FramePack
Generate 1-minute videos quickly with only 6GB of low VRAM.
Features
System Requirements
Minimum 16GB RAM.
Note: The model requires significant storage space - 60GB+ free disk space recommended.
macOS 15+: M-series chips required.
Windows 10/11: NVIDIA GPU with 6GB+ VRAM required
Note: For NVIDIA GPUs, install a newer driver.Introduction
FramePack is a project for video generation. It is the official implementation and desktop software for the paper "Packing Input Frame Context in Next-Frame Prediction Models for Video Generation".
The core of FramePack is a neural network structure based on next-frame (next-frame-section) prediction, which can generate videos progressively. One of its remarkable features is that it can compress the input context into a fixed length, making the generation workload invariant to the video length. This means that even on a laptop GPU, it can process a large number of frames using a 13-billion-parameter model. Moreover, it can be trained with a batch size similar to that of image diffusion training.
In terms of usage, this project has many advantages. It supports multiple operating systems, such as Windows and Linux. Regarding hardware requirements, it is recommended to use Nvidia RTX 30XX, 40XX, 50XX series GPUs that support fp16 and bf16. A minimum of 6GB of GPU memory is sufficient to generate a 1-minute, 30fps video using a 13B model. The installation process is relatively simple. There is a one-click installation package for Windows systems, and for Linux systems, users just need to install the dependencies according to the steps. The software provides a graphical user interface (GUI). Users can generate videos simply by uploading an image and entering a prompt, and they can also view the generation progress and latent previews.
However, there are also some points to note when using it. For example, due to the inverted sampling, the starting actions of the video may be generated later than the ending actions, so users need to be patient. In addition, functions like TeaCache and sage-attention can improve the speed but may have a certain impact on the results. It is recommended to use them for trying ideas first and then use the full diffusion process to obtain high-quality results. At the same time, it should be noted that the only official website of this project is the above GitHub repository.