Reproduce “MagicFight: Personalized Martial Arts Combat 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
Jiancheng Huang, Mingfu Yan, Songyan Chen, Yi Huang, Shifeng Chen
Researchers from SIAT, Chinese Academy of Sciences, developed MagicFight, a framework for generating personalized two-person martial arts combat videos using adapted diffusion models. The system, along with the new KungFu-Fiesta dataset, achieves higher fidelity and temporal consistency compared to previous single-person video generation approaches on combat video benchmarks.
Key insights
- Generating complex, dynamic multi-person interactions requires specialized datasets and architectural components, as existing single-person models are insufficient.
- The Personalized Attention Layer (ID-attn) is crucial for maintaining distinct character identities and avoiding feature mixing during close interactions.
- Combining a custom martial arts dataset (KFF) with recreated two-person fashion videos through mixture dataset finetuning significantly improves realism and aesthetic quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.