Research map / Image generation / Fast diffusion sampling
Physics in 2-Steps: Locking Motion Priors Before Visual Refinement Erases Them
Woojung Han, Seil Kang, Youngjun Jun, Min-Hung Chen, Fu-En Yang, Seong Jae Hwang
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 PhaseLock Latent Delta Guidance | CogVideoX-5B · Physics-IQ | score: 30.8 -> 36.0 | Improved |
| add PhaseLock Latent Delta Guidance | Wan 2.1 14B · Physics-IQ | score: 20.9 -> 28.7 | Improved |
| add PhaseLock to step-distilled model | Wan 2.1 Distilled 4-step · Physics-IQ | score: 27.7 -> 29.4 | Improved |
| add PhaseLock | CogVideoX · PhyGenBench (FLUX-schnell) | average score: 0.46 -> 0.57 | Improved |
| add PhaseLock | Wan 2.1 · PhyGenBench (Gemini-2.5 Flash) | average score: 0.46 -> 0.52 | Improved |
| add PhaseLock | Wan 2.1 · VBench aesthetic quality | aesthetic quality: 0.475 -> 0.451 | Hurt |
| add PhaseLock | CogVideoX-5B · VBench aesthetic quality | aesthetic quality: 0.467 -> 0.489 | Improved |
| increase denoising from 2 to 50 steps | base video diffusion · Physics-IQ | score: 34.02 -> 30.32 | Hurt |
| increase denoising from 2 to 50 steps | base video diffusion · LPIPS vs GT | LPIPS: 0.23 -> 0.19 | Improved |
| extend denoising from 50 to 100 steps (ablation) | base video diffusion · Physics-IQ | score: about 1 point | No clear effect |
| raise few-step NFE for motion prior above 2 (ablation) | PhaseLock · Physics-IQ | score | Hurt |
| raise motion strength above 0.05 (ablation) | PhaseLock · Physics-IQ | score | Hurt |
Closest papers by meaning
- SAND: Stage-Aware Noise Decomposition for Training-Free Diffusion Guidance (ECCV 2026 · similarity 0.88)
- PhyCo: Learning Controllable Physical Priors for Generative Motion (CVPR 2026 · similarity 0.88)
- LaMo: Self-Supervised Latent Motion Priors for Physical Realism in Video Generation (NeurIPS 2026 · similarity 0.87)
- Identifying and Solving Conditional Image Leakage in Image-to-Video Diffusion Model (NeurIPS 2024 · similarity 0.87)
- Cyclic Denoising Reveals Ultrastable Memories in Diffusion Models (NeurIPS 2026 · similarity 0.87)
- FlowMo: Variance-Based Flow Guidance for Coherent Motion in 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.
