Research map / Video / Video generation
InterCMDM: Block-Causal Diffusion for Autoregressive Human Interaction Generation
Qing Yu, Kent Fujiwara
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use dual-stream instead of single-stream blocks (causal B=3) (ablation) | InterCMDM · InterHuman ablation | R-Precision Top-1: 0.484 -> 0.529 | Improved |
| use dual-stream instead of single-stream blocks (causal B=3) (ablation) | InterCMDM · InterHuman ablation | FID: 5.068 -> 4.492 | Improved |
| use dual-stream in all layers instead of hybrid (ablation) | InterCMDM · InterHuman ablation | R-Precision Top-1: 0.499 -> 0.529 | Improved |
| use separate per-person weights instead of shared in dual-stream (ablation) | InterCMDM · InterHuman ablation | R-Precision Top-1: 0.506 -> 0.529 | Improved |
| use bidirectional instead of causal attention (ablation) | InterCMDM (dual) · InterHuman ablation | R-Precision Top-1: 0.529 -> 0.549 | Mixed |
| use block size 3 instead of 1 (ablation) | InterCMDM · InterHuman ablation | R-Precision Top-1: 0.514 -> 0.529 | Improved |
| use block size 5 instead of 3 (ablation) | InterCMDM · InterHuman ablation | R-Precision Top-1: 0.529 -> 0.532 | Mixed |
| use latent rollout instead of decode-encode composition (ablation) | InterCMDM · Long-horizon InterHuman (B/NB ratio) | B/NB ratio: 1.91 -> 1.66 | Improved |
| use latent rollout instead of naive concatenation (ablation) | InterCMDM · Long-horizon InterHuman (B/NB ratio) | B/NB ratio: 98.58 -> 1.66 | Improved |
| use latent rollout instead of naive concatenation (ablation) | InterCMDM · Long-horizon InterHuman (segment FID) | FID: 18.37 -> 25.104 | Hurt |
| train with multi-task masks instead of simultaneous only (ablation) | InterCMDM · Reaction generation InterHuman | FID: 2.585 -> 2.121 | Improved |
| weighted (60/20/10/10) instead of even mask sampling (ablation) | InterCMDM · InterHuman ablation | FID: 5.247 -> 4.492 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- TIMotion — TIMotion: Temporal and Interactive Framework for Efficient Human-Human Motion Generation (CVPR 2025) · better on 6 of 6 shared results
- InterMask — InterMask: 3D Human Interaction Generation via Collaborative Masked Modeling (ICLR 2025) · better on 6 of 6 shared results
- InterGen — InterGen: Diffusion-Based Multi-human Motion Generation Under Complex Interactions (IJCV 2024) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Auto-Regressive Diffusion for Generating 3D Human-Object Interactions (AAAI 2025 · similarity 0.87)
- InterDreamer: Zero-Shot Text to 3D Dynamic Human-Object Interaction (NeurIPS 2024 · similarity 0.87)
- ReGenNet: Towards Human Action-Reaction Synthesis (CVPR 2024 · similarity 0.86)
- InterMask: 3D Human Interaction Generation via Collaborative Masked Modeling (ICLR 2025 · similarity 0.86)
- AMD: Autoregressive Motion Diffusion (AAAI 2024 · similarity 0.86)
- InterControl: Zero-shot Human Interaction Generation by Controlling Every Joint (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
