Research map / Video / Video generation
NS-Diff: Fluid Navier-Stokes Guided Video Diffusion via Reinforcement Learning
Zijun Deng, Yuxin Peng
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-guided RL with latent injection to OpenSora2-11B | OpenSora2-11B · PhysVideoBench test | Jerk error (ΔJ): 0.69 -> 0.25 | Improved |
| add physics-guided RL with latent injection to OpenSora2-11B | OpenSora2-11B · UCF-101 | FVD: 110 -> 85 | Improved |
| add physics-guided RL with latent injection to OpenSora2 | OpenSora2 · WebVid-10M text-to-video | FVD: 297 -> 275 | Improved |
| add PPO optimization on top of the physics losses (ablation) | NS-Diff-DiT 1B · PhysVideoBench (1B DiT ablation) | Jerk error (ΔJ): 0.58 -> 0.33 | Improved |
| add physics-conditioned latent injection (ablation) | NS-Diff-DiT 1B · PhysVideoBench (1B DiT ablation) | Jerk error (ΔJ): 0.82 -> 0.33 | Improved |
| add physics constraints (rigid and fluid losses) (ablation) | NS-Diff-DiT 1B · PhysVideoBench (1B DiT ablation) | Jerk error (ΔJ): 1.21 -> 0.33 | Improved |
| add adaptive activation scheduler (ablation) | NS-Diff-DiT 1B · PhysVideoBench (1B DiT ablation) | Jerk error (ΔJ): 0.67 -> 0.33 | Improved |
| use adaptive scheduler instead of linear scheduler (ablation) | NS-Diff-DiT 1B · PhysVideoBench (1B DiT ablation) | FVD: 207 -> 183 | Improved |
| add fluid Navier-Stokes loss (ablation) | NS-Diff-DiT 1B · PhysVideoBench (1B DiT ablation) | Fluid divergence error (Ldiv, x10^-3): 10.4 -> 2.9 | Improved |
| add rigid jerk loss (ablation) | NS-Diff-DiT 1B · PhysVideoBench (1B DiT ablation) | Jerk error (ΔJ): 1.33 -> 0.33 | Improved |
| remove fluid loss on jerk metric (ablation) | NS-Diff-DiT 1B · PhysVideoBench (1B DiT ablation) | Jerk error (ΔJ): 0.33 -> 0.49 (w/o Lfluid) | Mixed |
| estimate flow on decoded RGB proxies instead of directly in latent space (ablation) | ARFlow · PhysVideoBench (1B DiT ablation) | FVD: 205 -> 183 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PhysGen — PhysGen: Rigid-Body Physics-Grounded Image-to-Video Generation (ECCV 2024) · better on 4 of 4 shared results
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
- LaMo: Self-Supervised Latent Motion Priors for Physical Realism in Video Generation (NeurIPS 2026 · similarity 0.87)
- DreamPhysics: Learning Physics-Based 3D Dynamics with Video Diffusion Priors (AAAI 2025 · similarity 0.87)
- PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation (NeurIPS 2025 · similarity 0.87)
- FlowMo: Variance-Based Flow Guidance for Coherent Motion in Video Generation (NeurIPS 2025 · similarity 0.86)
- Lost in Latent Space: An Empirical Study of Latent Diffusion Models for Physics Emulation (NeurIPS 2025 · similarity 0.86)
- MotionCraft: Physics-Based Zero-Shot Video Generation (NeurIPS 2024 · similarity 0.85)
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.
