DanceTogether! Identity-Preserving Multi-Person Interactive Video Generation
Junhao Chen, Mingjin Chen, Jianjin Xu, Xiang Li, Junting Dong, Mingze Sun, Puhua Jiang, Hongxiang Li, Yuhang Yang, Hao Zhao, Xiaoxiao Long, Ruqi Huang
2025-05-26
Summary
This paper talks about DanceTogether, a new AI system that can create realistic videos of several people interacting and dancing together, starting from just one photo of each person and some information about their movements.
What's the problem?
The problem is that making high-quality videos with multiple people who look like themselves and move naturally is very difficult for current AI models, especially when you only have a single picture of each person to start with.
What's the solution?
The researchers built DanceTogether, which uses a special technique called diffusion to generate long and lifelike videos where each person keeps their unique appearance and interacts smoothly with others. The system uses both the original images and pose information to guide the video creation, and it works better than previous methods.
Why it matters?
This is important because it opens up new possibilities for creating custom videos for entertainment, social media, or virtual reality, allowing people to see themselves and their friends in creative group videos without needing to film anything in real life.
Abstract
DanceTogether, an end-to-end diffusion framework, generates long, photorealistic multi-actor interaction videos from single reference images and pose-mask streams, outperforming existing systems.