Reproduce “YingVideo-MV: Music-Driven Multi-Stage 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

Jiahui Chen, Weida Wang, Runhua Shi, Huan Yang, Chaofan Ding, Zihao Chen

YingVideo-MV is a music-driven multi-stage framework that generates high-quality, long-duration music performance videos featuring expressive virtual avatars. It achieves precise control over facial expressions, body dynamics, and camera movements, demonstrating superior visual quality, identity preservation, and lip-sync accuracy compared to existing models.

Key insights

  • A multi-stage cascaded architecture, combining high-level narrative planning with low-level video synthesis, effectively enables comprehensive music-driven video generation with cinematic quality.
  • Explicitly conditioning diffusion models on camera parameters via Plücker embeddings within the latent space provides precise and smooth control over camera movements.
  • Optimizing the generation process with Direct Preference Optimization (DPO) and a Timestep-aware Dynamic Window Range Strategy significantly enhances video quality, multi-modal synchronization, and temporal coherence for long sequences.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.