Research map / Video / Video generation
AAMDM: Accelerated Auto-regressive Motion Diffusion Model
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 AAMDM, an auto-regressive motion diffusion model that combines a Denoising Diffusion GAN generator with a diffusion polishing module in a learned low-dimensional embedded pose space, achieving near-AMDM200 quality at roughly 40x its speed.
Problem
Interactive character motion synthesis needs diverse, high-quality, contextually responsive animation at real-time rates, but diffusion models require expensive reverse diffusion steps and Motion Matching scales poorly.
Approach
Generate a draft motion with a few Denoising Diffusion GAN steps (generation module), then refine it with two auto-regressive diffusion steps (polishing module), operating in a learned 64-dim embedded space rather than the 338-dim full pose space, with guided diffusion for user control.
Key results
- AAMDM achieves FID 14.051 and DIV 11.574 on random synthesis, close to AMDM200 (FID 12.132, DIV 11.165) while running at 173 FPS vs 4.72 FPS (about 40x faster).
- AAMDM outperforms LMM (FID 49.706), MVAE (FID 22.981) and AMDM5 (FID 18.741) on random motion metrics.
- Ablation: removing the polishing module (T_ADM=0) causes divergence in long-horizon generation; replacing the embedded space with the full pose space raises FID to 128.412.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AAMDM instead of AMDM200 | AMDM200 · LaFAN1 random motion synthesis | FPS: 4.72 -> 173 | Helped |
| Use AAMDM instead of AMDM200 | AMDM200 · LaFAN1 random motion synthesis | FID: 12.132 -> 14.051 | Hurt |
| Add a polishing module rather than using no polishing steps (ablation) | AAMDM · LaFAN1 random motion synthesis | FID: N/A -> 14.051 | Helped |
| Increase polishing steps from 2 to 10 (ablation) | AAMDM · LaFAN1 random motion synthesis | FID: 14.051 -> 11.779 | Helped |
| Increase polishing steps from 2 to 10 (ablation) | AAMDM · LaFAN1 random motion synthesis | FPS: 173 -> 47 | Hurt |
| Use 2 rather than 3 generation steps (ablation) | AAMDM · LaFAN1 random motion synthesis | FID: 14.051 -> 28.476 | Hurt |
| Use 4 rather than 3 generation steps (ablation) | AAMDM · LaFAN1 random motion synthesis | FID: 14.051 -> 14.534 | No clear effect |
| Use 10 rather than 3 generation steps (ablation) | AAMDM · LaFAN1 random motion synthesis | FID: 14.051 -> 16.312 | Hurt |
| Remove the embedded transition space and learn in full pose space (ablation) | AAMDM · LaFAN1 random motion synthesis | FID: 14.051 -> 128.412 | Hurt |
| Remove the polishing module (ablation) | AAMDM · LaFAN1 long-horizon motion synthesis | motion quality | Hurt |
| Use AAMDM instead of AMDM5 | AMDM5 · LaFAN1 random motion synthesis | FID: 18.741 -> 14.051 | Helped |
| Use AAMDM instead of LMM | LMM · LaFAN1 random motion synthesis | FID: 49.706 -> 14.051 | Helped |
Surprising findings
- Using only DD-GAN generation steps without the polishing module fails to produce reliable long-horizon trajectories.
- Increasing DD-GAN generation steps to 10 gave the worst quality, attributed to the simple MLP architecture.
Limitations
- Slower than single feed-forward baselines such as LMM and MVAE due to five network evaluations per frame.
- Performance degrades with larger numbers of DD-GAN generation steps given the simple MLP networks.
- Controllability relies on gradient-based sampling guidance rather than a learned control mechanism.
Experimental setup
Datasets: LaFAN1 Squ-9 (synthetic 2D Gaussian many-to-many dataset)
Benchmarks: LaFAN1
Metrics: Diversity FID FFR FPS TE-UC FID-UC FFR-UC
Baselines: Learned Motion Matching MotionVAE AMDM5 AMDM200
Builds on: Denoising Diffusion GANs DDPM Autoregressive Motion Diffusion Model (AMDM) Learned Motion Matching Motion VAE
Closest papers by meaning
- Single Motion Diffusion (ICLR 2024 · similarity 0.88)
- Diffusion Models Are Real-Time Game Engines (ICLR 2025 · similarity 0.88)
- Seamless Human Motion Composition with Blended Positional Encodings (CVPR 2024 · similarity 0.86)
- MotionDreamer: One-to-Many Motion Synthesis with Localized Generative Masked Transformer (ICLR 2025 · similarity 0.86)
- Generative Rendering: Controllable 4D-Guided Video Generation with 2D Diffusion Models (CVPR 2024 · similarity 0.86)
- Lost in Latent Space: An Empirical Study of Latent Diffusion Models for Physics Emulation (NeurIPS 2025 · 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.