whereismyai
TurboDiffusion
TurboDiffusion is an open-source video generation acceleration framework that speeds up end-to-end diffusion model inference by 100 to 200 times.
TurboDiffusion provides an official implementation for accelerating video diffusion models. It employs SageAttention, Sparse-Linear Attention (SLA), and rCM for timestep distillation to achieve inference speeds 100 to 200 times faster than original models. The framework supports models such as Wan2.1 and Wan2.2, allowing for both text-to-video and image-to-video generation at 480p or 720p resolutions. Users can run inference on high-end hardware like H100 or RTX 5090 GPUs, with support for quantized and unquantized checkpoints. It is designed for researchers and developers requiring high-performance video generation. The project also includes training code and provides infrastructure support including FSDP2 and Ulysses CP for custom model training.
TurboDiffusion at a glance
- Best for: Software engineer
- Runs on: Open source, CLI, Self-hosted
- Pricing: Free
- Released: December 2025
- Category: AI video tool
- Website: github.com/thu-ml/TurboDiffusion
Key features
- Accelerated video generation using Sparse-Linear Attention
- Timestep distillation with rCM
- Support for text-to-video and image-to-video inference
- Quantized and unquantized model checkpoint support
- Training code with FSDP2 and Ulysses CP support
- Multi-turn video generation via terminal interface
- ComfyUI integration via community efforts
Use it for
- Generating 480p or 720p videos from text prompts
- Creating videos from images with I2V models
- Accelerating existing Wan-2.1 and Wan-2.2 model inference
- Training video diffusion models with sparse-linear attention
Alternatives to TurboDiffusion
- FastVideo — Provides end-to-end post-training and real-time inference for accelerated video generation.
- fal — Access 1,000+ generative models and high-performance serverless GPUs.
- Self Gradient Forcing Plus — It enables continuous video generation for up to 24 hours from 5-second training windows.
- FlashVSR Plus — Enables spatially untiled 4K video super-resolution on a single GPU.
- Bernini — Combines a semantic planner and DiT renderer for video generation and editing tasks.
- stable-diffusion.cpp — Runs diffusion models efficiently in pure C and C++ with no external dependencies.
- Kandinsky 6 — Generates synchronized 5-second video clips and 44 kHz audio using diffusion models.
- Diffusers — A modular toolbox for inference and training of diffusion models.
- FreeVideo — Runs MiniMax H3 locally on consumer GPUs with as little as 8 GB of VRAM.
- H3.c — Runs native MiniMax-H3 video and audio generation locally on Apple Silicon.
- Lottie Skill — It allows AI agents to edit, animate, and manage complex video projects via code.
- Wan — Provides native 30-second video generation and pixel-perfect consistency.
All TurboDiffusion alternatives