Pyramid Flow

Pyramid Flow generates video autoregressively using flow matching over a pyramid of resolutions: instead of running every denoising step at full resolution, it interpolates between latents at different scales and noise levels, which cuts compute substantially. A single Diffusion Transformer, trained in about 20,700 A100 GPU hours exclusively on open datasets including WebVid-10M, OpenVid-1M, and Open-Sora Plan, produces 10 second clips at 768p and 24 fps or 5 second clips at 384p, supports image-to-video with text conditioning, and includes a miniFLUX variant for image generation up to 1024p. Inference fits in as little as 8 GB of VRAM with sequential CPU offloading, and on the VBench benchmark the model surpasses the open-source baselines it was compared against while approaching commercial systems such as Kling and Gen-3 Alpha on quality and motion smoothness scores. Code and checkpoints are released under the MIT license, attracting researchers and hobbyists interested in reproducible video generation built on open data.