SkyReels-A1
Project lead · 2025
Expression-driven portrait animation on video diffusion transformers, addressing identity drift and facial dynamics.
- Video diffusion transformer
- Portrait animation
- Expression transfer
Portrait animation tends to fail in three recognisable ways: identity drifts across the sequence, the background becomes unstable, and facial dynamics look subtly wrong. All three get worse in head-only animation. Extending the same model to different body proportions usually introduces visual inconsistency and unnatural articulation on top of that.
SkyReels-A1 addresses these inside a video diffusion transformer. The work is written up as a paper and the implementation is open source.