Introduction

pi-H3 is a new few-step version of the open MiniMax H3 video model from Hansheng Chen, Shengqu Cai, Gordon Wetzstein and Leonidas Guibas at Stanford, with co-authors Jiaming Song and Xinchen Yan (part of the work was done at Luma AI). It generates video with stereo audio in just 2 or 4 sampling steps, and on the last step it paints pixels directly, so the heavy video VAE decoder drops out of the pipeline entirely.

The announcement went up on November 10. You can try it today in the free pi-H3 demo Space on Hugging Face. Open weights and inference code are scheduled for November 16–22, with ComfyUI support coming through the existing ComfyUI-piFlow extension.

The Speed Numbers

The team measured complete requests (text encoding through decoded video and audio) for a 1344×768, 24 fps clip with stereo audio on a single NVIDIA B200, using FP8 weights:

  • 5-second clip: 3.81 s at 2 steps, 7.3 s at 4 steps, versus 9.48 s for Sol-H3 at 4 steps
  • 10-second clip: 9.16 s at 2 steps, 17.56 s at 4 steps, versus 24.31 s for Sol-H3
  • 15-second clip: 18.58 s at 2 steps, 36.41 s at 4 steps, versus 44.96 s for Sol-H3

That works out to 2.4–2.7× faster than Sol-H3 at 2 steps and 1.2–1.4× faster at 4 steps. The 5- and 10-second clips finish faster than they play. One fairness note: the Sol-H3 figures come from its paper on a GB200, while pi-H3 was timed on a B200.

On a single H100 80GB, the per-stage breakdown shows where the win comes from. For a 5-second 768p clip, pi-H3 at 2 steps totals about 6.2 s, while a 2-step PDMD model plus the standard H3 VAE takes about 9.9 s, and almost 4 of those seconds go to the decoder alone.

How It Works

pi-H3 stacks two techniques:

  • pi-Flow v2: an upgrade to the team's pi-Flow distillation method (ICLR 2026). It trains few-step models without extra data, keeps rich texture, and holds structure together much better at 2 steps.
  • DWD (Decoding Without a Decoder): instead of handing final latents to a separate VAE, the denoiser itself predicts clean pixels at the last step. That final step costs slightly more than a normal one, but it replaces the entire decode stage.

The authors are upfront that DWD on H3 is still experimental. H3 packs an unusually large patch of pixels into each token, and they say they haven't had enough compute to fully tune for it yet.

Quality Claims

The project page runs side-by-side clips using the same prompts and the same starting noise, with no seed cherry-picking. At 4 steps, the team says pi-H3 beats LightX2V Turbo, FastH3 and larryvrh's H3-Turbo v4 with more dynamic motion, sharper texture, and fuller audio with crisp speech. At 2 steps, they compare it with the 3-step TaoMate-H3 and 2-step PDMD, and suggest that 2 steps is already enough for simpler scenes.

The gallery covers text to video at 2 and 4 steps, image to video, and first-and-last-frame to video, all at up to 768p with stereo audio.

How to Try It Today

  1. Open the pi-H3 demo Space and run a short text-to-video prompt at 2 steps, then rerun the same prompt at 4 steps to compare motion.
  2. Watch the LakonLab GitHub repo, where the weights and inference code are due November 16–22.
  3. Star ComfyUI-piFlow for the ComfyUI nodes once the weights land.

A starter prompt for the Space:

A street violinist plays under a neon awning in light rain at night, the camera slowly pushing in as droplets fall from the brim of her hat, warm shop light on her face, the violin melody echoing off wet pavement with soft traffic in the background.

What We Don't Know Yet

The license for the pi-H3 weights hasn't been posted yet, and the model repo isn't public. Training code is promised "later in November." Until the weights land, the B200 and H100 numbers are the team's own measurements, so treat them as a strong preview rather than a guarantee for your GPU.

Conclusion

The H3 speed race has mostly been about shaving denoising steps. pi-H3 also goes after the decoder, the step that quietly eats seconds at the end of every render. If the open weights match the demo, a 2-step H3 that renders faster than it plays could make local preview loops feel almost live.

Want to play with the full-quality base model while you wait? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=pi-h3-minimax-h3-video-2-4-steps-no-vae-decoder

Original Source

—Aurelia ♡