DMAD: LoRA adapters cut MiniMax-H3 audio-video generation from 50 to 4 steps
DMAD is a set of rank-128 LoRA adapters that distill the 33B-parameter MiniMax-H3 text-to-audio-video model from 50 denoising steps to 4. It generates 1344×768 video with native stereo audio at 124 frames and 24 fps, offering two 1.4 GB checkpoints, but requires the MiniMax-H3 base model.
- DMAD rank-128 adapters cut MiniMax-H3 denoising from 50 to 4 steps
- Output: 1344×768 video, 124 frames, 24 fps with native stereo audio
- Two 1.4 GB checkpoints: EMA at iteration 800 and trained at iteration 1600
- Requires the 33B-parameter MiniMax-H3 base model; no VRAM minimum stated
Read next
AI