ModelTC/Minimax-H3-Turbo

Distill Minimax-H3 into 4 steps

What it solves

It provides a set of distilled LoRA checkpoints for the MiniMax-H3 model to enable high-quality video and audio generation with significantly fewer inference steps (NFE), reducing the time and computational cost required to produce content.

How it works

The project uses distillation to create "Turbo" versions of the model. These checkpoints allow for fast inference (typically 4 to 8 steps) while supporting multiple tasks including text-to-video-and-audio (T2VA), image-to-video-and-audio (FL2VA), and reference-to-video-and-audio (Ref2VA). It includes specific configurations for different resolutions (544p and 768p) and provides specialized reference-image resizing policies (match, max, and diffusers) to ensure the reference images align with the training resolution.

Who it’s for

  • AI Artists and Creators: Those using ComfyUI or Diffusers to generate videos with audio.
  • Developers: Those looking to integrate fast video generation into applications via the provided API.
  • Researchers: Users interested in distilled model checkpoints for efficient inference.

Highlights

  • Fast Inference: Supports 4-step and 8-step distilled models.
  • Multi-Task Support: Handles T2VA, FL2VA, and Ref2VA workflows.
  • Flexible Integration: Compatible with both Diffusers and ComfyUI.
  • Resolution Options: Checkpoints available for both 544p and 768p outputs.
  • Reference Image Control: Includes three distinct resizing policies for reference images to optimize visual consistency.

Related

  • Dispatch
    SDXL and Stable Diffusion Fast Inference with Latent Consistency LoRAs

    Hugging Face introduces LCM LoRAs, a method to enable high-quality image generation in 4 to 8 steps for SDXL and Stable Diffusion models, significantly reducing inference time.

  • Project
    hao-ai-lab/FastVideo

    FastVideo is an open‑source framework for training, fine‑tuning, and real‑time inference of video diffusion models. It offers sparse‑attention distillation, scalable multi‑GPU training, and cross‑platform support (NVIDIA GPUs, Apple Silicon via MLX, ARM DGX Spark). The library includes a CLI, Python API, and the Dreamverse app for live video editing. Install via `uv` and start generating videos with a few lines of Python.

  • Project
    huangserva/ComfyUI_MiniMaxH3_Director

    A collection of ComfyUI workflows for MiniMax H3 Director, enabling text-to-video, image-to-video, and advanced video-to-video editing with reference materials.

  • Project
    OpenVDN/vdn-minimax-h3

    A hybrid-attention model based on MiniMax H3 that significantly speeds up video generation while maintaining high visual quality.

  • Project
    antirez/h3.c

    A native Metal inference engine for MiniMax-H3 on Apple Silicon, enabling local text-to-video and text-to-audio generation with advanced memory and performance optimizations.