Research map / Video / Video generation
AMD: Autoregressive Motion Diffusion
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 AMD, an autoregressive motion diffusion model that generates long compound 3D human motions from sequential text prompts, plus a new HumanLong3D dataset pairing long text with complex motions.
Problem
Text-conditioned human motion generation struggles with long prompts and complex compound motions due to scarce paired data and a many-to-many mapping between text and motion distributions.
Approach
Autoregressively generates motion segments by conditioning a diffusion model on the current text prompt plus the previous prompt and previous motion segment, with a duration predictor for adaptive length and explicit geometric losses (height, position, rotation, velocity, foot sliding).
Key results
- On HumanLong3D compound motion, Ours reaches FID 0.225 versus 1.249 for T2M-GPT and 27.348 for MDM, with R-Precision 0.158.
- On AIST++ music-to-dance, Ours achieves BAS 0.2302, second only to Bailando (0.2332), with FID_k 30.28 versus Bailando's 28.16.
- Ours ranks top-3 on four of five HumanLong3D metrics, with FID and Diversity ranking first.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AMD instead of T2M-GPT for compound motion generation | Text-to-motion generation · HumanLong3D | FID: 1.249^{\pm.026} -> 0.225^{\pm.013} | Helped |
| Use AMD instead of T2M-GPT for compound motion generation | Text-to-motion generation · HumanLong3D | R-Precision (Top3): 0.159^{\pm.005} -> 0.158^{\pm.006} | No clear effect |
| Use AMD instead of Bailando without its customized reinforcement learning module | Music-to-dance generation · AIST++ | BAS: 0.2332 -> 0.2302 | Hurt |
| Use AMD instead of Bailando | Music-to-dance generation · AIST++ | FID_k: 28.16 -> 30.28 | Hurt |
| Use AMD instead of Bailando | Music-to-dance generation · AIST++ | Div_g: 6.34 -> 6.29 | No clear effect |
Surprising findings
- Joint prediction (Ours-J) has the highest MultiModality score yet very poor FID, which the authors attribute to chaotic motions at low quality.
- T2M-GPT performs well on short single motions but its generated long compound motions cut off early because it predicts the end token prematurely.
Limitations
- Music-to-dance results trail Bailando on FID_k and BAS-related performance, not SOTA on all metrics.
- MultiModality score for the proposed method (1.242) is lower than several baselines, indicating less diversity across samples for the same prompt.
Experimental setup
Datasets: HumanLong3D HumanML3D HumanMusic AIST++ AMASS HumanAct12
Benchmarks: HumanLong3D HumanML3D AIST++
Metrics: FID R-Precision (Top 3) MultiModal Dist Diversity MultiModality FID_k FID_g BAS
Baselines: MDM MotionDiffuse T2M-GPT MLD DanceNet DanceRevolution FACT Bailando
Builds on: MDM T2M CLIP TEACH
Closest papers by meaning
- Rethinking Diffusion for Text-Driven Human Motion Generation: Redundant Representations, Evaluation, and Masked Autoregression (CVPR 2025 · similarity 0.91)
- AMD: Anatomical Motion Diffusion with Interpretable Motion Decomposition and Fusion (AAAI 2024 · similarity 0.91)
- Human Motion Diffusion as a Generative Prior (ICLR 2024 · similarity 0.90)
- MotionMix: Weakly-Supervised Diffusion for Controllable Motion Generation (AAAI 2024 · similarity 0.89)
- InterDreamer: Zero-Shot Text to 3D Dynamic Human-Object Interaction (NeurIPS 2024 · similarity 0.89)
- Seamless Human Motion Composition with Blended Positional Encodings (CVPR 2024 · similarity 0.89)
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.