DMAD gives MiniMax H3 users a new reason to investigate faster local audio-video generation. The important question is which execution path you are using: the authors' reference implementation, their Diffusers example, or a community conversion for ComfyUI. A matching LoRA name and a four-step setting do not establish that those paths behave identically.
This article explains how to choose and evaluate a path. It is based on the released sources reviewed on October 13, 2026; H3Vid has not benchmarked DMAD or added it as an online generation preset.
AI-generated concept illustration made with Codex imagegen; not a DMAD output or interface screenshot.
What is DMAD for MiniMax H3?
DMAD means Distribution Matching as Adversarial Distillation. The authors' repository releases a distilled student of MiniMax H3's text-to-audio-video model as a transformer LoRA. Its headline is four model evaluations instead of the teacher's 50-step generation, with video and stereo sound generated together. This is an acceleration adapter, rather than a character or visual-style LoRA.
That distinction changes what you should test. A style adapter is usually evaluated for appearance. A distillation adapter also needs its intended sampling procedure: reducing the step count on a different setup is not sufficient evidence that you reproduced the released method.
Which checkpoint should you start with?
The official weight card lists two rank-128 adapters, each approximately 1.4 GB:
| File suffix | Role described by the authors |
|---|---|
4step_lora_critic.safetensors | Paper checkpoint, using exponential moving average weights |
4step_full_critic.safetensors | Full-critic variant, reported to score higher on AVGen-Bench |
For a first reference run, choose the paper checkpoint because it has an explicitly associated example in the release. Save the complete filename with your output. Evaluate the second checkpoint separately with the same prompt and dimensions if you want to compare them. The authors' benchmark statement is not a guarantee that one variant wins for your scenes.
These files are adapters, so they still require the base model components. The model card also identifies the weights as derivatives governed by the MiniMax H3 Community License Agreement. Check the linked terms for your intended use rather than treating the code repository's license as the weights' license.
Why four steps can mean different settings
The reference inference code separates prompt encoding, transformer sampling, and video/audio decoding. Its default settings include four evaluations, video time shift 12, audio time shift 2, and 124 frames at 1344 by 768. It selects the re-noise rollout unless the Euler option is enabled, then writes both the output and a settings record.
By contrast, the Diffusers runner passes steps + 1 to the pipeline while describing the desired count as model evaluations. Its comments explicitly distinguish the pipeline's Euler rule from the reference re-noise rule. Therefore, a displayed count of five in that specific pipeline should not automatically be interpreted as a different five-evaluation student.
When comparing outputs, record the runner, scheduler, shifts, checkpoint, dimensions, frame count, and seed. Inspect the code path before changing a number to match someone else's screenshot. This avoids confusing an interface setting with the underlying schedule.
Can you use the ComfyUI conversion?
Kijai's experimental collection lists minimax_h3_DMAD_4step_full_lora_avg_rank_39_bf16.safetensors, approximately 367 MB. Its filename and size differ from the authors' full checkpoint. Save that precise identity when evaluating the conversion; advice shown for a different distillation adapter is not automatically a DMAD recipe.
For an existing graph, identify the base model loader and LoRA loader first. Then inspect the sampler and video/audio branches. A successful load confirms that the graph accepts the file, but does not confirm that its sampler follows the reference algorithm. Consult the converter's notes and the workflow author's version requirements before assuming compatibility.
Keep your working base-model graph available. Change one setting at a time and save each graph alongside its output. If the converted adapter is faster but loses fine detail, that is a tradeoff to investigate in that particular configuration. It does not establish the quality of the authors' reference route.
For general setup and model placement, start with the existing MiniMax H3 ComfyUI overview. This article addresses the adapter and sampling distinction rather than repeating a complete installation guide.
Hardware: a small adapter is not a small model
The release documentation now describes a streamed consumer-GPU path in addition to its resident reference route. The October 10 release notes extend it to 8 GB GPUs. The small adapter still requires the base components: roughly 170 GB of downloads are listed, so the adapter size is not the installation size.
| Documented GPU preset | Clip boundary | Execution choices |
|---|---|---|
--low-vram 8 | 5 seconds, 124 frames | 8192-row chunks and video VAE offloading |
--low-vram 12 | 15 seconds, 362 frames | The same chunk and decode strategy |
--low-vram 16 | 15 seconds | 16384-row chunks; VAE stays resident when enough memory is free |
These are the authors' supported configurations, not our verified minimum hardware requirements. Their published peak-memory measurements were taken on an H200 with memory constrained, rather than on every card in this table. GPU architecture, desktop allocations, host memory and storage performance still matter. The authors did not measure pre-Ampere cards, whose emulated BF16 path may be much slower.
The low-memory route streams text-encoder and transformer layers instead of keeping the complete models on the GPU. --weights-in-ram retains a 36 GB pinned transformer copy; the documentation recommends it with at least 48 GB host memory. --weights-int8 halves that weight traffic and host copy, but changes samples. Do not group it with the authors' bit-identical streaming/offload claim. VAE offloading preserves output in their comparison but adds about 1–1.5 minutes of decode time on the measured H200.
The repository also links a five-second ComfyUI graph for 8 GB GPUs. Keep its checkpoint and sampler identity separate from an older community conversion; a shared DMAD label does not establish equivalent execution. Save the exact graph and review its current instructions before using the new memory boundary.
How to evaluate a useful speedup
Use a small set of scenes you actually need: a person speaking, a moving camera, and a scene with fine background texture are useful starting points. Keep the prompt, dimensions, frame count, and audio requirements fixed across the baseline and adapter runs. Review motion and sound, not only the sharpest still frame.
Measure the time from submission to the playable file, and separately record loading, sampling, decoding, and encoding when logs allow it. Repeat after a cold start and with models already loaded. A faster sampler can be hidden by download or offload costs; a warm run can look unusually fast because setup has already finished.
Finally, decide whether the result serves your task. Fast previews can justify a visible detail tradeoff, while final dialogue shots may require more reliable sound and faces. Keep the best usable result and its saved graph or settings. That evidence will be more helpful than converting the 50-to-four step ratio into an assumed end-to-end speedup.