# ali-vilab/TeaCache

Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model

Repository: https://github.com/ali-vilab/TeaCache
Canonical: https://ross.abutalabs.com/products/teacache
Homepage: https://liewfeng.github.io/TeaCache/
Language: Python
License: Apache-2.0
License Family: permissive
Topics: diffusion-models, inference-acceleration, video-generation, cogvideox, hunyuan-video, latte, open-sora, open-sora-plan
Last push: 2025-06-08T14:29:09+00:00

## Health v2 (maintenance only)
Score: 33/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 25, release rhythm 35, longevity 46
- inputs: {"age_days": 644, "days_push": 451, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1369, forks 60 (observed 2026-08-28T04:04:31.742935+00:00)

## What it is
TeaCache is a training-free caching approach that accelerates inference for video diffusion models by estimating output differences across timesteps using timestep-embedding-modulated model inputs. It supports models like CogVideoX, HunyuanVideo, Latte, Open-Sora, and Open-Sora-Plan, achieving up to 4.41x speedup with negligible visual quality loss.

## Use cases
- speed up video diffusion model inference
- accelerate CogVideoX video generation
- reduce latency of HunyuanVideo denoising
- cache model outputs in Open-Sora pipelines
- training-free acceleration for video generation models
- make text-to-video generation faster on GPU

## When to choose
- you need faster video diffusion inference without retraining or fine-tuning
- you use a supported model like CogVideoX, HunyuanVideo, Latte, or Open-Sora
- you can tolerate minimal visual quality degradation for large speedups

## When to avoid
- you need exact, bit-identical outputs from the original diffusion model
- your video model is not among the supported architectures
- you need image-only diffusion acceleration rather than video

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, llm-inference, video-processing, caching
- domain: deep-learning, artificial-intelligence, image-processing
- platform: python
- tags: diffusion-models, inference-acceleration, video-generation, training-free, cogvideox, hunyuan-video, open-sora, latte, cvpr-2025, video, gpu, linux

## Member repositories
- ali-vilab/TeaCache (main) score 33

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:31.742935+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:40:57.660045+00:00, confidence not recorded.
  - readme: https://github.com/ali-vilab/TeaCache (fetched 2026-08-28T04:04:31.742935+00:00, sha ce2b6b4580b8)
  - homepage: https://liewfeng.github.io/TeaCache/ (fetched 2026-08-29T11:57:44.859554+00:00, sha 9bf7b864859f)
- Data as of 2026-08-30T08:39:29.467469+00:00.