Research map / Video / Video generation
PhysDiff: Physics-Guided Human Motion Diffusion Model
Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, Jan Kautz
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| add physics-based projection during diffusion sampling to MDM | MDM · HumanML3D text-to-motion | Phys-Err: 31.572 -> 4.111 | Improved |
| add physics-based projection during diffusion sampling to MDM | MDM · HumanML3D text-to-motion | FID: 0.544 -> 0.433 | Improved |
| add physics-based projection to MotionDiffuse | MotionDiffuse · HumanML3D text-to-motion | Phys-Err: 30.652 -> 2.690 | Improved |
| add physics-based projection to MDM on action-to-motion | MDM · UESTC action-to-motion | FID: FID 12.81 -> 13.27; Phys-Err 28.371 -> 1.463 | Mixed |
| increase number of projection steps beyond about four (ablation) | PhysDiff · HumanML3D text-to-motion | FID / R-Precision | Hurt |
| increase number of projection steps (ablation) | PhysDiff · HumanML3D text-to-motion | Phys-Err | Improved |
| place projection steps consecutively at end (End 4, Space 1) instead of uniform spread (ablation) | PhysDiff · HumanML3D text-to-motion | Phys-Err: 4.921 -> 4.111 | Improved |
| place projection steps at start of diffusion (Start 3, End 1) (ablation) | PhysDiff · HumanML3D text-to-motion | FID: 0.433 -> 0.510 | Hurt |
| reduce spacing between end projections from 3 to 1 (ablation) | PhysDiff · HumanML3D text-to-motion | Phys-Err: 4.689 -> 4.111 | Improved |
| apply physics projection as multi-step post-processing instead of inside diffusion (ablation) | MDM · HumanML3D text-to-motion | motion quality / physical plausibility | Hurt |
| add physics simulation to sampling (inference cost) | MDM · inference time, single motion | seconds: 19.6s -> 51.6s | Hurt |
Later papers that compared with this method
2 later accepted papers report PhysDiff as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- FreqPhys: Repurposing Implicit Physiological Frequency Prior for Robust Remote Photoplethysmography (ECCV 2026 · better on 5 of 5 shared results)
- PHASE-Net: Physics-Grounded Harmonic Attention System for Efficient Remote Photoplethysmography Measurement (CVPR 2026 · better on 1 of 1 shared result)
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
- Human Joint Kinematics Diffusion-Refinement for Stochastic Motion Prediction (AAAI 2023 · similarity 0.90)
- Nonisotropic Gaussian Diffusion for Realistic 3D Human Motion Prediction (CVPR 2025 · similarity 0.88)
- Human Motion Diffusion as a Generative Prior (ICLR 2024 · similarity 0.88)
- Optimizing Diffusion Noise Can Serve As Universal Motion Priors (CVPR 2024 · similarity 0.87)
- MotionMix: Weakly-Supervised Diffusion for Controllable Motion Generation (AAAI 2024 · similarity 0.87)
- PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation (NeurIPS 2025 · 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.
