FramePack

Minute-long AI video on a 6 GB laptop GPU is FramePack's claim to fame, a next-frame prediction architecture from lllyasviel, the developer behind ControlNet. It compresses the context of previously generated frames to a constant length, so memory use stays flat no matter how long the video runs, letting a 13B-parameter model generate 1,800 frames, 60 seconds at 30 fps, on 6 GB of VRAM. On an RTX 4090 generation runs about 2.5 seconds per frame, or 1.5 seconds with TeaCache, and supported hardware is NVIDIA RTX 30, 40, and 50 series cards with fp16 or bf16, including laptop GPUs. A one-click Windows package bundles CUDA 12.6 and PyTorch 2.6, Linux installs via pip, and a Gradio interface shows frames as they stream out, making video diffusion feel closer to image diffusion in practice. Code is Apache-2.0, the repository has around 17,200 stars, and F1 and P1 follow-up variants extended the approach through 2025. Hobbyists on consumer hardware are the core audience, since little else generates videos this long at this VRAM budget.