Pyramid-Flow

Why we included this project

Pyramid Flow generates short videos from text prompts or images using a flow-matching scheme that builds frames up across resolutions in stages. That pyramid structure is what keeps training costs down; the authors trained it on open datasets yet it still produces 10-second 768p clips at 24 FPS. It also comes with real tooling, not just a paper: pretrained checkpoints on Hugging Face, finetuning code for the DiT backbone, VAE training code, and multi-GPU inference with CPU offloading that can run on under 8GB of GPU memory. For teams prototyping text-to-video or image-to-video pipelines, it's a complete reference implementation you can actually pull apart and adapt.