2-step pi-h3 generates video without a VAE decoder — 5s clip in 84s on an RTX 5080
NoCockroach2245 · reddit · 2026-11-22
Lakonik's released pi-h3 generates video in just 2 steps without a VAE decoder. Though trained with first-frame conditioning, it works fine in ref mode too. End-to-end generation of a 5-second, 1280x736 video takes 84.11 seconds on an RTX 5080.
The post links the weights (Lakonik/pi-H3 and SulphurAI/Sulphur-3 on Hugging Face) plus a ready-made workflow. The author notes there's no free lunch — the speed has trade-offs — but it clearly has its use cases.
Related event: pi-H3 open-sources 2-step video generation, 5s HD in 3.8s on B200(3 posts)→
More from Multimodal
- ComfyUI launches model usage rankings: Qwen Image 2.1 tops generations at 16% — cremebait15 · 2026-11-22
- pi-H3, a flow-matching video model built on MiniMax-H3, trends on Hugging Face — Lakonik · 2026-11-22
- GPT 6.2 Astra composes Beethoven-style scherzo, with all 604 key strikes from the score — imjustnewatai · 2026-11-22
- Musk demos real-time AI video generation on half of one AI4 TRIP accelerator — jamesdouma · 2026-11-22
- RAEA engine lets AI agents create videos, 3D scenes and games with React — Promptmethus · 2026-11-22
- Do adaptive optimizers still work? Reddit users doubt it for flow-matching models — Inevitable_Pen9043 · 2026-11-22