lllyasviel /FramePack

About FramePack

FramePack brings a still image to life as a longer video on your own computer. It works on Nvidia RTX 30XX, 40XX or 50XX cards, even laptops with 6GB of memory. Windows setup pulls over 30GB of models, and the speed-up options change your results.

Best for: Nvidia RTX 30XX, 40XX or 50XX owners who want to animate a still image into a longer clip at home.

• Packs input context to a fixed length, so longer videos don't cost more compute
• Runs a 13B (13 billion parameter) video model on GPUs with as little as 6GB of VRAM
• Shows frames and rough previews as each section finishes
• Makes clips like 5, 6 or 60 seconds at 30fps
• Offers optional TeaCache and attention kernels (xformers, flash-attn, sage-attention)
• Ships a Windows one-click package and a Linux pip install

1. 1
   Windows: download the one-click package, uncompress it, run update.bat, then run.bat.
2. 2
   Linux: use a separate Python 3.10 and install PyTorch with the command below.

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126

   Copy
3. 3
   Install the rest with the command below.

   pip install -r requirements.txt

   Copy
4. 4
   Start the GUI with python demo_gradio.py (it also takes --share, --port, --server).
5. 5
   Skip attention kernels at first; if you want one, try the command below.

   pip install sageattention==1.0.6

   Copy
6. 6
   Run the README sanity checks with the sample image and prompt before your own inputs.