^{1}$  Reza Shokri  $^{*2}$
+
+# Abstract
+
+How effectively can LLM-based AI assistants utilize their memory (context) to perform various tasks? Traditional data benchmarks, which are often manually crafted, suffer from several limitations: they are static, susceptible to overfitting, difficult to interpret, and lack actionable insights-failing to pinpoint the specific capabilities a model lacks when it does not pass a test. In this paper, we present a framework for automatically generating a comprehensive set of tests to evaluate models' abilities to use their memory effectively. Our framework extends the range of capability tests beyond the commonly explored (passkey, key-value, needle in the haystack) search, a dominant focus in the literature. Specifically, we evaluate models on atomic tasks such as searching, recalling, editing, matching, comparing information in context memory, performing basic operations when inputs are structured into distinct blocks, and maintaining state while operating on memory, simulating real-world data. Additionally, we design composite tests to investigate the models' ability to perform more complex, integrated tasks. Our benchmark enables an interpretable, detailed assessment of memory capabilities of LLMs.
+
+# 1. Introduction
+
+What capabilities should we expect from AI assistants? The AI assistants are provided with a (large) input context containing all the available information that is potentially relevant to the user's request (e.g., all the prior emails, messages, and confirmed calendar events). This input, commonly referred to as the context (for the LLM), encapsulates what the AI assistant knows about the world in which it is tasked to operate. This representation of the world, expressed in natural language, functions as the model's memory. In this
+
+*Equal contribution  ${}^{1}\mathrm{M}{365}$  Research,Microsoft  ${}^{2}$  National University of Singapore. Correspondence to: Menglin Xia <molyxia@microsoft.com>, Reza Shokri <reza@comp.nus.edu.sg>.
+
+Proceedings of the  $42^{nd}$  International Conference on Machine Learning, Vancouver, Canada. PMLR 267, 2025. Copyright 2025 by the author(s).
+
+paper, we address a fundamental question that is critical to improving AI assistants: What specific capabilities do large language models demonstrate in utilizing their memory?
+
+One common approach to testing models is through data benchmarks. However, evaluating model capabilities using static data benchmarks—based on some user queries, their data, and expected outcomes—can be costly, imprecise, and lacks scalability. Additionally, performing tasks in realistic scenarios often requires multiple capabilities, making it challenging to identify which specific capability a model lacks when it fails a data benchmark. This limitation reduces the effectiveness of these benchmarks in designing better models. Moreover, blindly optimizing models to improve on such benchmarks risks overfitting, often rendering the data benchmarks obsolete over time.
+
+To address some of these concerns, recently, there have been many attempts to test models using automatically generated benchmarks, but these efforts have primarily focused on evaluating basic search capabilities (e.g., passkey or key-value search) in long-contexts (Kamradt, 2023; Liu et al., 2024; Zhang et al., 2024; Anthropic, 2024; Wu et al., 2024; Li et al., 2024; Hsieh et al., 2024). In this paper, we go beyond simple search tasks and introduce a framework to test a comprehensive range of memory-related capabilities in LLMs. Note that, we deliberately avoid conflating the memory usage capabilities with the complex reasoning abilities (e.g., complex mathematical or logical reasoning) in language models, as the latter is a separate skill that is currently the focus of extensive study (Clark et al., 2018; Cobbe et al., 2021; Hendrycks et al., 2021b; Suzgun et al., 2022).
+
+What are memory-usage capabilities? We define these as the abilities to retrieve relevant information, compose it for the instructed task, and recall key details when synthesizing the output. This process also involves creating associations between the instruction and the stored information, as well as among different parts of the memory itself. Without extracting these relationships, the memory remains flat and formless, rendering it less useful. Consequently, the model must be able to recognize differences, identify similarities, and take appropriate actions based on them. We design a series of atomic tests aimed at evaluating each of these individual capabilities in isolation, to the extent that isolating such capabilities is possible.
+
+To evaluate the more complex scenarios in memory usage, we construct composite tests that reflect real-world scenarios, where memory is divided into multiple compartments (i.e., information relevant to distinct contexts). The model is expected to recognize the boundaries of these compartments, trace the knowledge contained within them, and perform operations such as information retrieval, memory association, and other tasks while respecting these boundaries. The complexity increases further when there is interaction between compartments. In such cases, information must flow across boundaries—for example, when stories about two parallel events converge at a particular moment, when interactions occur between different events, or when information known to certain entities in memory is shared with others. Examples include AI assistants managing calendar events, tracking financial transactions, or suggesting medical diagnoses. Handling these scenarios is highly challenging, yet it is essential for AI assistants to achieve practical and reliable performance. The core challenge is due to the fact that context memory, as provided to the model, flattens data from multiple parallel and potentially interrelated memory compartments. This requires the model to disentangle the content by leveraging the available labels and clues, and tracking the state of relevant information throughout the memory scanning, while performing the task.
+
+Table 1 presents an overview of the types of memory tests included in our benchmark. For each test, we use efficient parametric programs to generate randomized test cases. We run a comprehensive evaluation of several major open-source and black-box models (e.g., GPT-4(o), Cohere, Gemma, LLaMA, Mistral, Phi). Our experimental results show that while models perform relatively well on simple search tasks, they exhibit significant disparities across context utilization capabilities even at a context length of 4k tokens. This indicates that strong performance in basic retrieval does not necessarily translate to other context processing abilities. Our framework goes beyond search-based tests by incorporating atomic tests that pinpoint distinct capabilities, providing a more nuanced picture of the strengths and weaknesses of models in context processing. Moreover, composite tests, which combine multiple atomic capabilities, resulted in substantial performance drops for all models. These tests present the limitations of current models and provide valuable insights for guiding future model training and development.
+
+The code and data will be available at https://github.com/microsoft/minerva_memory_test.
+
+# 2. Benchmark for Memory Tests
+
+We define the entirety of context data available to a large language model (LLM) as its memory. Users of an AI assistant leveraging the LLM can instruct the model to parse
+
+this memory and execute potentially complex retrieval tasks. Accordingly, we expect the AI assistant to demonstrate specific capabilities in memory utilization, including accurate retrieval, effective synthesis of relevant information, and adaptability to evolving context. These capabilities include:
+
+- Information Retrieval and Localization: The ability to efficiently locate, search for, and extract relevant information from the memory (i.e., input data) based on specific instructions or user queries.  
+- Processing and Basic Reasoning: The capability to perform modifications, computations, and logical operations on the input data, including identifying patterns, recognizing repetitions, and understanding relationships within the memory.  
+- Content Transfer and Synthesis: The ability to copy, rephrase, or generate synthesized output by integrating both original and modified elements from the input.  
+- Structural Awareness and Organization: The capacity to interpret the spatial, structural, or organizational layout of the memory, such as distinguishing labeled fragments of text, sets, lists, or hierarchical structures.
+
+In our benchmark, we focus on isolating the atomic memory-related capabilities of LLMs. Success or failure in these tests provides a clear and interpretable assessment of the strengths and limitations of the models. We design multiple atomic tests to measure fundamental skills without interference from other factors. These tests are simple, targeted, and structured to assess specific abilities with clarity. Our benchmark includes the existing basic tests, notably the needle-in-the-haystack tests and its variations, but it goes beyond the search methods and includes a diverse set of atomic capabilities. In our benchmark, we focus on some fundamental capabilities: search, recall and edit, match and compare, spot the differences, compute on sets and lists, and stateful processing.
+
+We also develop composite tests to evaluate how effectively models can perform more complex, integrated tasks. These tests assess the integration of multiple atomic capabilities to simulate real-world scenarios. By combining elements such as retrieval, reasoning, synthesis, and structural awareness, composite tests measure how well an AI assistant can coordinate different skills to execute complex operations. We provide two class of composite tests: processing data blocks, and composite-state tracking (theory of mind). Our objective here is to test the composition of various atomic operations and the ability to interpret segments of data (e.g., messages or paragraphs associated with a particular person or topic). These composite tests evaluate if the model can make sense of the "spatial" structure of the memory, and also keep track of its "temporal" changes (e.g., information
+
+that gets updated across many emails). This can become particularly challenging for the current architecture of major LLMs because the context has a flat structure.
+
+Table 1 presents the list and description of representative tests in each category. Appendix A presents the exact templates for all our tests. Our benchmark differs from traditional data benchmarks by allowing the generation of fresh, randomized test cases for each category. Each test acts as a programmable script that measures the model's capability while adjusting the hyperparameters that influence test difficulty. The programmable tests also enable us composing them easily, which is one of the key advantages of our benchmark. New categories, and new tests, can be easily added to this framework enabling a more diverse set of tests.
+
+# 3. Evaluation
+
+# 3.1. Experimental Setup
+
+We use the proposed framework to evaluate nine widely used language models on a fixed snapshot of 1110 randomly generated test samples. For all tests, we fixed the context length to 4k tokens, except in the Stateful Processing category, where the context length depends on the number of operation steps. We set the number of steps as 200 for quantity state and 100 for set state, corresponding to an approximate context length of 1.5k tokens. For evaluation, we use exact match accuracy for binary tasks, ROUGE-L(Lin, 2004) for tests that require sequence overlap measurement, and Jaccard similarity (Jaccard, 1901) for set overlap. Further details on the number of examples, hyperparameter configurations, and evaluation metrics for the tests are provided in Appendices B and C.
+
+The evaluated models are divided into two groups:
+
+Black-box models: GPT-4-turbo, GPT-4o, GPT-4o-mini, and Cohere-command-rplus.
+
+Open-source models: Mistral-7b-instruct-v02, Phi-3-small-128k-instruct (7B), LLaMA-3.1-8b-instruct, Gemma-2-9b, and Phi-3-medium-128k-instruct (14B).
+
+We set the max output token to 4096, temperature to 0, and top_p to 1 for all model inference.
+
+# 3.2. Model Performance Overview
+
+Figure 1 summarizes the overall performance of the evaluated models on the memory test snapshot within 4k context length. Notably, this context length is usually considered short for context utilization benchmarks, and many models are expected to perform perfectly at this length. However, our evaluation reveals significant disparities in performance across the capabilities, even within this manageable context length. Overall, the GPT-4-turbo/GPT-4o
+
+models show stronger all-around performance across the capabilities. In contrast, other models excel at the search task but struggle significantly in other areas, leading to a widening performance gap compared to stronger models. This is especially evident in the Stateful Processing tasks, where models exhibit steep performance drops. Even within the GPT-4(o) models, there were noticeable variations in performance across different tasks, despite them being the best-performing models. This suggests that strong performance in simple retrieval tasks does not imply effective context processing, highlighting that using NIAH-like tests alone for evaluating context utilization is not sufficient to capture the full spectrum of model capabilities. Our framework instead reveals significant variability in performance across distinct capability categories, offering a more nuanced understanding of model limitations.
+
+The following sections analyze each test type in detail, highlighting key insights from the evaluations.
+
+# 3.3. Analysis on Atomic Tests
+
+Search All models performed relatively well on Search tasks, which is unsurprising given the 4k context length. However, even at this length, model performance varied significantly depending on the specific search type (see Table 2). For example, in the binary String Search task, models handled individual word searches well but struggled with subsequence searches, where queries consisted of multi-word sequences. The performance drop can be attributed to two factors: (1) length of query affects the difficulty of precise memory access; (2) negative samples are created by replacing a single word in present subsequences, making absent longer subsequence more distracting.
+
+Figure 2 further analyzes subsequence search performance for GPT-4o, Mistral, and Phi-3-medium. These models exhibit distinct error patterns as the length of the subsequence increases: GPT-4o has no false negative errors (it never misses a present subsequence) but makes more false positive errors as the subsequence length grows, suggesting it overestimates presence in more ambiguous cases. Mistral also makes no false negative errors but exhibits a decreasing false positive rate, implying it struggles more with shorter distractors. Phi-3-medium, in contrast, makes few false positive errors (rarely identifies an absent sequence as present), but struggles more with false negatives, indicating a general tendency to deny presence. These differing patterns suggest that the models may employ different search strategies, affecting their susceptibility to different types of errors.
+
+For Batch Search and Key-Value Search tasks (analogous to multi-NIAH and NIAH, respectively), models like Mistral, Phi-3, and Cohore show a notable performance drop, revealing their limitations in handling multiple memory accesses effectively.
+
+Search  
+
+<table><tr><td>String search (binary test)</td><td>Is the string x in memory? x could be a word or a sequence.</td></tr><tr><td>Key-value search</td><td>What is the word or phrase that is paired with keyword x?</td></tr><tr><td>Batch search</td><td>For each of the keywords in the batch x1, x2, ···, xk, perform the search and return the batch of corresponding responses.</td></tr></table>
+
+Recall and Edit  
+
+<table><tr><td>Snapshot</td><td>Share a snapshot of the entire memory (as a verbatim copy).</td></tr><tr><td>Replace all</td><td>Share the entire memory after replacing all occurrences of x with y.</td></tr><tr><td>Overwrite positions</td><td>Share the entire memory after overwriting the words that are on particular positions (e.g., every kth word) with y.</td></tr><tr><td>Functional updates</td><td>Update every x with the output of a function f(x).</td></tr></table>
+
+Match and Compare  
+
+<table><tr><td>Compare positions (binary test)</td><td>Does x appear before y?</td></tr><tr><td>Find duplicates</td><td>Which word/string has duplicates in memory?</td></tr><tr><td>Count</td><td>How many times is x repeated in memory?</td></tr><tr><td>Check association (binary test)</td><td>Check if both x and y are associated with the same tag in memory (assuming every term is associated with a tag).</td></tr></table>
+
+Spot the Differences  
+
+<table><tr><td>Compare two lists</td><td>Give two lists X and Y of the same length (e.g., same number of words), report the difference (i.e., X - Y).</td></tr><tr><td>Identify the odd group</td><td>Given multiple sets, identify which one is different (assuming n - 1 identical sets with shuffled elements, and 1 set with some differences).</td></tr><tr><td>Patch the difference</td><td>A sequence of words is repeated multiple times, and then there is a partial sequence. What is the next kth element in the sequence?</td></tr></table>
+
+Compute on Sets and Lists  
+
+<table><tr><td>Group membership</td><td>Given k sets, identify which set includes x.</td></tr><tr><td>Group association</td><td>Check if x and y belong to the same set.</td></tr><tr><td>Iterate</td><td>Given k lists, return the last element in each list.</td></tr></table>
+
+Stateful Processing  
+
+<table><tr><td>Quantity</td><td>Keep track of the total quantity of items, based on a sequence of addition and subtraction operations (e.g., “add 10, subtract 2, add 7, ...”).</td></tr><tr><td>Set</td><td>Keep track of the items in a set, based on a sequence of addition and removal operation (e.g., “add apple, pear; add orange; remove apple; add lime, ...”).</td></tr></table>
+
+Processing Data Blocks  
+
+<table><tr><td>Search, recall, and edit</td><td>The input contains alternating labeled lists of elements (e.g., “L1: a, b, c; L2: h, f, i; L1: d, z, k; ...”). For a given list label (e.g., L1 or L2) and a specified element within that list, return all the elements that appear after that specified element in the same list.</td></tr></table>
+
+Composite-State Tracking (Theory of Mind)  
+
+<table><tr><td>State tracking across data blocks</td><td>Perform “Stateful Processing” for multiple agents, and report the final set state for each agent. The input provides a list of operations by k agents over time, including both independent actions (add/remove) and interactive actions (swap) (e.g., “Alice: add apple, pear, remove orange, add banana; Bob: add peach, berry, remove kiwi; Charley: add lime; Bob: remove peach, swap berry with Alice for banana; ...”).</td></tr></table>
+
+Table 1. List of memory tests. We divide the tests into different categories based on the core expected capability for passing the test. Most the initial tests are atomic, i.e., the expected capability cannot be broken down into other meaningful capabilities. The tests at the bottom of the list are composite tests and require the model to have multiple atomic capabilities at the same time in order to succeed.
+
+Recall and Edit Figure 3 presents the results for the Recall and Edit tasks. While models performed well on ba
+
+sic recall (Snapshot), their performance dropped sharply when tasked with making regular edits. A closer analysis
+
+![](images/9510c684a11f111076a3cb460c5797351f5544dd8af42dbcab9663eb573e38a1.jpg)  
+(a) Performance of the black-box models.
+
+![](images/e4cae25f2a050f3779853160bb5b7ed62ea2cc9a8b0650b965419bd308dcec5c.jpg)  
+(b) Performance of the open-source models.
+
+<table><tr><td>Models</td><td>Word</td><td>Subsequence</td><td>Key-value</td><td>Batch</td></tr><tr><td>gpt-4-turbo</td><td>0.94</td><td>0.94 (-0.00)</td><td>1.00</td><td>1.00 (-0.00)</td></tr><tr><td>gpt-4o</td><td>1.00</td><td>0.82 (-0.18)</td><td>1.00</td><td>1.00 (-0.00)</td></tr><tr><td>gpt-4o-mini</td><td>0.98</td><td>0.64 (-0.34)</td><td>1.00</td><td>0.96 (-0.04)</td></tr><tr><td>cohere-command-rplus</td><td>1.00</td><td>0.85 (-0.15)</td><td>0.98</td><td>0.87 (-0.11)</td></tr><tr><td>mistral-7b</td><td>0.78</td><td>0.80 (+0.02)</td><td>0.92</td><td>0.47 (-0.45)</td></tr><tr><td>phi-3-small</td><td>0.94</td><td>0.84 (-0.10)</td><td>0.94</td><td>0.77 (-0.17)</td></tr><tr><td>phi-3-medium</td><td>1.00</td><td>0.55 (-0.45)</td><td>1.00</td><td>0.72 (-0.28)</td></tr><tr><td>gemma-2-9b</td><td>1.00</td><td>0.60 (-0.40)</td><td>1.00</td><td>1.00 (-0.00)</td></tr><tr><td>llama-3.1-8b</td><td>1.00</td><td>0.57 (-0.43)</td><td>1.00</td><td>0.99 (-0.01)</td></tr></table>
+
+Table 2. Results for the Search tasks. The four columns represent: String Search (with word), String Search(with subsequence), Key-value Search, and Batch Search. Numbers in parentheses indicate comparative performance differences between String Search (with subsequence vs. word) and Batch Search vs. Key-Value Search.  
+
+<table><tr><td>Model</td><td>String Search (word)</td><td>Snapshot</td></tr><tr><td>gpt-4-turbo</td><td>1.00 (0.06)</td><td>1.00 (0.04)</td></tr><tr><td>gpt-4o</td><td>1.00 (0.00)</td><td>1.00 (0.00)</td></tr><tr><td>gpt-4o-mini</td><td>0.94 (-0.04)</td><td>1.00 (0.00)</td></tr><tr><td>cohere</td><td>1.00 (0.00)</td><td>1.00 (0.26)</td></tr><tr><td>mistral-7b</td><td>1.00 (0.22)</td><td>0.96 (0.00)</td></tr><tr><td>phi-3-small</td><td>1.00 (0.06)</td><td>0.99 (0.04)</td></tr><tr><td>phi-3-medium</td><td>0.98 (-0.02)</td><td>0.87 (-0.09)</td></tr><tr><td>gemma-2-9b</td><td>0.96 (-0.04)</td><td>0.96 (0.05)</td></tr><tr><td>llama-3.1-8b</td><td>0.98 (-0.02)</td><td>1.00 (0.00)</td></tr></table>
+
+Table 3. Ablation study with gibberish context.
+
+of the generated outputs reveals that models struggled with maintaining coherence during edits, often getting trapped in repetitive word loops. For the Functional Update task, we deliberately selected simple numerical updates, such as "Subtract 1 from every number," to ensure the edits were within the models' capabilities. Nevertheless, when comparing performance on Snapshot (with numbers) to Functional Updates, all models exhibited a steep decline, especially for smaller ones. Analysis of generated outputs revealed
+
+![](images/475300e91b83ac61f553ede29d0afe5f0935bc666325d2c06dbdebabdcbe8c5d.jpg)  
+Figure 1. Overall performance of nine models on a snapshot within 4k context length of Minerva.  
+Figure 2. Analysis on String Search (with subsequence) across increasing subsequence lengths. This figure examines the behavior of models on positive samples (where the subsequence is present) and negative samples (where the subsequence is absent).
+
+that these models frequently deviated from instructions over longer sequences, suggesting difficulties in maintaining consistent rule applications over extended contexts.
+
+Additionally, we conducted a separate ablation study on Snapshot and String Search. In this study, we replaced meaningful words in the context with gibberish tokens consisting of randomly generated alphabetical characters. As shown in Table 3, performance remained largely unchanged, suggesting that semantic meaning was not a significant distractor in these tasks.
+
+Match and Compare As shown in Figure 4, model performance in the Match and Compare tasks was relatively consistent across different model sizes. Given that counting is a well-known weakness in LLMs, it is unsurprising that all models struggled significantly with the counting task, though GPT models performed slightly better than others. However, models generally succeeded in identifying the
+
+<table><tr><td>Model 
+(Compared against)</td><td>Group membership 
+(Sub-string search)</td><td>Group association 
+(Group membership)</td><td>Group assoc. (alternating) 
+(Group association)</td><td>Iterate 
+(Iterate (last))</td></tr><tr><td>gpt-4-turbo</td><td>0.96 (0.02)</td><td>0.75 (-0.21)</td><td>0.68 (-0.07)</td><td>0.83 (-0.17)</td></tr><tr><td>gpt-4o</td><td>0.98 (-0.02)</td><td>0.65 (-0.33)</td><td>0.52 (-0.13)</td><td>0.86 (-0.14)</td></tr><tr><td>gpt-4o-mini</td><td>0.96 (-0.02)</td><td>0.68 (-0.28)</td><td>0.52 (-0.16)</td><td>0.67 (-0.33)</td></tr><tr><td>cohere-command-rplus</td><td>0.93 (-0.07)</td><td>0.70 (-0.23)</td><td>0.72 (0.02)</td><td>0.10 (-0.9)</td></tr><tr><td>mistral-7b</td><td>0.50 (-0.28)</td><td>0.57 (0.07)</td><td>0.52 (-0.05)</td><td>0.04 (-0.25)</td></tr><tr><td>phi-3-small</td><td>0.52 (-0.42)</td><td>0.55 (0.03)</td><td>0.68 (0.13)</td><td>0.04 (-0.7)</td></tr><tr><td>phi-3-medium</td><td>0.60 (-0.4)</td><td>0.72 (0.12)</td><td>0.50 (-0.22)</td><td>0.04 (-0.69)</td></tr><tr><td>gemma-2-9b</td><td>0.80 (-0.2)</td><td>0.60 (-0.2)</td><td>0.62 (0.02)</td><td>0.14 (-0.65)</td></tr><tr><td>llama-3.1-8b</td><td>0.84 (-0.16)</td><td>0.78 (-0.06)</td><td>0.82 (0.04)</td><td>0.05 (-0.43)</td></tr></table>
+
+Table 4. Results for Compute on Sets and Lists. The numbers in parentheses indicate the performance difference compared to the corresponding tasks they are evaluated against.
+
+![](images/bfed666d4676da2348fdcb3ebfd1599b9978f798af67592eee35bdce0b310dee.jpg)  
+Figure 3. Results for the Recall and Edit tasks.
+
+![](images/1f44acdd738fc52e86337ce640164bdec3514556791e0e9d2b23647130d1a669.jpg)
+
+![](images/5360e54ee4c97384134042af7e2d84f791224ebddfece1b514edc3748e181d7e.jpg)  
+Figure 4. Results for the Match and Compare tasks.
+
+![](images/2f82a22800f2c8d4004f8f4f5da42986cbaceb010e3081fc4735f4701748db64.jpg)
+
+duplicates (in Find duplicates), and primarily struggled with the counting aspect, which requires tracking and updating an integer state, a skill that is more similar to stateful processing. This suggests that relying solely on counting-based tests (Song et al., 2024) could overly bias the evaluation and fail to capture broader model capabilities. The results also indicate that models exhibit some ability to recognize relative positions and group associations, but their accuracy remains limited (ranging between 0.6-0.8). A closer examination of model generations reveals an overwhelming tendency for the models to produce false positive errors – models often answer “yes” when the correct answer is “no”, while making very few false negative errors. This means that when the relationship is correct, the models can more
+
+reliably identify it. This may stem from a combination of their inherent inclination to agree and the difficulty in recognizing relative comparisons and associations.
+
+![](images/98588932f21bbd6e51bbab27cc6e2b8482131d2061efedc0d1365f84fad07fc5.jpg)  
+Figure 5. Results for Spot the Differences tasks.
+
+Spot the Differences As shown in Figure 5, performance across all models are poor on Compare Two Lists, suggesting inherent difficulties in cross-referencing information across long contexts, even for larger models. GPT-4o and the LLaMA model significantly outperform the others in the Identify the Odd Group task, highlighting a general weakness in detecting contextual differences by the other models. However, an 8B LLaMA model outperforms both equivalently-sized models and even GPT-4 in this task, suggesting that model size alone was not the determining factor. This indicates that architectural differences, training objectives, or specific inductive biases may contribute to improved performance in comparative memory utilization.
+
+Compute on Sets and Lists The tasks in this category require models to recognize and process group structures within the context, and performance gradually declines as the complexity of the task increases (see Table 4). For instance, in comparing the Group Membership task with the
+
+<table><tr><td>Model</td><td>Quantity state</td><td>Set state</td></tr><tr><td>gpt-4-turbo</td><td>0.8</td><td>0.80</td></tr><tr><td>gpt-4o</td><td>1.0</td><td>0.65</td></tr><tr><td>gpt-4o-mini</td><td>0.7</td><td>0.24</td></tr><tr><td>cohere</td><td>0.0</td><td>0.58</td></tr><tr><td>mistral-7b</td><td>0.0</td><td>0.08</td></tr><tr><td>phi-3-small</td><td>0.0</td><td>0.13</td></tr><tr><td>phi-3-medium</td><td>0.0</td><td>0.11</td></tr><tr><td>gemma-2-9b</td><td>0.0</td><td>0.24</td></tr><tr><td>llama-3.1-8b</td><td>0.0</td><td>0.13</td></tr></table>
+
+Table 5. Results for Stateful Processing tasks.  
+
+<table><tr><td>Model</td><td>Processing Data Blocks</td><td>Theory of Mind</td></tr><tr><td>gpt-4-turbo</td><td>0.31</td><td>0.26</td></tr><tr><td>gpt-4o</td><td>0.37</td><td>0.38</td></tr><tr><td>gpt-4o-mini</td><td>0.31</td><td>0.21</td></tr><tr><td>cohere</td><td>0.26</td><td>0.18</td></tr><tr><td>mistral-7b</td><td>0.18</td><td>0.16</td></tr><tr><td>phi-3-small</td><td>0.21</td><td>0.20</td></tr><tr><td>phi-3-medium</td><td>0.18</td><td>0.10</td></tr><tr><td>gemma-2-9b</td><td>0.26</td><td>0.12</td></tr><tr><td>llama-3.1-8b</td><td>0.15</td><td>0.03</td></tr></table>
+
+Table 6. Results for the composite tests.
+
+![](images/a8a16f12e3c0e3ad377b3b3aced108022cd7db25de86c0ad4f106f145e75585a.jpg)  
+Figure 6. Effect of context length (number of operation steps) on performance in the quantity state (left) and set state (right) tasks.
+
+String Search task, where the former requires identifying which list a word belongs to rather than simply determining its presence, the performance of open-source models drops considerably. Similarly, in comparing the Group Association task with the Group Membership task, where the former requires determining whether two words belong to the same group, all models exhibit a noticeable decline in performance. The decline becomes even more pronounced when comparing the Group Association (alternating) variant of the task to the standard Group Association task. Here, the context involves alternating repeated groups rather than simple group structures, which further challenges the models' abilities to handle partitioned contexts effectively.
+
+An interesting observation was found during the Iterate task. In an ablation study, we modified the task to require returning the first words in each list instead of the last words (making it more similar to the Batch Search task). The performance sharply declines when models are asked to return the last words, despite their strong information-fetching capabilities. This suggests that, while the models can retrieve information effectively, they struggle to accurately recognize and process partitions within the context.
+
+Stateful Processing Table 5 presents the results for the Stateful Processing tasks, where performance gaps among models are the most pronounced. The GPT-4(o) models perform well on integer state tracking, while most other
+
+models struggle (near zero accuracy). For set state tracking, larger models generally perform better.
+
+We conducted an ablation study to examine how the number of operation steps influences performance of five selected models (Fig. 6). For quantity state tracking, GPT-4(o) models perform well within fewer than 200 steps but experience a sharp decline in accuracy beyond this threshold. For set state tracking, the performance decline is more gradual. The differences in performance drop between the two tasks can be attributed to the nature of the two tasks. While tracking an integer state might seem simpler than tracking a set, it actually requires the model to maintain and apply every operation sequentially to compute the final value. In contrast, for set state, the fixed size of the set makes more recent operations more relevant to the final state, reducing the need for exhaustive step-by-step tracking. Nevertheless, even in this scenario, all models show a clear inability to handle longer or more complex operation sequences effectively. Interestingly, GPT-4 model outperformed GPT-4o at this task, suggesting potential optimization trade-offs may have affected its ability to manage set-based updates.
+
+Overall, while larger models like GPT-4(o) exhibit some ability to track state over time, their effectiveness rapidly deteriorates as task complexity increases. Smaller models, in particular, struggle to track operations over time, pointing to significant gaps in their ability to manage and process sequential dependencies critical for state tracking tasks.
+
+# 3.4. Results on Composite Tests
+
+The composite tests significantly challenge the models by combining multiple atomic capabilities into a single test. In the Processing Data Blocks task, the context is fixed at 4k tokens, while for the Theory of Mind task, the number of operation steps is set to 100. As shown in Table 6, model performance on both tasks are generally low, showing a broad inability to handle the more complex scenarios. Performance across all models drop substantially on composite tasks compared to their performance on individual capability tasks, such as search, recall, and group processing.
+
+Interestingly, some smaller models, like Mistral and Phi-3-small, exhibit slightly better performance on the Theory of Mind task than on the set state tracking task. This anomaly likely stems from their already weak state tracking ability, which limits their performance across both tasks. Additionally, these models tend to generate longer answers in the set state task which reduces the set overlap.
+
+Notably, even the most capable models, such as GPT-4-turbo and GPT-4o, struggle, showing that scaling model size alone is not enough for solving these composite tasks. Additionally, the variation in performance among smaller models suggests that their limitations stem not only from size but also from underlying architectural or training differences. This indicates that smaller models require more targeted care to bridge the gap in effective memory use.
+
+# 3.5. Extending the benchmark to other configurations
+
+Our benchmark is fully programmable and supports flexible experimentation across a wide range of configurations, including varying context lengths, evaluation criteria, and prompt phrasing. In this section, we illustrate how the benchmark can be adapted beyond the default setup used in the main paper. These examples highlight its versatility in probing model behavior under diverse conditions.
+
+Context length In the main experiments, we fixed the context length to 4K tokens to emphasize that models already exhibit notable failures at this moderate length across many memory tasks. However, the benchmark is scalable to longer contexts. Table 7 presents additional results for two representative tasks Functional updates and Counting evaluated at various context lengths up to 16K tokens.
+
+In both cases, we observe that model performance begins to degrade significantly well before reaching what is typically considered a "long" context window. These failures reveal underlying limitations in how models manage long-range memory beyond simple retrieval. In contrast, models tend to maintain strong performance on retrieval-style tasks (e.g., String search) even at extended lengths, making such tasks less effective at distinguishing model capabilities. Additional results and task examples are provided in Appendix D.
+
+Prompt Variation We also investigated the model performance sensitivity to minor variations in prompt phrasing. Specifically, we tested different phrasings of task instructions while keeping the underlying task logic unchanged. For instance, in the String search task, we compared "Given the context, determine if XXX is present (Var 1)" versus "Is XXX present in the context? (Var 2)". Similarly, in the Group association task, we tested "Determine if word 'AAA' and word 'BBB' are in the same list (Var 1)" versus "Check
+
+if the words 'AAA' and 'BBB' belong to the same list (Var 2).
+
+As shown in Table 8, performance differences between prompt variants are generally small, suggesting that instruction interpretation is not the primary bottleneck in these tasks, rather, the main challenge lies in actually executing the task correctly. Nevertheless, we recognize that more intensive prompt engineering could potentially affect model performance. Given these findings, we standardized the prompts to a single, simple version (as shown in Appendix A) for all experiments in this paper to ensure consistency and comparability across models. However, because the benchmark is programmable, researchers can easily swap in alternate prompts to explore additional prompt settings. Appendix D provides additional examples on more tests for prompt variations.
+
+# 4. Related Work
+
+LLM Evaluation with Benchmarks The evaluation of LLMs has traditionally relied on static benchmarks, from early benchmarks for perplexity-based evaluation (Marcus et al., 1993) to datasets focused on specific downstream tasks such as question answering (Kwiatkowski et al., 2019), summarization (Gliwa et al., 2019), math reasoning (Cobbe et al., 2021), and code generation (Chen et al., 2021). As LLMs began to address a broader range of tasks across various domains (Wu et al., 2023; Wang et al., 2023), more comprehensive benchmark suites (Hendrycks et al., 2021a; Zhong et al., 2023) were developed to assess general capabilities rather than individual task performance. Recent advancements in LLM evaluation have introduced the concept of LLM-as-a-judge, enabling the use of open-ended benchmarks without predefined answers (Zheng et al., 2023). However, these benchmarks remain static in nature and can easily get overfit. Recently, platforms like ChatBot Arena (Zheng et al., 2023) utilize crowdsourcing to rank LLM responses and provides more dynamic evaluations. However, its reliance on human annotation makes it less scalable. Moreover, despite their utility, existing benchmarks primarily assess downstream applications that usually require multiple capabilities, making it difficult to debug and understand model weaknesses.
+
+Tests and Benchmarks for Evaluating Context Utilization As LLMs become capable of processing increasingly long inputs, designing automated tests to evaluate their ability to utilize context has become an area of active research. A notable example is the needle-in-a-haystack (NIAH) task<sup>1</sup>, where a small piece of information (the "needle") is hidden within a long document, and the model needs to re-
+
+<table><tr><td>Model</td><td>500</td><td>1K</td><td>2K</td><td>4K</td><td>8K</td></tr><tr><td>gpt-4o</td><td>1.00</td><td>1.00</td><td>0.99</td><td>0.93</td><td>0.59</td></tr><tr><td>gpt-4o-mini</td><td>0.69</td><td>0.66</td><td>0.42</td><td>0.24</td><td>0.10</td></tr><tr><td>phi-3-small</td><td>0.49</td><td>0.45</td><td>0.21</td><td>0.07</td><td>0.03</td></tr></table>
+
+(a) Functional updates (ROUGE-L)  
+
+<table><tr><td>Model</td><td>1K</td><td>2K</td><td>4K</td><td>8K</td><td>16K</td></tr><tr><td>gpt-4-turbo</td><td>0.52</td><td>0.44</td><td>0.40</td><td>0.32</td><td>0.28</td></tr><tr><td>cohere</td><td>0.36</td><td>0.24</td><td>0.20</td><td>0.20</td><td>0.12</td></tr><tr><td>phi-3-medium</td><td>0.20</td><td>0.16</td><td>0.12</td><td>0.08</td><td>0.04</td></tr></table>
+
+Table 7. Performance across context lengths on two representative tasks.
+
+(b) Counting (Exact match)  
+
+<table><tr><td>Model</td><td>Var 1</td><td>CI95%</td><td>Var 2</td><td>CI95%</td></tr><tr><td>gpt-4o</td><td>1.00</td><td>(0.93, 1.00)</td><td>1.00</td><td>(0.93, 1.00)</td></tr><tr><td>gpt-4o-mini</td><td>0.98</td><td>(0.90, 1.00)</td><td>0.98</td><td>(0.90, 1.00)</td></tr><tr><td>phi-3-medium</td><td>1.00</td><td>(0.93, 1.00)</td><td>1.00</td><td>(0.93, 1.00)</td></tr></table>
+
+(a) String search (Exact match)  
+
+<table><tr><td>Model</td><td>Var 1</td><td>CI95%</td><td>Var 2</td><td>CI95%</td></tr><tr><td>gpt-4o</td><td>0.65</td><td>(0.50, 0.78)</td><td>0.63</td><td>(0.47, 0.76)</td></tr><tr><td>cohere</td><td>0.70</td><td>(0.55, 0.82)</td><td>0.75</td><td>(0.60, 0.86)</td></tr><tr><td>phi-3-small</td><td>0.55</td><td>(0.40, 0.69)</td><td>0.55</td><td>(0.40, 0.69)</td></tr></table>
+
+(b) Group association (Exact match)
+
+Table 8. Prompt variation performance on the String search task (with 50 samples) and the Group association task (with 40 samples). Variation 1 and Variation 2 differ slightly in phrasing but preserve task intent.
+
+trieve it. Similar tests include key-value retrieval (Liu et al., 2024) and passkey retrieval (Mohtashami & Jaggi, 2023). The simplicity and interpretability of NIAH have made it a standard for evaluating LLM context utilization, and it has since inspired various methods for improving long-context processing (Mohtashami & Jaggi, 2023; Ding et al., 2024; Xiong et al., 2023; Behrouz et al., 2024).
+
+However, these tests focus solely on basic information retrieval, without capturing more complex aspects of context processing. To address this limitation, other tests have been proposed. Needlebench (Li et al., 2024) extends simple retrieval tasks to include multi-needle reasoning and a ancestral trace task which requires navigating chains or graphs of information. Song et al. (2024) introduce the Counting Stars task, which involves tallying numbers of stars embedded in phrases. Ruler (Hsieh et al., 2024) proposes additional tasks such as variable tracking and frequent word extraction. While these tests increase task complexity or broaden the range of evaluated tasks, they remain limited in scope for systematically evaluating contextual processing.
+
+Beyond individual tests, several benchmarks explicitly target long-context processing, including InftyBench (Zhang et al., 2024), L-Eval (An et al., 2024), and LongBench (Bai et al., 2024). These benchmarks use NIAH-like tasks alongside question answering, summarization, and code generation over long contexts. However, like other benchmarks, they remain static and primarily measure end-to-end performance rather than systematically dissecting capabilities.
+
+Analogies to Human Cognitive Testing Memory tests are widely used in cognitive research to assess specific functions. Such assessments often involve evaluating short term memory via recall tests (Crannell & Parrish, 1957; Towse et al., 2008), inductive reasoning via pattern recognition tasks, or attention via instruction-following (Kane et al.,
+
+2007; Nasreddine et al., 2005). By isolating distinct abilities while minimizing confounding factors like attention, memory span, and reasoning (Kane et al., 2007), such tests provide detailed profiles of cognitive functions, guiding interventions and shaping broader theories of human thought. Inspired by this approach, we design atomic tests that systematically isolate core aspects of LLM context processing, aiming for a fine-grained understanding of LLM memory-usage capabilities – analogous to memory testing in humans.
+
+# 5. Conclusions
+
+AI assistants powered by LLMs are expected to handle numerous operations involving memory. However, simple tests reveal that they often fall short of meeting user expectations, even in basic retrieval and processing tasks. For instance, retrieving vacation schedules for each team member from a message history that includes evolving plans over time proves challenging. To enable targeted improvements, it is essential to establish a comprehensive benchmark that tests each capability in isolation while also allowing for programmable composition to evaluate more complex scenarios. Our benchmark provides a straightforward yet effective approach to achieving this goal. We primarily focus on short-context scenarios to demonstrate that current limitations are not solely attributable to the models' challenges with parsing long contexts. Addressing these issues demands attention beyond merely solving the "attention" problem.
+
+# Impact Statement
+
+This paper contributes to the advancement of Machine Learning by introducing a systematic and programmable evaluation framework for assessing the contextual processing capabilities of large language models. Our work provides insights into model strengths and limitations in han
+
+dling various atomic and composite tasks, offering a structured way to analyze model behavior. These contributions can guide future research in improving model efficiency and reliability.
+
+We do not foresee any direct ethical concerns or negative societal consequences arising from this work. Our evaluation methodology is designed to be model-agnostic and does not involve sensitive data or high-stakes applications.
+
+# References
+
+An, C., Gong, S., Zhong, M., Zhao, X., Li, M., Zhang, J., Kong, L., and Qiu, X. L-eval: Instituting standardized evaluation for long context language models. In Ku, L.-W., Martins, A., and Srikumar, V. (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 14388-14411, Bangkok, Thailand, August 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.acl-long.776. URL https://aclanthology.org/2024.acl-long.776/.  
+Anthropic. Introducing the next generation of claude. https://www.anthropic.com/news/claude-3-family, 2024. Accessed: 2024-03-27.  
+Bai, Y., Lv, X., Zhang, J., Lyu, H., Tang, J., Huang, Z., Du, Z., Liu, X., Zeng, A., Hou, L., Dong, Y., Tang, J., and Li, J. LongBench: A bilingual, multitask benchmark for long context understanding. In Ku, L.-W., Martins, A., and Srikumar, V. (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 3119-3137, Bangkok, Thailand, August 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.acl-long.172. URL https://aclanthology.org/2024.acl-long.172/.  
+Behrouz, A., Zhong, P., and Mirrokni, V. Titans: Learning to memorize at test time, 2024. URL https://arxiv.org/abs/2501.00663.  
+Chen, M., Tworek, J., Jun, H., Yuan, Q., de Oliveira Pinto, H. P., Kaplan, J., Edwards, H., Burda, Y., Joseph, N., Brockman, G., Ray, A., Puri, R., Krueger, G., Petrov, M., Khlaaf, H., Sastry, G., Mishkin, P., Chan, B., Gray, S., Ryder, N., Pavlov, M., Power, A., Kaiser, L., Bavarian, M., Winter, C., Tillet, P., Such, F. P., Cummings, D., Plappert, M., Chantzis, F., Barnes, E., Herbert-Voss, A., Guss, W. H., Nichol, A., Paino, A., Tezak, N., Tang, J., Babuschkin, I., Balaji, S., Jain, S., Saunders, W., Hesse, C., Carr, A. N., Leike, J., Achiam, J., Misra, V., Morikawa, E., Radford, A., Knight, M., Brundage, M., Murati, M., Mayer, K., Welinder, P., McGrew, B., Amodei, D., McCandlish, S., Sutskever, I., and Zaremba
+
+W. Evaluating large language models trained on code. 2021.  
+Clark, P., Cowhey, I., Etzioni, O., Khot, T., Sabharwal, A., Schoenick, C., and Tafjord, O. Think you have solved question answering? try arc, the ai2 reasoning challenge, 2018. URL https://arxiv.org/abs/1803.05457.  
+Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., and Schulman, J. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021.  
+Crannell, C. and Parrish, J. A comparison of immediate memory span for digits, letters, and words. The Journal of Psychology, 44(2):319-327, 1957.  
+Ding, Y., Zhang, L. L., Zhang, C., Xu, Y., Shang, N., Xu, J., Yang, F., and Yang, M. Longrope: Extending llm context window beyond 2 million tokens, 2024.  
+Gliwa, B., Mochol, I., Biesek, M., and Wawer, A. Samsum corpus: A human-annotated dialogue dataset for abstractive summarization. In Proceedings of the 2nd Workshop on New Frontiers in Summarization. Association for Computational Linguistics, 2019. doi: 10.18653/v1/d19-5409. URL http://dx.doi.org/10.18653/v1/D19-5409.  
+Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., and Steinhardt, J. Measuring massive multitask language understanding. In International Conference on Learning Representations, 2021a. URL https://openreview.net/forum?id=d7KBjmI3GmQ.  
+Hendrycks, D., Burns, C., Kadavath, S., Arora, A., Basart, S., Tang, E., Song, D., and Steinhardt, J. Measuring mathematical problem solving with the math dataset. NeurIPS, 2021b.  
+Hsieh, C.-P., Sun, S., Kriman, S., Acharya, S., Rekesh, D., Jia, F., and Ginsburg, B. RULER: What's the real context size of your long-context language models? In First Conference on Language Modeling, 2024. URL https://openreview.net/forum?id=kIoBbc76Sy.  
+Jaccard, P. Étude comparative de la distribution florale dans une portion des alpes et des jura. Bull Soc Vaudoise Sci Nat, 37:547-579, 1901.  
+Kamradt, G. Needle in a haystack - pressure testing llms. GitHub, 2023. URL https://github.com/gkamradt/LLMTest_NeedleInAHaystack/tree/main.
+
+Kane, R. L., Roebuck-Spencer, T., Short, P., Kabat, M., and Wilken, J. Identifying and monitoring cognitive deficits in clinical populations using automated neuropsychological assessment metrics (anam) tests. Archives of Clinical Neuropsychology, 22(Suppl_1):S115-S126, 2007.  
+Kwiatkowski, T., Palomaki, J., Redfield, O., Collins, M., Parikh, A., Alberti, C., Epstein, D., Polosukhin, I., Devlin, J., Lee, K., et al. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:453-466, 2019.  
+Li, M., Zhang, S., Liu, Y., and Chen, K. Needlebench: Can llms do retrieval and reasoning in 1 million context window?, 2024. URL https://arxiv.org/abs/2407.11963.  
+Lin, C.-Y. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pp. 74-81, Barcelona, Spain, July 2004. Association for Computational Linguistics. URL https://aclanthology.org/W04-1013/.  
+Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., and Liang, P. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12:157-173, 2024. doi: 10.1162/tacl_a_00638. URL https://aclanthology.org/2024.tacl-1.9/.  
+Marcus, M. P., Santorini, B., and Marcinkiewicz, M. A. Building a large annotated corpus of English: The Penn Treebank. Computational Linguistics, 19(2):313-330, 1993. URL https://aclanthology.org/J93-2004/.  
+Mohtashami, A. and Jaggi, M. Random-access infinite context length for transformers. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=7eHn64wOVy.  
+Nasreddine, Z. S., Phillips, N. A., Bédirian, V., Charbonneau, S., Whitehead, V., Collin, I., Cummings, J. L., and Chertkow, H. The montreal cognitive assessment, moca: a brief screening tool for mild cognitive impairment. Journal of the American Geriatrics Society, 53(4):695-699, 2005.  
+Song, M., Zheng, M., and Luo, X. Counting-stars: A multi-evidence, position-aware, and scalable benchmark for evaluating long-context large language models, 2024. URL https://arxiv.org/abs/2403.11802.  
+Suzgun, M., Scales, N., Schärli, N., Gehrmann, S., Tay, Y., Chung, H. W., Chowdhery, A., Le, Q. V., Chi, E. H., Zhou, D., and Wei, J. Challenging big-bench tasks and
+
+whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261, 2022.  
+Towse, J. N., Cowan, N., Hitch, G. J., and Horton, N. J. The recall of information from working memory: Insights from behavioural and chronometric perspectives. Experimental Psychology, 55(6):371-383, 2008.  
+Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., and Anandkumar, A. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv: Arxiv-2305.16291, 2023.  
+Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., Jiang, L., Zhang, X., Zhang, S., Liu, J., Awadallah, A. H., White, R. W., Burger, D., and Wang, C. Autogen: Enabling next-gen llm applications via multi-agent conversation, 2023. URL https://arxiv.org/abs/2308.08155.  
+Wu, Y., Hee, M. S., Hu, Z., and Lee, R. K.-W. Longgen-bench: Benchmarking long-form generation in long context llms. arXiv preprint arXiv:2409.02076, 2024.  
+Xiong, W., Liu, J., Molybog, I., Zhang, H., Bhargava, P., Hou, R., Martin, L., Rungta, R., Sankararaman, K. A., Oguz, B., Khabsa, M., Fang, H., Mehdad, Y., Narang, S., Malik, K., Fan, A., Bhosale, S., Edunov, S., Lewis, M., Wang, S., and Ma, H. Effective long-context scaling of foundation models, 2023. URL https://arxiv.org/abs/2309.16039.  
+Zhang, X., Chen, Y., Hu, S., Xu, Z., Chen, J., Hao, M., Han, X., Thai, Z., Wang, S., Liu, Z., and Sun, M. Infty-bench: Extending long context evaluation beyond 100K tokens. In Ku, L.-W., Martins, A., and Srikumar, V. (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 15262-15277, Bangkok, Thailand, August 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.acl-long.814. URL https://aclanthology.org/2024.acl-long.814/.  
+Zheng, L., Chiang, W.-L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., Zhang, H., Gonzalez, J. E., and Stoica, I. Judging LLM-as-a-judge with MT-bench and chatbot arena. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. URL https://openreview.net/forum?id=uccHPGDlao.  
+Zhong, W., Cui, R., Guo, Y., Liang, Y., Lu, S., Wang, Y., Saied, A., Chen, W., and Duan, N. Agieval: A human-centric benchmark for evaluating foundation models, 2023. URL https://arxiv.org/abs/2304.06364.
+
+# A. Test Templates
+
+In this appendix, we provide the templates of the test prompts. Placeholder context words such as “aaa,bbb,ccc,” etc., are used for illustration purposes. During testing, these context words are uniformly sampled from an English dictionary. Variable tokens in the instruction part are marked with the bold font.
+
+<table><tr><td colspan="2">Search</td></tr><tr><td>Task name</td><td>Prompt</td></tr><tr><td>String search (with word)</td><td>Context: 
+aaa,bbb,ccc,... 
+Instruction: 
+Given the context, determine if the word “bbb” is present in the context. Answer with “yes” or ‘no’. 
+Answer:</td></tr><tr><td>String search (with subsequence)</td><td>Context: 
+aaa,bbb,ccc,... 
+Instruction: 
+Given the list of words in the context, determine if the sequence “bbb, xxx, ddd” appears in the context. Answer with ‘yes’ or ‘no’. 
+Answer:</td></tr><tr><td>Key-value search</td><td>Context: 
+aaa:bbb, ccc:ddd, ... 
+Instruction: 
+Given a list of word pairs formatted as “word.1: word.2” in the context, return the second word associated with the provided first word. For the first word “aaa”, the corresponding second word is:</td></tr><tr><td>Batch search</td><td>Context: 
+aaa:bbb, ccc:ddd, ... 
+Instruction: 
+Given a list of word pairs formatted as “word.1: word.2” in the context, return the second word associated with the provided first words. For the first words: aaa, ccc, ..., the corresponding second words are:</td></tr><tr><td colspan="2">Recall and Edit</td></tr><tr><td>Task name</td><td>Prompt</td></tr><tr><td>Snapshot</td><td>Context: 
+aaa,bbb,ccc,... 
+Instruction: 
+Repeat the previous context exactly as it is, without making any additions or deletions. 
+Answer:</td></tr><tr><td>Replace all (x to y)</td><td>Context: 
+aaa,bbb,aaa,ccc,aaa,ddd,... 
+Instruction: 
+Repeat the previous context and replace the word “aaa” with “zzz” each time it appears. 
+Answer:</td></tr><tr><td>Replace all (x to null)</td><td>Context: 
+aaa,bbb,aaa,ccc,aaa,ddd,... 
+Instruction: 
+Repeat the previous context but skip the word “aaa” each time it appears. 
+Answer:</td></tr><tr><td>Overwrite positions (nth to y)</td><td>Context: 
+aaa,bbb,ccc,... 
+Instruction: 
+Repeat the previous context and replace every third word with “zzz”. 
+Answer:</td></tr><tr><td>Overwrite positions (nth to null)</td><td>Context: 
+aaa,bbb,ccc,... 
+Instruction: 
+Repeat the previous context and skip every other word. 
+Answer:</td></tr><tr><td>Functional updates</td><td>Context: 
+111,222,333,... 
+Instruction: 
+Add 3 to every number in the previous context. 
+Answer:</td></tr><tr><td colspan="2">Match and Compare</td></tr><tr><td>Task name</td><td>Prompt</td></tr><tr><td>Compare positions</td><td>Context: 
+aaa,bbb,ccc,... 
+Instruction: 
+Given the list of words in the context, determine the relative positions of two words. 
+Does the word “aaa” appear before the word “ccc” in the list? Answer “yes” or “no”. 
+Answer:</td></tr><tr><td>Find duplicates</td><td>Context: 
+aaa,bbb,aaa,... 
+Instruction: 
+A word is repeated multiple times in the context. Your task is to identify the word that is repeated. 
+The repeated word is:</td></tr><tr><td>Count</td><td>Context: 
+aaa,bbb,aaa,... 
+Instruction: 
+Count the number of times the word “aaa” appeared in the context. 
+Answer: The word “aaa” appeared</td></tr><tr><td>Check association</td><td>Context: 
+aaa:attribute 1,bbb:attribute 2,ccc: attribute 2,ddd: attribute 1,... 
+Instruction: 
+Given the list of words and their respective attributes in the format of “word:attribute”, determine if the word “aaa” and the word “ggg” have the same attribute. Answer with “yes” or “no”. 
+Answer:</td></tr><tr><td colspan="2">Spot the Differences</td></tr><tr><td>Task name</td><td>Prompt</td></tr><tr><td>Compare two lists</td><td>Context: 
+List 1: aaa,bbb,ccc,... 
+List 2: aaa,ddd,ccc,... 
+Instruction: 
+There are two lists of words in the context. The first list contains the original words. The second list is similar to the first but has some words replaced with different ones. Your task is to identify the words in the first/second list that are different from those in the other list. Provide the different words as your answer. 
+Answer:</td></tr><tr><td>Identify the odd group</td><td>Context: 
+List 1: aaa,bbb,ccc,... 
+List 2:bbb,aaa,ccc,... 
+List 3:aaa,zzz,ccc,... 
+List 4:ccc,aaa,bbb,... 
+Instruction: 
+Given the lists of words in the context, identify the list that is different from the others. Provide the list number as your answer. For example, if the Nth list is different, provide “List N” as your answer. 
+Answer:</td></tr><tr><td>Patch the difference</td><td>Context: 
+aaa,bbb,ccc,aaa,bbb,ccc,... 
+Instruction: 
+Given the sequence of words that follows a specific pattern in the context, predict the Nth word that appears after the final word in the given sequence. 
+Answer: The Nth word that appears after the final word in the given sequence is</td></tr><tr><td colspan="2">Compute on Sets and Lists</td></tr><tr><td>Task name</td><td>Prompt</td></tr><tr><td>Group membership</td><td>Context: 
+List 1: aaa,bbb,ccc,... 
+List 2:ddd,eee,fff,... 
+... 
+Instruction: 
+Given the lists of words in the context, determine which list contains the word “fff”. If the word is not present in either list, answer “no”. 
+Answer:</td></tr><tr><td>Group association</td><td>Context: 
+List 1: aaa,bbb,ccc,... 
+List 2:ddd,eee,fff,... 
+... 
+Instruction: 
+Given the lists of words in the context, determine if the word “aaa” and the word “eee” are in the same list. Answer with “yes” or “no”. 
+Answer:</td></tr><tr><td>Group association (al-ternating)</td><td>Context:Role A: aaa,bbb,...Role B:ccc,ddd,...Role A: eee, fff,...Role B: ggg,hhh,...Instruction:Given the context with alternating roles and their respective context words,determine if the word “aaa” and the word “ggg” are in the same role. Answer with “yes” or “no”. Answer:</td></tr><tr><td>Iterate</td><td>Context:List 1: aaa,bbb,ccc,...List 2: ddd, eee, fff,...Instruction:Given the lists of words in the context, identify and recall the last word from each list.Provide your answer as a list of these words separated by commas.Answer:</td></tr></table>
+
+<table><tr><td colspan="2">Stateful Processing</td></tr><tr><td>Task name</td><td>Prompt</td></tr><tr><td>Quantity state</td><td>Context:
+Begin with the number xx. Perform the following operations:
+1. Add xx 2. Subtract xx 3. ...
+Instruction:
+In the context, you are given an initial number and a series of operations to perform on that number. Your task is to determine the final result of the operations. Write your final answer after the text “FINAL ANSWER:”. For example, “FINAL ANSWER: 42”.
+FINAL ANSWER:</td></tr><tr><td>Set state</td><td>Agent actions:
+Agent draws aaa,bbb,ccc
+Agent discardsbbb,ccc
+Agent draws ddd, fff
+Agent discards ddd
+...
+Instruction:
+Given the actions of the agent, your task is to determine the final list of words the agent ends up with after a series of actions. Write your final answer after the text “FINAL ANSWER:”. For example, “FINAL ANSWER: word1, word2, word3”.
+FINAL ANSWER:</td></tr><tr><td colspan="2">Processing Data Blocks</td></tr><tr><td>Task name</td><td>Prompt</td></tr></table>
+
+<table><tr><td>Processing Blocks</td><td>Data</td><td>Context: 
+Role 1: aaa,bbb,ccc,... 
+Role 2: ddd, eee, fff,... 
+Role 3: ggg,hhh,iii,... 
+Role 1: jjj,kkk,... 
+... 
+Instruction: 
+The context consists of a series of alternating roles, each associated with a list of words. Your task is to identify and recall all the words from the role labeled “Role 2” that appear after the word “zzz” in the sequence. Please write your answer after the text “Answer”: For example,“Answer:word1,word2,word3”. Answer:</td></tr><tr><td colspan="3">Composite-State Tracking (Theory of Mind)</td></tr><tr><td>Task name</td><td colspan="2">Prompt</td></tr><tr><td>Theory of Mind</td><td colspan="2">Agents actions:: 
+Agent A starts with the following words: aaa,bbb,ccc,... 
+Agent B starts with the following words: ddd, eee, fff,... 
+Agent B starts with the following words: ggg,hhh,iii,... 
+Agent B swaps the following words “ddd” with Agent C for the following words “hhh”. 
+Agent A discards the following words: ccc,zzz,... 
+Agent C draws the following words: xxx,yyy,... ... 
+Instruction: 
+Given the actions of the agents, your task is to determine the final list of words each agent ends up with after a series of actions. Write your final answer after the text “FINAL ANSWER:”. For example,“FINAL ANSWER:Agent A:word1,word2,word3\nAgent B:word4,word5”. 
+FINAL ANSWER:</td></tr></table>
+
+# B. Task Details
+
+This appendix provides details for each task, including the number of examples, evaluation metrics, and configurable hyperparameters. The context length is fixed at 4k for almost all tasks, apart from Stateful Processing, where the context is determined by number of operation steps and set to 200 for quantity state and 100 for set state, which maps to around 1.5k context tokens.
+
+Here is an example of number of examples calculation String search (with word): 5 (query depth) * 2 (labels) * 5 (samples per parameter setting) = 50.
+
+Table 10: Task Overview with Hyperparameters, Number of Examples, and Evaluation Metrics  
+
+<table><tr><td>Task Name</td><td>Hyperparameters</td><td># of Examples</td><td>Metric</td></tr><tr><td colspan="4">Search</td></tr><tr><td>String search (word)</td><td>query depth = [0, 0.25, 0.5, 0.75, 1], label = [positive, negative], samples = 5</td><td>50</td><td>exact-match</td></tr><tr><td>String search (sequence)</td><td>sequence length = [8, 16, 32, 64], label = [positive, negative], samples = 10</td><td>80</td><td>exact-match</td></tr><tr><td>Key-value search</td><td>query depth = [0, 0.25, 0.5, 0.75, 1], samples = 10</td><td>50</td><td>exact-match</td></tr><tr><td></td><td></td><td colspan="2">Continued on next page</td></tr><tr><td>Batch search</td><td>batch size = [4, 8, 16, 32], samples = 5</td><td>20</td><td>rouge-L_recall</td></tr><tr><td></td><td></td><td colspan="2">Number of Entries for Category: 200</td></tr><tr><td>Recall and Edit</td><td></td><td></td><td></td></tr><tr><td>Snapshot (words)</td><td>samples = 10</td><td>10</td><td>rouge-L</td></tr><tr><td>Replace all</td><td>density = [0.2, 0.4, 0.6, 0.8], y = [random word, null], samples = 5</td><td>40</td><td>rouge-L</td></tr><tr><td>Overwrite positions</td><td>nth = [2, 3, 4], y = [random word, null], samples = 5</td><td>30</td><td>rouge-L</td></tr><tr><td>Snapshot (numbers)</td><td>samples = 10</td><td>10</td><td>rouge-L</td></tr><tr><td>Functional updates</td><td>function type = [add (3), subtract (1), multiply (2)], samples = 5</td><td>15</td><td>rouge-L</td></tr><tr><td></td><td></td><td colspan="2">Number of Entries for Category: 105</td></tr><tr><td>Match and Compare</td><td></td><td></td><td></td></tr><tr><td>Compare positions</td><td>query 1 depth = [0, 0.25, 0.5, 0.75, 1], query 2 depth = [0, 0.25, 0.5, 0.75, 1], samples = 3</td><td>75</td><td>exact-match</td></tr><tr><td>Find duplicates</td><td>repetition = [2, 4, 8, 16, 32], samples = 5</td><td>25</td><td>exact-match</td></tr><tr><td>Count</td><td>repetition = [2, 4, 8, 16, 32], samples = 5</td><td>25</td><td>exact-match</td></tr><tr><td>Check association</td><td>n attribute = [2, 4, 8, 16, 32], label = [positive, negative], samples = 5</td><td>50</td><td>exact-match</td></tr><tr><td></td><td></td><td colspan="2">Number of Entries for Category: 175</td></tr><tr><td>Spot the Differences</td><td></td><td></td><td></td></tr><tr><td>Compare two lists</td><td>num different words = [1, 5, 10, 20], chosen list = [first, second], samples = 10</td><td>80</td><td>rouge-L_recall</td></tr><tr><td>Identify the odd group</td><td>words per group = [25, 50, 75, 100], percentage of difference = [0, 0.25, 0.5], samples = 5</td><td>60</td><td>exact-match</td></tr><tr><td>Patch the difference</td><td>pattern length = [2, 15, 30], cut off depth = [0, 0.5, 1], nth = [1, 3, 6], samples = 5</td><td>120²</td><td>exact-match</td></tr><tr><td></td><td></td><td colspan="2">Number of Entries for Category: 260</td></tr><tr><td>Compute on Sets and Lists</td><td></td><td></td><td></td></tr><tr><td>Group membership</td><td>number of groups = [4, 8, 16, 32], query depth = [0, 0.25, 0.5, 0.75, 1], samples = 5</td><td>100</td><td>exact-match</td></tr><tr><td>Group association</td><td>number of groups = [4, 8, 16, 32], label = [positive, negative], samples = 5</td><td>40</td><td>exact-match</td></tr><tr><td>Group association (alternating)</td><td>number of groups = [2, 4, 8, 16, 32], number of turns = 10, label = [positive, negative], sample = 5</td><td>50</td><td>exact-match</td></tr><tr><td>Iterate</td><td>number of groups = [4, 8, 16, 32], samples = 5</td><td>20</td><td>rouge-L</td></tr><tr><td></td><td></td><td colspan="2">Number of Entries for Category: 210</td></tr><tr><td>Stateful Processing</td><td></td><td></td><td></td></tr><tr><td>Set state</td><td>number of steps = 100, set size = [5, 10, 15, 20], samples = 10</td><td>40</td><td>jaccard_similarity</td></tr><tr><td>Quantity state</td><td>number of steps = 200, samples = 10</td><td>10</td><td>exact-match</td></tr><tr><td></td><td></td><td colspan="2">Number of Entries for Category: 50</td></tr><tr><td></td><td></td><td colspan="2">Continued on next page</td></tr><tr><td colspan="4">Composite</td></tr><tr><td>Processing data blocks</td><td>number of blocks = [2, 4, 8, 16, 32], number of 
+turns = 10, samples = 5</td><td>50</td><td>rouge-L</td></tr><tr><td>Theory of mind</td><td>number of steps = 100, number of agents = [2, 
+3, 4], samples = [10, 20]</td><td>60</td><td>jaccard_similarity</td></tr><tr><td></td><td></td><td colspan="2">Number of Entries for Category: 110</td></tr><tr><td></td><td></td><td colspan="2">Total Number of Entries: 1110</td></tr></table>
+
+# C. Evaluation Metrics
+
+In this appendix section, we provide details about the evaluation metrics we have used in the tests.
+
+- Exact Match: The exact match accuracy measures whether the generated answer exactly matches the reference answer. It is computed as follows:
+
+$$
+\text {E x a c t M a t c h} = \left\{ \begin{array}{l l} 1 & \text {i f r e f e r e n c e _ {-} a n s w e r} = \text {g e n e r a t e d _ {-} a n s w e r}, \\ 0 & \text {o t h e r w i s e}. \end{array} \right.
+$$
+
+- ROUGE-L / ROUGE-L-recall: ROUGE (Recall-Oriented Understudy for Gisting Evaluation) (Lin, 2004) measures the verbatim overlap between the reference and the generated answers. ROUGE-L specifically looks for the longest common subsequence (LCS) between the two texts, which reflects the structure of the text and the longest sequence of matching words. ROUGE-L recall focuses on the ability of the model to recall the content from the reference answer, and it emphasizes matching the longest subsequences.
+
+ROUGE-L-recall can be defined as:
+
+$$
+\text {R O U G E - L - r e c a l l} = \frac {\text {L C S (g e n e r a t e d - a n s w e r , r e f e r e n c e - a n s w e r)}}{\text {l e n g t h o f r e f e r e n c e - a n s w e r}}
+$$
+
+ROUGE-L is computed as the F1-score, which combines both precision and recall to provide a more balanced measure of overlap.
+
+- Jaccard Similarity: Jaccard similarity measures the overlap between two sets by comparing the intersection and union of the sets. It is computed as:
+
+$$
+\text {J a c c a r d S i m i l a r i t y} = \frac {| A \cap B |}{| A \cup B |}
+$$
+
+where  $A$  and  $B$  are sets representing the elements in the generated and reference answers, respectively. This metric is used for tasks involving set-based comparisons or when the goal is to measure the similarity between two sets of elements (e.g., word sets).
+
+# D. More Examples for Context Length Variation and Prompt Variation
+
+In this section, we present additional examples illustrating the effects of context length and minor variations in prompt instructions on selected models.
+
+Length variation Table 11 reports results from three additional tasks: string search (word), replace all, and iterate. The results reaffirm that while models tend to perform well on the search task even at long context lengths, their performance on other tasks degrades significantly at much shorter lengths. This highlights a fundamental limitation in how effectively these models utilize long-range context for non-search tasks.
+
+Minerva: A Programmable Memory Test Benchmark for Language Models  
+
+<table><tr><td>Model</td><td>2000</td><td>4000</td><td>8000</td><td>16000</td><td>32000</td></tr><tr><td>gpt-4o</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>0.98</td></tr><tr><td>gpt-4o-mini</td><td>0.98</td><td>0.98</td><td>0.94</td><td>0.90</td><td>0.78</td></tr><tr><td>phi-3-small</td><td>1.00</td><td>0.94</td><td>0.90</td><td>0.98</td><td>0.98</td></tr></table>
+
+(a) String search (word)
+
+<table><tr><td>Model</td><td>1000</td><td>2000</td><td>4000</td><td>8000</td><td>16000</td></tr><tr><td>gpt-4o</td><td>1.00</td><td>1.00</td><td>0.99</td><td>0.81</td><td>0.48</td></tr><tr><td>gpt-4o-mini</td><td>0.99</td><td>0.91</td><td>0.84</td><td>0.71</td><td>0.42</td></tr><tr><td>phi-3-small</td><td>0.87</td><td>0.67</td><td>0.49</td><td>0.32</td><td>0.06</td></tr></table>
+
+(b) Replace all  
+
+<table><tr><td>Model</td><td>1000</td><td>2000</td><td>4000</td><td>8000</td><td>16000</td></tr><tr><td>gpt-4o</td><td>1.00</td><td>0.97</td><td>0.86</td><td>0.70</td><td>0.57</td></tr><tr><td>gpt-4o-mini</td><td>0.91</td><td>0.87</td><td>0.67</td><td>0.43</td><td>0.28</td></tr><tr><td>phi-3-small</td><td>0.15</td><td>0.09</td><td>0.04</td><td>0.01</td><td>0.01</td></tr></table>
+
+Prompt variation. We provide additional results examining how small changes in prompt phrasing affect model performance across three tasks (see Table 12). In general, we find that minor wording changes (e.g., in replace all and quantity state) do not significantly affect performance. This suggests that task accuracy is primarily driven by the model's underlying capability to process memory rather than sensitivity to prompt wording. However, more substantial prompt changes, such as shown in check association task, can lead to notable differences in performance across models. To minimize the impact of prompt tuning, we standardize all prompts to the versions specified in the main paper.
+
+(c) Iterate  
+Table 11. Model performance across varying context lengths for three tasks.  
+
+<table><tr><td>Model</td><td>Var 1</td><td>CI95%</td><td>Var 2</td><td>CI95%</td></tr><tr><td>gpt-4o</td><td>0.99</td><td>(0.93, 1.00)</td><td>0.98</td><td>(0.93, 1.00)</td></tr><tr><td>gpt-4o-mini</td><td>0.84</td><td>(0.71, 0.94)</td><td>0.83</td><td>(0.71, 0.94)</td></tr><tr><td>phi-3-small</td><td>0.49</td><td>(0.32, 0.63)</td><td>0.51</td><td>(0.35, 0.65)</td></tr></table>
+
+(a) Replace all (ROUGE-L)  
+
+<table><tr><td>Model</td><td>Var 1</td><td>CI95%</td><td>Var 2</td><td>CI95%</td></tr><tr><td>gpt-4o</td><td>0.72</td><td>(0.58, 0.83)</td><td>0.76</td><td>(0.63, 0.86)</td></tr><tr><td>gpt-4o-mini</td><td>0.60</td><td>(0.46, 0.72)</td><td>0.58</td><td>(0.42, 0.69)</td></tr><tr><td>phi-3-small</td><td>0.70</td><td>(0.56, 0.81)</td><td>0.60</td><td>(0.46, 0.72)</td></tr></table>
+
+(b) Check association (Exact match)  
+
+<table><tr><td>Model</td><td>Var 1</td><td>CI95%</td><td>Var 2</td><td>CI95%</td></tr><tr><td>gpt-4o</td><td>1.00</td><td>(0.72, 1.00)</td><td>1.00</td><td>(0.72, 1.00)</td></tr><tr><td>gpt-4o-mini</td><td>0.70</td><td>(0.40, 0.89)</td><td>0.80</td><td>(0.49, 0.94)</td></tr><tr><td>phi-3-small</td><td>0.00</td><td>(0.00)</td><td>0.00</td><td>(0.00)</td></tr></table>
+
+(c) Quantity state (Exact match)
+
+Table 12. Effect of prompt variations on model performance across three tasks. We report accuracy (or ROUGE-L) and  $95\%$  confidence intervals.
+
+# Prompt variants used in the table above:
+
+# - Replace all
+
+- Var 1: "Repeat the previous context and replace the word aaa with bbb."  
+- Var 2: "Copy the previous context but replace the word aaa withbbb."
+
+# - Check association
+
+- Var 1: "Given the context with words and their assigned attributes in the format of word: ATT_N, determine if the word aaa has the same attribute as the wordbbb?"  
+- Var 2: "Given the context"
+
+# Quantity state
+
+- Var 1: "Your task is to determine the final result of the operations."  
+- Var 2: "Determine the final result after the operations."

2025/Minerva_ A Programmable Memory Test Benchmark for Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:294bea5337f505bd13619c198b4e7346107f92701008d03da3c3ec4c9c3d31cb
+size 1986356