Reproduce “Hierarchical Fine-grained Preference Optimization for Physically Plausible Video Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Harold Haodong Chen, Haojian Huang, Qifeng Chen, Harry Yang, Ser-Nam Lim
PhysHPO trains text-to-video generators to favor physically plausible clips using preferences at whole-video, boundary-state, motion, and semantic levels, alongside selection of real-world videos from an existing collection. On CogVideoX-5B, it raises VideoPhy Overall from 41.3 with vanilla DPO to 45.9. This measures joint passing of physical-law-consistency and semantic-coherence thresholds on 344 prompts, not universal physical correctness.
Key insights
- Achieving physical plausibility in video generation requires fine-grained, hierarchical preference optimization rather than only coarse-grained approaches.
- Automated selection of physically rich and diverse data from existing real-world datasets offers a more efficient and effective training strategy than extensive new data collection.
- Explicitly optimizing for physical consistency at multiple granularities improves a model's capacity for generalization, even enabling more coherent generation for semantically "impossible" scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.