FramePack

FramePack is an open-source next-frame-prediction neural network for video generation that compresses input contexts to a constant length, making the generation workload invariant to video length and enabling long video synthesis even on laptop GPUs.

What is FramePack?

FramePack is a next-frame (next-frame-section) prediction neural network structure for progressive video generation. It takes an input image and a text prompt and generates video frames one by one (or section by section). The model compresses input contexts to a constant length, so the computational workload does not increase with video length. FramePack is developed by Lvmin Zhang and colleagues, and the official implementation is hosted on GitHub at lllyasviel/FramePack. It runs on Linux and Windows with Nvidia GPUs (RTX 30XX/40XX/50XX, 6GB VRAM minimum) and supports 13B-parameter models.

Key Features

• Constant workload regardless of video length — The generation workload stays the same for 5 seconds or 60 seconds of video, unlike traditional autoregressive or diffusion models that scale linearly.
• Runs on laptop GPUs — Minimal requirement is 6GB VRAM. A 13B model can generate 1800 frames (1 minute at 30fps) on a laptop GPU.
• Progressive generation — Because it predicts the next frame or next section, users see the video building frame-by-frame in real time, getting immediate visual feedback.
• User-friendly GUI — The included demo_gradio.py provides a web interface for uploading images, writing prompts, and viewing generated videos with progress bars.
• Multiple attention backends — Supports PyTorch native attention, xformers, flash-attn, and sage-attention (e.g., pip install sageattention==1.0.6).
• TeaCache acceleration — Optional caching that speeds up generation by ~1.7× (2.5→1.5 sec/frame on RTX 4090) but may slightly affect quality.
• Available as a one-click Windows package — Download from GitHub Releases, uncompress, run update.bat then run.bat.
• Paper and project page — Published at NeurIPS 2025 (arXiv:2504.12626) and project page.

Who is it for?

• AI video researchers — To experiment with next-frame-prediction architectures and drift prevention techniques using a state-of-the-art open-source implementation.
• Content creators — To generate short animated videos from a single image and a motion prompt, e.g., animating a static portrait into a dancing character.
• GPU-constrained users — Those with only a laptop or low-VRAM GPU (6GB+) who want to run large video diffusion models locally.
• Open-source enthusiasts — Developers who want to integrate video generation into their own projects or extend the codebase (Apache-2.0 license).

What can you do with FramePack?

• Image-to-video generation — Upload any image (e.g., a person) and write a motion prompt like “The girl dances gracefully, with clear movements, full of charm.” FramePack will generate a video of configurable length (up to 60+ seconds).
• Long-video stress testing — Generate 1-minute videos (1800 frames) on a laptop GPU to evaluate anti-drifting capabilities. The project page shows results of pure text-to-video stress tests.
• Batch experimentation — Use the sanity-check examples provided in the README to verify hardware and software setup, comparing results with and without TeaCache or quantization.
• Prompt engineering — Use the ChatGPT template included in the README to generate concise, motion-focused prompts optimized for FramePack.

For installation, see the README. The software is licensed under Apache-2.0.