DMAD MiniMax-H3 4-Step: Free AI Video With Sound in 4 Steps (Plus the Alternatives US Creators Can Use)

A new distillation method called DMAD (Distribution Matching as Adversarial Distillation) just shipped a pair of free LoRAs that turn MiniMax-H3, the 33B text-to-audio-video model, into a 4-step generator. The DMAD MiniMax-H3 4-step student makes a 1344×768, 5.2-second clip with a synchronized stereo soundtrack in four model evaluations instead of the teacher’s 50. The paper went up on arXiv on October 1, 2026, from researchers at Texas A&M University and ByteDance.

Why a creator should care: fewer steps means cheaper, faster drafts of video with sound, and there’s already a free hosted demo you can try in a browser. Why you should keep reading before you build a pipeline on it: the weights inherit MiniMax-H3’s license, which excludes the US, EU, UK and South Korea. Below is what DMAD actually does, the free demo, the real hardware bill, and the 4-step alternatives you can use if you live in one of those regions.

DMAD MiniMax-H3 4-Step at a Glance

• What it is: two 1.4 GB rank-128 LoRAs that make MiniMax-H3 (33B) generate video plus stereo audio in 4 steps.
• What’s free: a ZeroGPU demo Space, the LoRA weights, and the Apache-2.0 training/inference code.
• What it costs to run locally: roughly 170 GB of base-model downloads, then either one 80 GB GPU or a 24 GB consumer card with the new --low-vram mode. The authors estimate 4–8 minutes of sampling per clip on an RTX 4090/5090.
• The catch: the weights fall under the MiniMax H3 Community License, which excludes the US, EU, UK and South Korea, outputs included.
• Best free fallback if you’re in those regions: SANA-Video 2.0 4-Step (Apache 2.0, free demo) or LTX-2.5’s distilled 8-step model (free under $10M revenue).

What DMAD Actually Changes

Most fast video models come from “distillation”: you train a small-step student to imitate a slow, many-step teacher. Popular methods like DMD2 need an extra diffusion model running alongside training to estimate how far the student’s outputs are from the teacher’s. DMAD swaps that for a classifier. A shared backbone with two discriminator heads learns to tell real data and teacher samples apart from student samples, and the student is trained directly on those logits. The authors describe it as recasting distribution matching as classification.

For you, the practical result is that the student needs only 4 steps with no classifier-free guidance, so each step is a single forward pass. The authors applied the same recipe to SDXL, Wan2.1-T2V-14B and MiniMax-H3, but only the MiniMax-H3 checkpoints are published on the DMAD model page right now. There are no 4-step Wan2.1 or SDXL weights to download yet.

Specs: What the 4-Step Student Outputs

• Resolution: 1344×768 by default (any multiple of 32).
• Length: 124 frames at 24 fps, about 5.2 seconds. Frame counts must follow 17*n + 5.
• Audio: 32 kHz stereo, muxed as H.264 + AAC. Your prompt describes the soundscape too: narration in quotes, music described in words, ambient sound included.
• Checkpoints: dmad_minimax_h3_4step_lora_critic.safetensors (the paper checkpoint) and dmad_minimax_h3_4step_full_critic.safetensors (trained longer, scores higher on AVGen-Bench). Each is 1.4 GB.
• Sampler: a “re-noise” rule by default, with video and audio time shifts of 12 and 2. An --euler flag switches to plain Euler.

How Good Is It? Who Measured What

Every number here comes from the authors. For audio-video, they ran a human preference study in which their 4-step MiniMax-H3 student was preferred 79.1% of the time over DMD2 and 84.6% over rCM, ties excluded. Other results in the paper: FID 1.04 for one-step ImageNet-64, FID 14.47 for 4-step SDXL on COCO-10K, and a VBench total of 85.15 for 4-step Wan2.1-T2V-14B. The authors call these the best among the few-step methods they compared.

Speed, also from the authors: about 9 seconds per model evaluation on an H200, or roughly 37 seconds per clip at the default size. On consumer cards they estimate (rather than measure) about a minute per evaluation on an RTX 4090 or 5090, and twice that on a 3090. They also note that 8–12 steps render fast motion (martial arts, a dunk) more cleanly than 4. What’s missing: no independent leaderboard (such as Artificial Analysis) has tested the DMAD checkpoints, and nobody has compared it with the existing community MiniMax-H3 Turbo LoRA, which already gets H3 down to 4–8 steps. Treat the preference numbers as promising, not settled.

DMAD vs. Other Fast AI Video Models

If you followed our coverage of MiniMax-H3 ControlNet Union 2.0, the pattern is familiar. The H3 ecosystem is moving fast, but the base license decides who can use any of it. For a quick look at how H3 itself holds up, watch our Seedance 2.5 vs MiniMax H3 prompt-to-video comparison on the Creative AI Show YouTube channel.

Try It Free: Demo, Weights and Exact Commands

No-GPU option (easiest): open the DMAD 4-step MiniMax-H3 Space. It runs on Hugging Face ZeroGPU, so it’s free with a daily GPU quota, and the quota is bigger if you’re logged in. Type a prompt that covers both picture and sound, click Generate, and the clip plays back with its soundtrack on the right. Settings like seed live under Advanced. Example prompts sit just below the button.

Local option (80 GB GPU, or 24 GB with --low-vram): the DMAD GitHub repo uses Python 3.12 and PyTorch 2.11 (CUDA 12.6). First accept the license on the MiniMax-H3 model page, then run:

conda create -n dmad_h3 python=3.12 -y && conda activate dmad_h3
git clone https://github.com/Yzmblog/DMAD && cd DMAD
pip install torch==2.11.0 torchvision==0.26.0 torchaudio==2.11.0 --index-url https://download.pytorch.org/whl/cu126
pip install -r requirements.txt
hf download MiniMaxAI/MiniMax-H3 --local-dir models/MiniMax-H3 --exclude "FL2VA/*" --exclude "Ref2VA/*" --exclude "transformer_ref/*"
hf download ZhengmingYu/DMAD --include "minimax_h3/*" --local-dir ckpt
python inference.py --model-dir models/MiniMax-H3 \
  --lora ckpt/minimax_h3/dmad_minimax_h3_4step_lora_critic.safetensors \
  --prompt-file prompts/dmad_sweater.txt --seed 42 --output-dir outputs/test --low-vram

On an 80 GB card, drop --low-vram. That flag, added to the repo this week, streams the 32B text encoder and 33B transformer from disk layer by layer, so peak GPU memory stays around 13–14 GiB, and the authors say the output is bit-identical to the full-memory path. Fast NVMe storage or lots of system RAM matters here: add --weights-in-ram (36 GB pinned) or --weights-int8 (18 GB, which changes the samples slightly) to cut disk reads. ComfyUI users get custom nodes in the repo’s comfyui/ComfyUI-DMAD folder, plus 2.0 GB ComfyUI-layout copies of both LoRAs in the comfyui/ folder of the Hugging Face repo. There’s also run_diffusers_pipeline.py if you’d rather use Diffusers. Note that the “Colab” and “Kaggle” buttons on the model page are generic Hugging Face launchers, not tested notebooks, and free Colab/Kaggle GPUs don’t have enough memory for a 33B model.

US, EU, UK or Korea? Use the free SANA-Video 2.0 4-Step demo (Apache 2.0, no audio). We broke it down in our SANA-Video 2.0 4-Step guide. If you need sound, use LTX-2.5 with its distilled 8-step checkpoint in ComfyUI.

Honest Trade-Offs

• The license is the deal-breaker for many readers. The MiniMax H3 Community License excludes the US, EU, UK and South Korea, and the DMAD LoRAs are licensed as derivatives. DMAD’s code is Apache 2.0, but that doesn’t change the weights. Elsewhere, products earning over $20M a year need written permission from MiniMax.
• Consumer GPUs work, slowly. 4 steps doesn’t shrink a 33B model. You still need about 170 GB of downloads, and on a 24 GB card the authors’ own estimate is 4–8 minutes of sampling per 5-second clip, and it gets slower still if you’re streaming weights from a slow drive.
• All quality and speed claims come from the authors. There’s no independent benchmark, and the consumer-GPU timings are estimates, not measurements.
• Only the H3 student is released. The Wan2.1 4-step results, which would be license-friendlier, have no public weights.
• ZeroGPU quotas run out. A long iteration session on the free Space will hit the daily cap.

Workflow Tips for 4-Step Video With Sound

1. Write the soundtrack into the prompt. Put dialogue in quotes and name the music and ambience (“mellow hip-hop beat, distant city chatter”). Leave it out and the model guesses.
2. Lock a seed, then iterate the words. At 4 steps, the seed has a big effect on composition. Keep it fixed while you tune wording so you’re comparing prompts, not luck. For fast action, raise --steps to 8–12.
3. Try both checkpoints. The full_critic variant scored higher on AVGen-Bench, but the paper’s lora_critic may suit your style better. They’re the same size, so swapping costs nothing.
4. Draft fast, finish slow. Use 4-step output to choose shots, then re-render the keepers with a full-step model or upscale them. This is the same idea as the frame-propagation trick in our Viggle-Animate breakdown.
5. Check your region before you publish. If you’re in an excluded territory, use H3-family tools for testing only, and make client work with Apache-2.0 or LTX-licensed models.

FAQ

Is DMAD free?

Yes. The LoRA weights and code are free to download, and the demo Space runs free on ZeroGPU with a daily quota. Running it locally is free in software terms, but you need an 80 GB GPU.

Can I use DMAD commercially?

Only outside the US, EU, UK and South Korea, because the weights follow the MiniMax H3 Community License. Businesses over $20M in yearly revenue also need written permission from MiniMax. The DMAD code itself is Apache 2.0.

What are DMAD’s system requirements?

Either one 80 GB GPU or a 24 GB consumer GPU using --low-vram, plus Python 3.12, PyTorch 2.11 with CUDA 12.6, about 170 GB of disk for MiniMax-H3, and 1.4 GB per DMAD checkpoint. For the low-VRAM path, fast NVMe storage or 18–36 GB of spare system RAM keeps generation from crawling.

Does DMAD work in ComfyUI?

Yes, through custom nodes. The repo’s comfyui/ComfyUI-DMAD folder adds nodes for DMAD’s re-noise sampler and sigma schedule, and the Hugging Face repo hosts ComfyUI-format copies of both LoRAs (2.0 GB each). It isn’t built into ComfyUI core.

How is DMAD different from the MiniMax-H3 Turbo LoRA?

Both cut H3 to around 4 steps, and both have ComfyUI nodes. DMAD is a research distillation method with a paper and human-preference results against DMD2 and rCM. The Turbo LoRA is a community release that recommends 6–8 steps for best quality. Nobody has compared the two head to head yet.

Accurate as of October 6, 2026. Licenses, demo quotas and model support change fast, so check the linked model pages before you commit a project. For more free AI tool breakdowns, browse Creative AI Show.