#5757th of 89 in Image and videoFor image, video and design, video and animation

turbodiffusion

Speeds up Wan video diffusion models with distilled 4-step sampling

Stars
3.9k
License
Apache-2.0
Last commit
Nov 2026
Language
Python
Stars, last 29 weeks+432
25 Apr 2026: 3,433 stars2 May 2026: 3,443 stars9 May 2026: 3,454 stars16 May 2026: 3,462 stars23 May 2026: 3,471 stars30 May 2026: 3,482 stars6 Jun 2026: 3,493 stars13 Jun 2026: 3,501 stars20 Jun 2026: 3,507 stars27 Jun 2026: 3,516 stars4 Jul 2026: 3,522 stars11 Jul 2026: 3,537 stars18 Jul 2026: 3,555 stars25 Jul 2026: 3,567 stars1 Aug 2026: 3,576 stars8 Aug 2026: 3,588 stars15 Aug 2026: 3,594 stars22 Aug 2026: 3,604 stars29 Aug 2026: 3,614 stars5 Sept 2026: 3,624 stars12 Sept 2026: 3,629 stars19 Sept 2026: 3,728 stars26 Sept 2026: 3,832 stars3 Oct 2026: 3,838 stars10 Oct 2026: 3,841 stars17 Oct 2026: 3,847 stars24 Oct 2026: 3,848 stars31 Oct 2026: 3,856 stars7 Nov 2026: 3,859 stars14 Nov 2026: 3,865 stars

3,433 on 25 Apr 20263,865 on 14 Nov 2026

From GitHub's daily star counts.

Checked on GitHub.

Overview

TurboDiffusion accelerates text-to-video and image-to-video generation for Wan2.1 and Wan2.2 models. It combines SageAttention and Sparse-Linear Attention (SageSLA), W8A8 quantized linear layers, and rCM timestep distillation to sample in 1-4 steps. The README reports a 5-second 480p clip in 1.9s versus 184s originally, and a 720p I2V run in 38s versus 4549s, both on one RTX 5090.

Who it is for: Engineers running Wan video generation on RTX 5090/4090 or H100 GPUs

Strengths

  • Reported 100-200x end-to-end speedup on one RTX 5090, excluding text encoding and VAE decoding
  • Quantized checkpoints for RTX 5090/4090 and unquantized ones for 40GB+ GPUs like H100
  • Covers Wan2.1 T2V (1.3B, 14B) and Wan2.2 I2V A14B at 480p and 720p
  • Terminal interactive mode allows multi-turn generation without reloading the model

Weaknesses

  • Models trained only on long English prompts; other prompts need augmentation
  • README says checkpoints and paper are not finalized and will be updated
  • Builds with pip --no-build-isolation; torch 2.8.0 recommended, newer versions may OOM
  • H3 model license excludes the US, EU, UK and Republic of Korea

What it needs

  • GPU required
  • Needs PyTorch >=2.7.0 (2.8.0 recommended), Python >=3.9, SageAttention, SpargeAttn (optional, for SageSLA), Wan2.1 VAE and umT5 text encoder checkpoints
  • Models: Wan2.1-T2V-1.3B-480P, Wan2.1-T2V-14B-480P, Wan2.1-T2V-14B-720P, Wan2.2-I2V-A14B-720P, TurboH3-FL2VA-I2VA

Also in Image and video

See all 89
Also in Image and video
RankProjectScore
1frigateLocal NVR with realtime object detection for IP cameras37.6k starsMIT93 out of 100
2MoneyPrinterTurboGenerates short videos from a topic, with script, footage, voiceover and subtitles144.3k starsMIT88 out of 100
3toonflow-appInfinite-canvas tool for AI short drama and comic video production22.1k starsMIT88 out of 100
#4InvokeAIWeb UI for diffusion image generation with inpainting canvas and node workflows29.1k starsApache-2.087 out of 100
#5ComfyUINode-graph editor for running diffusion image, video, audio and 3D workflows147.4k starsGPL-3.085 out of 100