4-step MiniMax-H3 Text-to-Audio-Video Generation
DMAD (Distribution Matching as Adversarial Distillation) distills a 50-step teacher into a 4-step student model, generating high-quality videos with synchronized stereo audio in just 4 denoising steps.
Distilled from 50 steps down to just 4 denoising steps for fast inference
Generates synchronized stereo soundtrack alongside video
1344x768 resolution at 24fps, ~5 seconds of video
Based on MiniMax-H3 omni-modal text-to-audio-video architecture