ali-vilab/TeaCache

Implements timestep embedding-aware output caching that estimates and reuses intermediate model states across diffusion steps, eliminating the need for training. The approach dynamically exploits varying output similarities at different timesteps to reduce redundant computations without quality degradation. Supports multiple diffusion modalities including video (CogVideoX, HunyuanVideo, LTX-Video), image (FLUX, Lumina), and audio models through a unified caching framework.

1,290 stars. No commits in the last 6 months.