TeaCache

Consecutive diffusion timesteps often produce similar outputs, and TeaCache exploits that: the Timestep Embedding Aware Cache estimates how much a model's output will change from its timestep embeddings and reuses cached computation when the predicted difference is small, accelerating inference with no retraining. Reported results include up to 4.41x acceleration on Open-Sora-Plan with a negligible 0.07 percent VBench score drop, around 2.1x on HunyuanVideo, and up to 2.25x on FLUX. The official ali-vilab repository ships per-model integrations for more than a dozen systems, spanning video models such as Wan 2.1, CogVideoX, LTX-Video, Mochi, and Cosmos, image models including FLUX, Lumina-Image 2.0, and HiDream-I1, and the TangoFlux audio model, each in its own subdirectory with example scripts, and a single threshold parameter trades speed against fidelity. The paper was accepted to CVPR 2025 as a highlight. Code is Apache-2.0 licensed with around 1,350 stars, and the technique has been adopted by downstream inference stacks such as xDiT and numerous ComfyUI workflows.