thu-ml/TurboDiffusion

TurboDiffusion: 100–200× Acceleration for Video Diffusion Models

What it solves

TurboDiffusion addresses the high computational cost and slow generation speeds of video diffusion models. It significantly reduces end-to-end generation latency—by 100 to 200 times on a single RTX 5090—while attempting to maintain the original video quality.

How it works

The framework achieves acceleration through a combination of three primary technologies:

  • Attention Acceleration: It utilizes SageAttention and Sparse-Linear Attention (SLA) to speed up the attention mechanism.
  • Timestep Distillation: It employs rCM for timestep distillation to reduce the number of sampling steps required.
  • Quantization: It provides quantized checkpoints for linear layers to enable high-performance generation on consumer GPUs like the RTX 4090 and 5090.

Who it’s for

It is designed for developers and researchers working with video generation models (specifically the Wan series) who need near-instantaneous inference speeds on a single GPU.

Highlights

  • Massive Speedup: Reduces generation time from 184 seconds to 1.9 seconds for certain models on an RTX 5090.
  • Versatile Support: Supports both Text-to-Video (T2V) and Image-to-Video (I2V) workflows.
  • Flexible Resolutions: Capable of generating videos at 480p and 720p resolutions.
  • Multi-GPU Compatibility: Offers both quantized and unquantized checkpoints to suit different hardware (e.g., RTX series vs. H100).

Related

  • Project
    hao-ai-lab/FastVideo

    A unified post-training and real-time inference framework designed to accelerate video generation through sparse distillation and attention optimizations.

  • Project
    dc-ai-projects/DC-Gen

    DC‑Gen is a post‑training framework that compresses the latent space of any pre‑trained diffusion model, aligns embeddings, and fine‑tunes a tiny LoRA module to achieve up to 54× faster inference while keeping image/video quality. It supports text‑to‑image, text‑to‑video, image‑to‑video, and editing, with guides for models like FLUX, Wan2.1, Qwen‑Image‑Edit, and Z‑Image.

  • Project
    NVlabs/FastGen

    A PyTorch-based framework for accelerating diffusion models using various distillation and acceleration techniques for image and video generation.

  • Dispatch
    Optimizing Stable Diffusion XL (SDXL) for Inference Speed and Memory

    Hugging Face explores several optimization techniques for Stable Diffusion XL (SDXL), demonstrating how to reduce memory usage from 28GB to as low as 11.47GB and decrease inference latency from 72.2 seconds to approximately 10.3 seconds.

  • Project
    ModelTC/LightX2V

    LightX2V is an open‑source inference framework for fast image and video generation (T2V, I2V, T2I, I2I, T2AV, etc.). It supports many state‑of‑the‑art models (MiniMax‑H3, Wan, HunyuanVideo, Qwen‑Image, etc.) and provides speed‑up techniques such as 4‑step distilled LoRAs, FP8/NVFP4 quantization, tensor/sequence parallelism, and disaggregated deployment on a wide range of hardware. Benchmarks show up to 3.9× faster inference than competing frameworks on H100 GPUs. The project offers Docker images, pip install, Gradio/ComfyUI front‑ends, extensive docs, and an online demo.