FramePack

FramePack is a video generation architecture by Lvmin Zhang (of ControlNet/Fooocus fame) and Maneesh Agrawala at Stanford, published in 2025. Instead of processing every frame at full cost, it packs input frame contexts into fixed-length representations, so a 13B image-to-video model can generate a full minute of 30fps video using only about 6GB of GPU memory — hardware that chokes on competing open video models. Because it predicts the next frame (or section) iteratively, you watch frames appear in real time and get immediate visual feedback. It plugs into the Hugging Face diffusers FramePackPipeline, ComfyUI, and various community Gradio apps.

The honest picture: generation is slow on consumer hardware — roughly 1.5–2.5 seconds per frame on an RTX 4090, several times slower on laptop GPUs — so a minute of video can take the better part of an hour. It is image-to-video only (no native text-to-video), needs an Nvidia GPU with fp16/bf16 support, and quality depends heavily on sampling parameters. But for local, private, long-form video generation on modest hardware, nothing else in the open ecosystem comes close.

Clone the official repo and follow its README (Nvidia GPU, 6GB+ VRAM, Linux or Windows), or use a community Gradio/ComfyUI wrapper for a simpler start.