pi-H3 generates video in 2 to 4 steps with no VAE decoder

pi-H3 generates video in 2 or 4 sampling steps and skips the video VAE decoder entirely, a combination the authors say makes it the fastest version of MiniMax's open H3 video model yet. The project is a research release from Stanford, announced in a blog post on November 10, 2026, and the two things that set it apart from the other fast H3 variants are the step count and the missing decoder.

The authors are Hansheng Chen, Shengqu Cai, Jiaming Song, Gordon Wetzstein, Xinchen Yan and Leonidas Guibas, all listed at Stanford University. The page notes that Chen did part of the work during an internship at Luma AI, and that Song and Yan did their work while at Luma AI. Only one piece of it is usable today: a demo. The weights are promised, not shipped.

What is pi-H3?

pi-H3 is a few-step variant of MiniMax H3, the open multimodal video model MiniMax launched on July 31, 2026 that generates video with native stereo sound, up to 15 seconds at 2K resolution. pi-H3 does not replace H3. It is a distillation that trains the model to reach a finished clip in 2 or 4 passes, where its own comparison set runs other H3 variants at 4 to 8 steps.

Two techniques do the work, the authors say. pi-Flow v2 is a data-free few-step distillation the page describes as a full upgrade of pi-Flow, the method Chen co-authored for ICLR 2026. DWD, short for Decoding Without a Decoder, takes the VAE decoder out of the pipeline.

The framing matters if you have been watching this model family. SLOP TV has covered other fast H3 releases, including Pika H3 Flex, a Pika post-training of H3 down to six steps, and the FastH3 Omni and Taobao PTD releases that cut H3 to 4 or 8 steps. pi-H3 is a separate project with different authors, a different university team, and a different trick.

How does DWD skip the VAE decoder?

DWD reuses the denoiser to predict clean pixels at the last step, so the final latents never pass through a separate VAE decoder network. A diffusion video model normally finishes by handing its latents to a decoder, and that decoder is a real cost in the pipeline. pi-H3 folds pixel prediction into the last denoising step instead, which the authors say is where a chunk of the speedup comes from.

The page is upfront that DWD on MiniMax H3 is still experimental. H3 packs pixels at an unusually large compression ratio, 4×32×32 3-channel pixels per token, and the authors write "we have not yet had enough compute to fully optimize for it, leaving room for further improvement."

What speed does pi-H3 claim?

On a single NVIDIA B200, the authors measured a 5-second 768p video with stereo audio at 3.8 seconds end to end on the 2-step setting. They call that "sub real-time" in the announcement and say both the 5-second and 10-second cases run faster than the video plays. The 10-second figure is 9.2 seconds, and the 15-second figure in the opening paragraph is 18.6 seconds.

The end-to-end chart covers one 1344×768, 24 fps video with stereo audio, measured as the median of seven requests, and it includes text encoding, all denoising steps, pixel output, RGB conversion and audio decoding. Model loading, compilation, warm-up and MP4 encoding are excluded.

Against Sol-H3, another accelerated H3 variant, the page reports pi-H3 at 2.4 to 2.7 times faster across 5 to 15 second videos with 2 steps, and 1.2 to 1.4 times faster with 4 steps. Two details in that comparison sit on the page itself. Sol-H3's numbers are labeled "official, from its paper," and they were measured on one NVIDIA GB200 GPU while pi-H3's were measured on one B200, so the two were not run on the same hardware.

The page's own chart gives the raw seconds, and they divide out the same way. A 5.2-second clip takes pi-H3 3.81 seconds with 2 steps against Sol-H3's 9.48, a factor of 2.5. A 15.1-second clip takes pi-H3 18.58 seconds against 44.96, a factor of 2.4. The speedup the authors attribute to 4 steps is smaller, because four denoising passes cost more than two.

What hardware does pi-H3 need?

Every figure on the page comes from a data-center GPU. The end-to-end numbers were measured on a single NVIDIA B200, and the per-stage latency breakdown on a single NVIDIA H100 80GB at batch size 1. The page lists the speedup as coming from fewer denoising steps, decoder removal and FP8 quantization, an 8-bit weight-and-activation format.

A second chart breaks the pipeline into a text encoder, the denoiser, a video decoder and an audio decoder, and it shows no video decoder bar for pi-H3. The page's caption explains why: "pi-H3's last denoising step costs slightly more than a regular step because it also produces pixels, replacing the video decoder." The same caption notes that every method in the chart shares the same text encoder, audio decoder and FP8 MiniMax H3 denoiser, so the differences come from the number of denoising steps and the decoder.

How does pi-H3 compare with other few-step H3 models?

At 4 steps, the page sets pi-H3 against LightX2V Turbo at 4 and 8 steps, FastH3 Dense/DataFree at 4 steps, and larryvrh's H3-Turbo v4 at 6 steps. At 2 steps, it sets pi-H3 against the 3-step TaoMate-H3 and 2-step PDMD. Every model in each comparison gets the same prompts and the same Gaussian noise, the page says, and there is no seed cherry-picking. The authors describe pi-H3's output as more dynamic movement, sharper frames and fuller audio, and they say 2-step generation is already enough for simpler scenes.

These are the authors' own side-by-side clips, and the page publishes no score from an outside benchmark or arena. No independent test of pi-H3 was available when this was written, so every number here is the team's own, measured on its own hardware.

When can you get pi-H3?

As of November 11, 2026, the demo is live and nothing else is. A Hugging Face Space runs the model, published under the account Lakonik, which is Chen's. The page labels its model and code links "coming soon," and the linked Hugging Face model page for the weights, MiniMax-H3-piFlowV2-DWD, was not publicly available when checked on November 11, 2026.

The schedule the authors give runs in three parts: the demo now, model weights and inference code in LakonLab during November 16 to 22, and training code "later in November, once we finish cleaning up the codebase." ComfyUI support is planned through the ComfyUI-piFlow extension.

LakonLab is Chen's repository, the official codebase for AsymFlow, pi-Flow and GMFlow under the Apache 2.0 license. ComfyUI-piFlow is the corresponding ComfyUI extension, also Apache 2.0, and its changelog already records fixes to MiniMax H3 workflows. Neither repository listed a pi-H3 checkpoint as of November 11, and the announcement does not state a license for the weights or a price.

If the schedule holds, pi-H3 would be the latest in a run of few-step variants built on the same open base model, each one shaving the render time. Whether a 2-step clip holds up on harder scenes than the demo gallery is the open question, and the page's own comparison clips are the only evidence so far.

Start with the demo at huggingface.co/spaces/Lakonik/pi-H3; the weights and inference code are promised for the week of November 16 in the LakonLab repository, with training code later in the month.