Dev trains video diffusion model from first principles with motion-first approach
pixlpa · x · 2026-09-11
Developer pixlpa shares hands-on progress training a video generation model from first principles using a motion-first approach: the diffusion model now generates frame motion data good enough to mimic source clips. Next steps include running it in realtime, exposing control knobs, and applying it to more interesting feedback processes.
Related event: Dev builds optical-flow diffusion model overnight to learn video motion(4 posts)→
More from Multimodal
- MiniMax H3 trick: 1:2 first-frame image generates frozen-in-time video with subtle motion — Hailuo_AI · 2026-09-11
- Creator chains Codex+Remotion motion graphics into MiniMax H3 video generation — Hailuo_AI · 2026-09-11
- MiniMax Design H3 + Blender workflow cracks seamless outdoor-to-indoor video scenes — Hailuo_AI · 2026-09-11
- Watercolor-style UE5 footage via Depth conversion + local MiniMax H3, four scenes built in one day with Astra — Hailuo_AI · 2026-09-11
- Creator builds Tallinn old town in Astra + Blender, iterates Itano Circus missile shots into anime with H3 — Hailuo_AI · 2026-09-11
- Flux Klein 2 Users Struggle to Change Clothes Fit Without Altering the Fabric — diond09 · 2026-09-11