Hansheng Chen, Shengqu Cai, Jiaming Song, Gordon Wetzstein, Xinchen Yan and Leonidas Guibas · Decoder-free few-step video

pi-H3

pi-H3 is a few-step distillation of MiniMax H3 that generates video with audio in 2 or 4 sampling steps. It predicts pixels directly at the final denoising step (DWD, Decoding Without a Decoder), so the video VAE decoder is skipped. On a single B200 it produces a 5-second 768p video with stereo audio in 3.8 s with 2 steps, and a 15-second video in 18.6 s. The authors report it as 2.4–2.7× faster than Sol-H3 at 2 steps and 1.2–1.4× faster at 4 steps, using their own measurements against Sol-H3's published figures. It combines pi-Flow v2, a data-free few-step distillation method, with DWD, which the authors call experimental on H3. It is an open-source research project: a Hugging Face demo is live, weights and inference code are due Nov 16–22 in LakonLab, and training code will come later in November.

Open the source ↗

Similar Video & Animation releases

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.