FramePack

Why we included this project

FramePack is a next-frame video diffusion model built to make long-form generation practical on ordinary consumer hardware. It compresses the input frame context to a fixed length, so the compute cost stays flat whether you ask for a few seconds or a full minute of 30fps footage. The authors show a 13B model producing that footage on as little as 6GB of GPU memory, which opens up animation from a single reference image plus a text prompt to indie creators and small teams who would rather not rent a cluster. The repository is a working desktop app with a Gradio interface, a one-click Windows package, and honest notes on speed and quality trade-offs such as TeaCache and attention kernels. Since it is the official reference implementation of the published paper, it is also a reasonable place to start if you want to study the technique or build on it.