Reproduce “Learning Explicit Physical Parameter Control and Benchmarking for 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
Yanxun Li, Hao Wen, Bingze Song, Jiashu Zhu, Aiming Hao, Chubin Chen, Jintao Chen, Jiahong Wu and 2 more
Researchers from Beihang University and Alibaba Group introduced a framework for image-to-video generation, enabling explicit, instance-level physical parameter control for rigid-body motion. They developed a large-scale, physically parameterized dataset, a physics-guided diffusion model, and a benchmark to evaluate physical law consistency, achieving superior motion consistency (FVMD 355.5) and higher human preference rates (up to 75.17% for semantic consistency) compared to general-purpose baselines.
Key insights
- Explicit physical parameter encoding combined with Physical Parameter Attention effectively enables fine-grained, instance-level control of physical attributes and prevents cross-condition interference in multi-object scenes.
- Semantic-structural feature-space supervision, utilizing pre-trained DINOv3 features, significantly improves object identity, shape, boundaries, and spatial displacement alignment, enhancing both physical consistency and temporal coherence.
- A dedicated, multi-level benchmark focused on physical law consistency, including metrics for long-tail scenarios, is crucial for diagnosing and advancing physically-grounded video generation models effectively.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.