ali-vilab/TeaCache
An inference acceleration technique for video diffusion models that caches intermediate computations based on timestep embeddings.
Not currently ranked — collecting fresh signals.
star history
TeaCache proposes a method to speed up video diffusion model inference by caching and reusing timestep-dependent features. It targets multiple video generation frameworks including CogVideoX, Open-Sora, and Latte. The approach analyzes timestep embeddings to determine when to reuse cached computations without sacrificing output quality, achieving significant speedups in text-to-video generation.
Frequently asked
- What is ali-vilab/TeaCache?
- An inference acceleration technique for video diffusion models that caches intermediate computations based on timestep embeddings.
- Is TeaCache open source?
- Yes — ali-vilab/TeaCache is open source, released under the Apache-2.0 license.
- What language is TeaCache written in?
- ali-vilab/TeaCache is primarily written in Python.
- How popular is TeaCache?
- ali-vilab/TeaCache has 1.4k stars on GitHub.
- Where can I find TeaCache?
- ali-vilab/TeaCache is on GitHub at https://github.com/ali-vilab/TeaCache.