OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
Image & Video AI

Dmad

DMAD is a diffusion-distillation research method, model collection and code repository for fast visual generation, including four-step MiniMax-H3 text-to-audio-video models.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Dmad — OpenViST project profile in Image & Video AI.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use DMAD when you want to examine adversarial diffusion distillation or work with the released few-step models, especially the MiniMax-H3 workflow for generating video and audio from text.

Who it is for

  • Researchers studying diffusion distillation and few-step generative models
  • Developers who want to run the documented MiniMax-H3 text-to-audio-video workflow locally
  • ComfyUI users seeking the repository’s MiniMax-H3 workflows

Documented uses

  • Generate video with audio from a text prompt using a four-step MiniMax-H3 student model.
  • Explore DMAD weights and training code for ImageNet-64, SDXL, Wan2.1 or MiniMax-H3.
  • Use the released students through the repository’s inference script, Diffusers pipeline code or ComfyUI materials.
Before you choose

Requirements & limits

  • MiniMax-H3 inference requires substantial model storage: the repository estimates about 170 GB for the needed base-model components, plus checkpoints.
  • The documented setup specifies Python 3.12, PyTorch 2.11.0 and CUDA 12.6.
  • The repository offers different use paths: command-line inference, a Diffusers pipeline, and ComfyUI workflows.
  • The README identifies the four-step setting as the trained configuration and describes default output as 1,344 × 768 video with stereo audio.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Set up the documented MiniMax-H3 inference environment

Prepare the Python environment and download the base model components and DMAD checkpoints required by the repository’s inference workflow.

Before you start

  • Python 3.12 and Conda
  • PyTorch 2.11.0, torchvision 0.26.0 and torchaudio 2.11.0 for CUDA 12.6
  • Accept the MiniMax-H3 Community License on Hugging Face
  • About 170 GB for the required MiniMax-H3 components

Steps

  1. Create and activate the dmad_h3 Conda environment with Python 3.12.

  2. Install the stated PyTorch packages for CUDA 12.6 and install the repository requirements.

  3. Download MiniMax-H3 into models/MiniMax-H3, excluding the listed FL2VA, Ref2VA and transformer_ref partitions.

  4. Download the MiniMax-H3 DMAD critic checkpoint files into ckpt.

Expected outcome

The documented model directory and checkpoint files are prepared for the repository’s inference script.

Guide limitations

  • This guide covers setup only; it does not verify that the environment runs on a particular machine.
  • The base-model download is about 170 GB, in addition to the DMAD checkpoint.
Documentation 1 ↗
Source-based practical guide

Generate a text-prompted audio-video clip with the inference script

Use a prepared MiniMax-H3 model and DMAD checkpoint to request a text-to-audio-video clip through the documented command-line script.

Before you start

  • The MiniMax-H3 components and DMAD checkpoint described in the repository setup
  • The repository’s Python environment and inference script

Steps

  1. Choose a DMAD checkpoint, such as dmad_minimax_h3_4step_lora_critic.safetensors.

  2. Run inference.py with --model-dir, --lora, a prompt supplied by --prompt or --prompt-file, and --output-dir.

  3. Optionally set --seed; the documented default is 42.

Expected outcome

The documented workflow writes generated MP4 output and its text into the chosen output directory, with a settings file recording arguments.

Guide limitations

  • The four-step setting is the trained configuration; the README says other step counts work but does not make them the trained setting.
  • The default output is 1,344 × 768, 124 frames and 32 kHz stereo audio; frame counts must follow the documented 17 * n + 5 constraint.
  • The repository describes an 80 GB GPU path and a low-VRAM option for consumer GPUs; actual resource needs depend on the selected configuration.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
14 October 2026
Original website title
Yzmblog/DMAD · Official homepage title · retained in its source language
How we review sourcesSuggest a correction