pi-H3 generates 768p video in 2-4 steps: 10s clip in 9.2s on a single B200
GordonWetzstein · x · 2026-11-11
Stanford's Gordon Wetzstein group, with Luma AI, announced pi-H3, an open-source project making the fastest MiniMax H3 yet.
Key technical points:
- Data-free distillation compresses sampling to 2 or 4 steps
- Skips the video VAE decoder entirely, predicting pixels directly at the final step
- With FP8 quantization, end-to-end generation on a single B200 takes 3.8s for a 5-second 768p video with stereo audio, and 9.2s for 10 seconds — both faster than playback
- 2.4-2.7× faster than Sol-H3 across 5-15s videos (2 steps), 1.2-1.4× with 4 steps
The Hugging Face demo is live; model weights and inference code land the week of Nov 16, training code later in November via LakonLab, plus ComfyUI support through ComfyUI-piFlow.
More from Multimodal
- Hailuo AI announces SOTA-quality model, strong across styles, open for everyone — Hailuo_AI · 2026-11-11
- Seedream 5.5 nails microvascular surgery imagery with sub-millimeter precision — aftahi_ai · 2026-11-11
- MiniMax H3-Image aces a text-heavy sign test: 4/4 generations, zero typos — Ok_annbae · 2026-11-11
- Fable 5.2 Max Turns Istanbul Into an Explorable 3D World for ~$46 — minchoi · 2026-11-11
- Claude Fable 5.2: 10 Wild One-Prompt Demos in Under 24 Hours — minchoi · 2026-11-11
- Fable 5.2 Shows Much More Musical Diversity in Video Spanning Four Decades of C-Pop — dotey · 2026-11-11