# thu-ml/TurboDiffusion

TurboDiffusion: 100–200× Acceleration for Video Diffusion Models

Repository: https://github.com/thu-ml/TurboDiffusion
Canonical: https://ross.abutalabs.com/products/turbodiffusion
Homepage: https://arxiv.org/pdf/2512.16093
Language: Python
License: Apache-2.0
License Family: permissive
Topics: ai-infra, diffusion-models, inference-acceleration, rcm, sageattention, sparse-linear-attention, video-generation, mlsystem, consistency-model, distillation
Last push: 2026-08-05T05:17:13+00:00

## Health v2 (maintenance only)
Score: 59/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 96, release rhythm 35, longevity 19
- inputs: {"age_days": 271, "days_push": 28, "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 3623, forks 275 (observed 2026-08-28T04:08:11.909295+00:00)

## What it is
TurboDiffusion is a Python framework that accelerates end-to-end video diffusion model generation by 100-200x using SageAttention, Sparse-Linear Attention, and rCM timestep distillation. It ships accelerated TurboWan checkpoints for text-to-video and image-to-video generation on consumer GPUs like the RTX 5090.

## Use cases
- generate videos from text prompts much faster
- speed up video diffusion model inference
- run video generation on a single consumer GPU
- accelerate Wan video models with distilled checkpoints
- reduce diffusion sampling steps for video generation
- generate 480p or 720p videos in seconds instead of minutes

## When to choose
- you need fast video generation with minimal quality loss on limited GPU hardware
- you want to use Wan 2.1/2.2 video models with dramatically reduced inference time
- you are doing research or building apps around accelerated video diffusion

## When to avoid
- you need video models other than the provided TurboWan checkpoints
- you require maximum video quality over generation speed
- you work with non-English prompts without prompt augmentation
- you need a finalized, production-hardened release (checkpoints and paper are still being updated)

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, machine-learning, video-processing, gpu-computing
- domain: deep-learning, artificial-intelligence, gpu-computing
- platform: python
- tags: diffusion-models, inference-acceleration, video-generation, sageattention, sparse-linear-attention, consistency-model, timestep-distillation, ai-infra, video, gpu, linux

## Member repositories
- thu-ml/TurboDiffusion (main) score 59

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:08:11.909295+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 2026-08-29T18:31:57.022396+00:00, confidence not recorded.
  - readme: https://github.com/thu-ml/TurboDiffusion (fetched 2026-08-28T04:08:11.909295+00:00, sha 6ce26767b989)
  - registry_pypi: https://pypi.org/pypi/turbodiffusion/json (fetched 2026-08-29T09:26:35.407441+00:00, sha bd03c6bf162a)
- Data as of 2026-08-30T08:39:29.467469+00:00.