FramePack is a progressive video generation architecture based on next-frame prediction. It compresses input context into a fixed-length format, ensuring that the computational workload stays consistent regardless of video length. This design enables even 13B models to handle long sequences efficiently—making it feasible to run on laptop GPUs.
However, the default Windows installation lacks Sage Attention support, which can significantly enhance performance. In this guide, I’ll show you how to add Sage Attention, xformers, and Flash Attention support to unlock the full potential of FramePack on your setup.
Installation
- Visit FramePack github page and follow the instructions to do the default installation.
- Download this zip file and extract the include and libs directories. Place them in system\python directory.
- Visit this page and download the package_installer.zip.
- Extract the package_installer.bat from the zip file and place it in the FramePack directory. Run the batch file and select the option to install required packages.
- After the packages are installed, you should see that xformers, Flash Attention, and Sage Attention are installed when you run FramePack.
- Speed increase –
Without Sage Attention: ~ 13.20 s/it
With Sage Attention: ~ 8.73 s/it
Examples
Prompt: The girl waves to the viewer and then touches her hair.
Prompt: The girl turns around and dances gracefully, with clear movements, full of charm.
Prompt: The girl dances gracefully, with clear movements, full of charm.
Prompt: The girl makes a heart gesture.
Conclusion
Enabling Sage Attention delivers a substantial performance boost for FramePack, reducing iteration time from approximately 13.20 seconds to just 8.73 seconds—a 34% speed improvement. If you’re working with larger models or longer videos, adding Sage Attention support is a must for smoother and faster generation, especially on resource-limited systems like laptops.
Further Reading
Leave a Reply