FramePack

The README states that the GitHub repository is the only official FramePack website and that the project runs no web services. It names framepack.ai, framepack.net, framepack.co, framepack.pro, framepack.cc and similar domains as spam and fake, and it tells readers not to pay money or download files from them. This page links only to the repository, its release and the paper.

On Windows, the README links a one-click 7z package (CUDA 12.6 and PyTorch 2.6) from the repository release, about 1.7 GB. Extract it, run update.bat, then run.bat. On Linux, install PyTorch for CUDA 12.6 and the requirements file in an independent Python 3.10 environment, then start the interface with demo_gradio.py.

The documented requirement is an NVIDIA RTX 30, 40 or 50 series GPU that supports fp16 and bf16, on Linux or Windows, with at least 6 GB of GPU memory. The README says 6 GB is enough for a one-minute, 30 fps video with the 13B model and that laptop GPUs are acceptable. GTX 10 and 20 series cards are not tested.

The app switches to a high-memory mode only when more than 60 GB of GPU memory is free; otherwise it offloads models between system memory and the GPU. If you hit out-of-memory errors, raise the preserved-memory setting, which starts at 6 GB, and expect slower generation. The README does not list macOS or AMD GPUs.

• GPU: NVIDIA RTX 30, 40 or 50 series with fp16 and bf16 support.
• GPU memory: at least 6 GB.
• Operating system: Windows or Linux.
• Disk: about 1.7 GB package plus more than 30 GB of models.

The author reports 2.5 seconds per frame unoptimized and 1.5 seconds per frame with TeaCache on an RTX 4090 desktop. Laptop GPUs such as a 3070 Ti or 3060 laptop run about four to eight times slower. The README links a troubleshooting thread for speeds far below these figures.

Because FramePack predicts the next section of frames, you see finished sections while the rest renders. The first minutes can be slower while the device warms up. A result that looks like a one-second clip early in the run is expected; wait for the remaining sections before judging length or motion.

There is no subscription or credit system, so the costs are local: a supported NVIDIA GPU, storage for about 1.7 GB of package files and more than 30 GB of models, and the electricity for long renders. The first launch downloads the models from Hugging Face automatically, so allow for bandwidth and time before the interface is usable.

The Apache-2.0 license covers the repository code. The weights come from separate Hugging Face repositories, including HunyuanVideo components and FramePack transformer weights, and their metadata does not state a license. If output is headed for commercial work, read the terms of the original model releases first. The –share option exposes the local interface through a public link, so use it deliberately.

The repository was created in April 2025. FramePack-F1 followed in May 2025, and June and July 2025 notes describe FramePack-P1 as the next version, with result pages on the project site but no separate download in the README. The latest commit, in October 2025, revised the paper title and citation.

With no code changes for roughly a year and hundreds of open issues, plan on community troubleshooting rather than provider support. Run update.bat after downloading the Windows package, because the README warns that skipping it can leave known bugs unfixed. The paper and project page document the method itself.

1. Record the GPU model, GPU memory, driver and whether update.bat was run.
2. Generate the 5-second clip and note the seconds per frame shown during sampling.
3. Compare the result with the reference video in the README; small differences between devices are expected.
4. Repeat with TeaCache on and note speed and visible changes.
5. Keep both MP4 files and the settings used.

Accept a clip that looks broadly similar to the reference at a speed consistent with your GPU class. Reject crashes, out-of-memory errors at default settings or output unrelated to the prompt, and check the troubleshooting thread before changing settings. No speed or quality result is recorded here.

Without a supported NVIDIA GPU, hosted generators such as Runway or Dzine avoid local hardware but charge credits and use different models.