Open Sora Plan

They acknowledge they currently don't have enough hardware access to achieve what they want. They are currently training on 8x A100-80G, and the input size is 8 frames * 128 * 128. It takes about a week to train enough to generate ucf-like videos.

I think there's a lot of room in FLOSS AI for "@home" style distributed training incorporating some features of federated learning. The limiting factor at the moment is that large models have a minimum VRAM requirement which is often higher than the 16-24GB that home users would be expected to have.