Skip to main content
Back

MiniMax-H3-Turbo-Lora Cuts Video and Audio Generation From 20 Steps to 4

A community LoRA accelerates MiniMax-H3 joint video and stereo audio generation to as few as 4 sampling steps, roughly 5x faster than default.

  • MiniMax-H3-Turbo-Lora cuts joint video plus audio sampling from ~20 steps to 4-8.
  • Over 680,000 downloads on Hugging Face, Apache-2.0 licensed, 744 MB bf16 weights.
  • Recommended checkpoint is v4_step600_ema, with v1 as fallback for 4-step heavy motion.
  • Works with all H3 base variants including pruned FP8, keep strength at 1.0 and scheduler on simple.
  • Runs via ComfyUI-MiniMax-H3-Turbo nodes or a standalone generate.py script.
  • Base model is ~33B parameters, 32 kHz stereo audio, up to 2K at 24 fps for 5 to 15 seconds.

A new LoRA adapter for MiniMax-H3, the open-weights omni-modal video model, has pulled in over 680,000 downloads on Hugging Face this month. MiniMax-H3-Turbo-Lora collapses the sampling loop for joint video plus synchronized stereo audio generation from roughly 20 steps down to as few as 4, without swapping the base model or retraining it.

The adapter targets a specific pain point: H3 is a heavy diffusion transformer that renders picture and sound in the same forward pass, and full-quality sampling is slow. Cutting to 4 steps delivers roughly a 5x speedup, with quality improving steadily as you add steps up to about 8.

What the base model actually does

MiniMax H3 jointly processes text, images, video, and audio in a single context, generating video with native stereo audio where voice, sound effects, and music are modeled together in one forward pass rather than layered on afterward. Output goes up to 2K, 24fps, and around 15 seconds. It is MiniMax's third-generation video model, following Hailuo 01 and Hailuo 02, and the first released with open weights.

The Turbo LoRA sits between the base loader and the sampler in a standard ComfyUI H3 workflow, alongside a custom sampler node that handles H3's dual-schedule quirk: video and audio latents ride different flow schedules and each must be integrated on its own clock.