Reproduce “DragAnything: Motion Control for Anything using Entity Representation” 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
Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li and 2 more
Researchers at Kuaishou Technology, Zhejiang University, and NUS Show Lab developed DragAnything, a framework for precise, user-friendly motion control of any entity in generated videos using a novel entity representation. It achieved an FID of 33.5 and an FVD of 494.8, outperforming previous trajectory-based methods like DragNUWA, and was preferred by human evaluators for both motion control and video quality.
Key insights
- Previous trajectory-based methods for video motion control often misinterpret user intent because single pixel points do not adequately represent an entire semantic entity.
- Directly dragging pixels disproportionately influences nearby areas, causing object deformation rather than unified, coherent movement.
- Leveraging deep latent diffusion features enables the extraction of robust semantic representations for diverse entities, facilitating more precise and intuitive control.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.