AIAny
AI Video2026
Icon for item

pi-H3: Video Generation in 2–4 Steps Without a VAE Decoder

Distills MiniMax H3 video generation into 2 or 4 steps, with an option to predict pixels directly and skip VAE decoding. Supports synchronized audio and first- or last-frame conditioning.

Introduction

For few-step video generation, decoding can become a major part of the latency: a fast denoising process still has to pass its output through a separate video decoder. This project tackles both sides of that bottleneck—fewer denoising steps and direct pixel prediction—so the speedup reaches the complete generation pipeline, not just sampling.

Key Capabilities
  • Two- and four-step adapters: pi-Flow v2 distills MiniMax H3 for short trajectories, targeting faster generation while retaining motion and structural coherence.
  • Decoder-free output: DWD reuses the denoiser to predict clean pixels at the final step instead of invoking a separate VAE decoder. A lightweight transformer refiner helps recover detail.
  • Video with audio and image conditioning: The adapters support synchronized audio and can condition generation on a first frame, a last frame, or both.
  • Measured end-to-end speed: On one NVIDIA B200, the reported pipeline generates a 5-second 768p video with stereo audio in 3.8 seconds using two steps. This is a hardware-specific result, not a general performance guarantee.
Who It Suits—and the Trade-off

A good fit for researchers and creators experimenting with fast MiniMax H3 workflows, especially when they can use the compatible model stack and high-end accelerators. Look elsewhere if you need broad hardware compatibility, a general-purpose video model, or licensing terms beyond the MiniMax H3 Community License. The direct-pixel path is still described as experimental, so output quality and speed should be checked on the target workflow.

Information

  • Websitehuggingface.co
  • OrganizationsStanford University, Luma AI
  • AuthorsHansheng Chen, Shengqu Cai, Jiaming Song, Gordon Wetzstein, Xinchen Yan, Leonidas Guibas
  • Published date2026/11/10

More Items

Hugging Face
AI Model2026

Generates video with audio from prompts using a MiniMax H3 fine-tune, with an uncensored tuning approach. Optional quantized checkpoints reduce memory demands; the weights are large and require local inference.

Hugging Face
AI Model2026

Uses a LoRA edit to reduce Inkling 2 Mini’s refusal behavior and provides GGUF builds for local use. It accepts text and images; evaluations show refusals drop, while MMLU falls by 1.5 points.

Hugging Face
AI Model2026

Runs text generation, summarization, and question answering directly on Windows devices as a 2.3B-parameter instruction-tuned model. It supports eight languages and a 32K context window, trading broad world knowledge and deep reasoning for local execution and privacy.