Research map / Video / Video generation
PhysRVG: Physics-Aware Unified Reinforcement Learning for Video Generative Models
Qiyuan Zhang, Biao Gong, Shuai Tan, Zheng Zhang, Xing Zhu, Yujun Shen, Yuyuan Li, kelu Yao, Chunhua Shen, Changqing Zou
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 MDcycle RL post-training with physics-grounded reward | Wan2.2 5B · PhysRVGBench | IoU / TO: IoU 0.15 -> 0.64; TO 162.78 -> 15.03 | Improved |
| add plain RL after full fine-tuning (ablation) | FT model · PhysRVGBench | TO: 46.27 -> 17.25 | Improved |
| replace plain RL by MDcycle (add flow-matching mimicry branch) (ablation) | FT+RL · PhysRVGBench | TO: 17.25 -> 15.03 | Improved |
| use full-parameter fine-tuning vs LoRA only (ablation) | Baseline · PhysRVGBench | IoU: LoRA 0.41 vs FT 0.38 | No clear effect |
| add collision-aware weights (1,2,3) vs uniform weights (ablation) | reward · PhysRVGBench | TO: 15.41 -> 15.03 | Improved |
| raise collision weight to 5 (ablation) | reward · PhysRVGBench | TO: 15.03 -> 16.17 | Hurt |
| restrict SDE sampling to high-noise window (75-100%) with 2 steps (ablation) | RL sampling · PhysRVGBench | TO: 27.24 -> 15.03 | Improved |
| set mimicry Threshold to 8 vs 4 or 12 (ablation) | MDcycle · PhysRVGBench | TO: 22.39 (T=4), 16.97 (T=12) -> 15.03 | Improved |
| use 16k vs 6k FT steps before MDcycle (ablation) | MDcycle · PhysRVGBench | TO: 37.23 -> 15.03 | Improved |
| use PEFT instead of full-parameter training inside RL (ablation) | MDcycle · training stability | reward curve | Improved |
| use PhysRVG vs Kling2.5 on VBench total score | video generator · VBench | Total Score: 78.77 -> 78.89 | No clear effect |
Closest papers by meaning
- PhysVid: Physics Aware Local Conditioning for Generative Video Models (CVPR 2026 · similarity 0.92)
- PhyCo: Learning Controllable Physical Priors for Generative Motion (CVPR 2026 · similarity 0.91)
- PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation (NeurIPS 2025 · similarity 0.91)
- PhyMotion: Structured 3D Motion Reward for Physics-Grounded Human Video Generation (NeurIPS 2026 · similarity 0.89)
- How Far Is Video Generation from World Model: A Physical Law Perspective (ICML 2025 · similarity 0.88)
- Force Prompting: Video Generation Models Can Learn And Generalize Physics-based Control Signals (NeurIPS 2025 · similarity 0.88)
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.
