A 'motion-first' video model: flow fields conditioned on 64px frames instead of images

pixlpa · x · 2026-09-23

pixlpa demos an unconventional video generation experiment: a flow model built motion-first rather than image-first, conditioned on a 64px image of the current frame for content awareness plus the previous 4 frames of flow for continuity — yielding a 'demented' but fascinating motion-first video model.

Related event: Dev Trains Motion-First Optical Flow Diffusion Model for Video Generation(2 posts)→

Original post →

More from Multimodal

Multimodal channel →