Dev Trains Motion-First Optical Flow Diffusion Model for Video Generation

Developer pixlpa trained an unconventional video pipeline that is motion-first: a diffusion model trained on optical flow maps warps generated frames to ensure continuity, with a small latent diffusion model conditioned on the current frame for image content.

2026-09-23 ~ 2026-09-23 · 2 related posts

• Self-trained optical flow diffusion model warps LDM frames for coherent video — pixlpa · 2026-09-23
• A 'motion-first' video model: flow fields conditioned on 64px frames instead of images — pixlpa · 2026-09-23