On the other site, we published an article testing FramePack, a locally run video generation AI.

Running FramePack locally

We generated a video from a still image using AI, albeit for a short period of time (4 seconds). The actual generated video is at the link.

I built FramePack on Windows11, WSL2, and Ubuntu, so that I can quickly migrate to a CUDA-optimized Linux bare-metal environment if I remember the steps to build on WSL2. In this sense, we built the environment on WSL2, not on Windows11.

Time required to generate a 4-second video

As noted in the link, it took approximately 22 minutes to generate a 640 x 608 pixel, 30.00 fps, 4 second video on an nVidia RTX3060 12GB environment.

If a 15-second video were to be generated, we estimate that it would take about 1.5 hours to generate a 15-second video, and about 3 hours for a 30-second video. It is not always possible to generate a satisfactory video in a single run, so if you want to generate a full-scale video, you will need more GPU power to do so.

Required Computer Resources

It states that a minimum of GPU 6GB memory is required.

In my environment, the main memory is 64 GB and the GPU 12 GB, but both were almost fully used up during video generation.

If you have less main memory or GPU memory than this, the generation time will be much slower because of the swap to SSD.

Musubi

I introduced an article on another site that installed and used FramePack, a video generation AI. I have already written an article about FramePack on my main blog. To avoid duplication, I will only introduce them on this site.