PKU-YuanGroup/ConsisID

Decomposes identity features across frequency components to maintain facial consistency without fine-tuning, operating as a plug-and-play module for DiT-based video diffusion models. The approach leverages frequency analysis insights from vision transformers to selectively preserve identity information while allowing stylistic variation. Integrates with Hugging Face Diffusers (v0.33.0+) and supports multiple backbone architectures including CogVideoX series, with optimized inference via TeaCache and xDiT frameworks.

835 stars. Actively maintained with 3 commits in the last 30 days.