Texas A&M's DMAD Releases 8-Step MiniMax-H3 LoRA for Higher Video-Audio Fidelity
A new 8-step DMAD LoRA raises the quality ceiling on MiniMax-H3 text-to-audio-video, keeping 8 GB GPU support and ComfyUI drop-in workflows.
Takeaways
- New 8-step DMAD LoRA for MiniMax-H3 delivers higher audio-video quality than the earlier 4-step students.
- Rank-128 LoRA, 1.4 GB, distills the 50-step 33B teacher down to 8 model evaluations.
- Runs 15 s of 1344x768 video with stereo audio on a 16 GB GPU; 5 s clips fit in 8 GB.
- Drops into existing ComfyUI workflows via a single file swap, no custom nodes needed.
- Based on the DMAD paper from Texas A&M and ByteDance, reframing distribution matching as classification.
- Training code and H3 training data are public on GitHub under base-model licenses.
The DMAD project has released an eight-step student LoRA for MiniMax-H3. Existing users of its four-step distillation can adopt the checkpoint by replacing the adapter and setting the sampler to eight denoising steps; the published ComfyUI graph, base model, resolution, and audio path stay the same. The files are available in the model repository.
MiniMax-H3 is a 33B-parameter model that generates video and stereo audio together. Its teacher normally uses 50 diffusion steps to produce a 1344 × 768 clip. The DMAD paper, from Texas A&M and ByteDance, describes how Distribution Matching as Adversarial Distillation trains a smaller sampling process to approximate that teacher. A LoRA stores the resulting behavior as compact adapter weights applied to the base model.
One file, eight denoising passes
The new Diffusers-key checkpoint is dmad_minimax_h3_8step_full_critic.safetensors, a 1.4 GB rank-128 LoRA trained from scratch for eight-step sampling. The full_critic label describes how the critic was trained during distillation. Generation still uses the MiniMax-H3 base model and the released LoRA.
| Checkpoint | Training recipe | Intended use |
|---|---|---|
4step_lora_critic |
Four-step moving-average checkpoint from the paper. The critic backbone stayed frozen while LoRA adapters were trained. | Paper reproduction and fast generation. |
4step_full_critic |
Four-step student trained with the full critic backbone updated. | Fast generation with higher reported AVGen-Bench scores than 4step_lora_critic. |
8step_full_critic |
Eight-step student trained from scratch with the full critic. | Higher reported audio-video fidelity at roughly twice the denoising work. |