Research map / Video / Video segmentation and tracking
TIMotion: Temporal and Interactive Framework for Efficient Human-Human Motion Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TIMotion, a temporal-then-interaction framework for two-person motion generation that models two person sequences as one causal interleaved sequence, adds role-evolving symmetric scanning and localized 1-D conv pattern amplification, and plugs into Transformer, Mamba, or RWKV mixers.
Problem
Human-human motion generation from text is typically done either by concatenating two persons into one sequence (single-person extension) or by modeling them separately with cross-attention, both of which model interactions inadequately and carry redundant parameters.
Approach
Interleave the two persons' per-frame motion tokens into one causal sequence (Causal Interactive Injection), concatenate a symmetric ordering so both persons can take active and passive roles (Role-Evolving Scanning), and add an AdaLN-conditioned 1-D convolution branch per person for short-term patterns (Localized Pattern Amplification), all within a diffusion denoiser.
Key results
- TIMotion+RWKV reaches FID 4.702 and Top-1 R-Precision 0.501 on InterHuman, versus 5.918 and 0.371 for InterGen.
- TIMotion+Transformer uses 65M parameters versus 182M for InterGen, with inference 0.632 s per sample versus 1.991 s.
- In the in-betweening editing task, TIMotion gets Top-1 0.516 and FID 3.590 versus InterGen's 0.461 and 4.700.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace separate modeling with Causal Interactive Injection (ablation) | Transformer human-human motion generation · InterHuman | R-Precision Top 1: 0.371±.010 -> 0.491±.005 | Helped |
| replace separate modeling with Causal Interactive Injection (ablation) | Transformer human-human motion generation · InterHuman | FID: 5.918±.079 -> 5.433±.080 | Helped |
| replace separate modeling with Causal Interactive Injection (ablation) | Mamba human-human motion generation · InterHuman | R-Precision Top 1: 0.420±.005 -> 0.431±.004 | Helped |
| replace separate modeling with Causal Interactive Injection (ablation) | Mamba human-human motion generation · InterHuman | FID: 7.221±.097 -> 6.142±.059 | Helped |
| replace separate modeling with Causal Interactive Injection (ablation) | RWKV human-human motion generation · InterHuman | R-Precision Top 1: 0.465±.005 -> 0.501±.005 | Helped |
| replace separate modeling with Causal Interactive Injection (ablation) | RWKV human-human motion generation · InterHuman | FID: 5.943±.079 -> 4.702±.069 | Helped |
| add Causal Interactive Injection (ablation) | InterGen with RWKV · InterHuman | R-Precision Top 1: 0.371±.010 -> 0.478±.005 | Helped |
| add Causal Interactive Injection (ablation) | InterGen with RWKV · InterHuman | FID: 5.918±.079 -> 5.410±.069 | Helped |
| add Role-Evolving Scanning to CII (ablation) | InterGen with RWKV · InterHuman | R-Precision Top 1: 0.478±.005 -> 0.494±.006 | Helped |
| add Role-Evolving Scanning to CII (ablation) | InterGen with RWKV · InterHuman | FID: 5.410±.069 -> 5.019±.077 | Helped |
| add Localized Pattern Amplification (ablation) | CII+RES · InterHuman | R-Precision Top 1: 0.494±.006 -> 0.501±.005 | Helped |
| add Localized Pattern Amplification (ablation) | CII+RES · InterHuman | FID: 5.019±.077 -> 4.702±.069 | Helped |
Surprising findings
- Single-person extension with Transformer performs worse than the separate modeling baseline on FID for Mamba-style mixers, showing the benefit of temporal modeling depends on the mixer.
Limitations
- Gains are reported only on two human-human motion datasets, InterHuman and Inter-X.
- The Mamba variant underperforms the RWKV variant, and the RWKV variant has more parameters than the Transformer variant.
Experimental setup
Datasets: InterHuman Inter-X
Benchmarks: InterHuman Inter-X
Metrics: FID R-Precision Diversity MultiModality MM Dist
Baselines: TEMOS T2M MDM ComMDM RIG InterGen
Builds on: InterGen MDM DDIM classifier-free guidance
Closest papers by meaning
- SemGeoMo: Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance (CVPR 2025 · similarity 0.89)
- InterControl: Zero-shot Human Interaction Generation by Controlling Every Joint (NeurIPS 2024 · similarity 0.89)
- HuTuMotion: Human-Tuned Navigation of Latent Motion Diffusion Models with Minimal Feedback (AAAI 2024 · similarity 0.88)
- ReGenNet: Towards Human Action-Reaction Synthesis (CVPR 2024 · similarity 0.87)
- Human Motion Diffusion as a Generative Prior (ICLR 2024 · similarity 0.87)
- InterDreamer: Zero-Shot Text to 3D Dynamic Human-Object Interaction (NeurIPS 2024 · similarity 0.87)
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.