pi-H3 generates video in 2–4 steps with no VAE decoder — 5s clip in 3.8s on one B200

Apprehensive_Sky892 · reddit · 2026-11-11

Open-source research project pi-H3 (built on MiniMax H3) generates video in just 2 or 4 sampling steps, predicting pixels directly at the final step and skipping the video VAE decoder entirely. On a single B200 it produces a 5-second 768p clip with stereo audio in 3.8s — faster than playback — and a 15-second video in 18.6s.

End-to-end speed (single B200, text encoding through decoded audio/video):

5s video in 3.8s (2 steps), 10s in 9.2s — both faster than real-time playback
2.4–2.7× faster than Sol-H3 at 2 steps; 1.2–1.4× at 4 steps across 5–15s videos
Speedup from fewer denoising steps, decoder removal, and FP8 quantization

Quality: the authors claim best overall fidelity among 4–8-step H3 variants (vs LightX2V Turbo, FastH3, H3-Turbo v4) — more dynamic motion, sharper textures, and fuller audio bass/crisper speech, with no seed cherry-picking in comparisons.

Two core techniques:

pi-Flow v2: a full upgrade of pi-Flow for data-free few-step distillation, improving structural coherence at 2 steps and simplifying inference with straighter trajectories
DWD (Decoding Without a Decoder): reuses the denoiser itself to output clean pixels at the last step; still experimental on H3's unusually large 4×32×32 compression ratio

Release schedule: HF demo live now; weights + inference code Nov 16–22; training code later in November, with support via the ComfyUI-piFlow extension.

Original post →

pi-H3 generates video in 2–4 steps with no VAE decoder — 5s clip in 3.8s on one B200

More from Multimodal