pi-H3 Released: 2-4 Step Video Generation Without a VAE Decoder, 3.8s for 5s HD on a B200
Apprehensive_Sky892 · reddit · 2026-11-22
pi-H3, the fastest version of MiniMax H3, is now open source: it generates video in just 2 or 4 sampling steps and directly predicts pixels at the final step, skipping the video VAE decoder. On a single B200 it produces 5-second HD video in 3.8 seconds (sub real-time) and 15-second clips in 18.6 seconds — over 2x faster than Sol-H3. Weights, inference code, and an HF demo are live, with ComfyUI support via the ComfyUI-piFlow extension; training code to follow.
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