hao-ai-lab /FastVideo

About FastVideo

FastVideo lets you run and retrain open video models yourself instead of paying a hosting bill. You get finetuning, LoRA training, and tricks that cut the number of denoising passes by more than half. Setup takes some work, since it assumes you have a decent GPU.

Best for: Hands-on makers with a solid GPU who would rather tune a model than rent someone else's.

• Finetunes open video models fully, or trains a smaller LoRA version
• Skips ahead during generation with DMD2 and sparse distillation
• Uses Video Sparse Attention and Sliding Tile Attention kernels
• Shares the load across GPUs with sequence parallelism
• Generates and edits live video in the browser via the Dreamverse app

1. 1
   Start a fresh environment: uv venv --python 3.12 --seed, then source .venv/bin/activate
2. 2
   If you have an NVIDIA card, install with the command below (swap in UV_TORCH_BACKEND=cu130 for newer drivers)

   UV_TORCH_BACKEND=cu126 uv pip install fastvideo

   Copy
3. 3
   Clone the repo and run uv pip install -e '.[mlx]' on an Apple Silicon Mac
4. 4
   Save a script named example.py that loads a generator and writes a clip
5. 5
   Generate your first video with python example.py