MiniMax H3 Turbo LoRA: 4-Step Audio-Video Generation Preview

A community LoRA for MiniMax H3 renders joint video plus synchronized stereo audio in 4 sampling steps instead of the usual ~20, a roughly 5x speedup in sampling time. Released on August 5 by larryvrh as an early preview, it is the first speed LoRA for the open-weights H3 model (read the open-weights story).

This is an early prototype. The weights and the ComfyUI tooling are a work in progress: 4 steps work but come out noticeably softer, and 6-8 steps are the current comfort zone for sharpness. Treat it as a preview, not a finished product.

Weights

All files are bf16, about 744 MB, and apply as a standard low-rank update (W_eff = W + lora_B @ lora_A, alpha equals rank, no extra scaling). The LoRA targets the H3 base transformer; the video and audio streams run on two different flow schedules, so a stock sampler over-steps the audio at 4 steps and breaks it. The custom sampler node handles this.

The LoRA requires a non-pruned base checkpoint: the bf16 or full INT8 convrot diffusion model. The pruned variants use a different time-conditioning layer and are not compatible. A pruned-model-compatible conversion is available separately (see below).

ComfyUI usage

The original release needs custom nodes (ComfyUI-MiniMax-H3-Turbo, installable via ComfyUI-Manager or git clone into custom_nodes) plus a minimax_h3_turbo_4step*.safetensors file in models/loras/:

1. Install the custom nodes.
2. Download a .safetensors into ComfyUI/models/loras/.
3. Start from the official MiniMax H3 open-weights workflow, insert the Turbo LoRA between the model loader and the sampler, and replace the sampler with the MiniMax-H3 Turbo Sampler (4-step) with scheduler set to simple.

A ready-made text-to-video workflow is bundled in the repo (minimax_h3_t2v_turbo.json) and in the node repo's example workflows.

This is not native ComfyUI support: the speed LoRA depends on the third-party ComfyUI-MiniMax-H3-Turbo custom nodes.

Pruned-model conversion

The pruned/curve-form H3 checkpoints used by ComfyUI need a key conversion. drbaph published minimax_h3_turbo_4step_ckpt500_pruned_comfyui.safetensors (plus an EMA variant) at MiniMax-H3-Turbo-Lora-ComfyUI, loading with ComfyUI's built-in MiniMax-H3 LoRA loader, and bundled an example workflow (fl_minimax_h3_turbo_lora_example_workflow.json). QrusherZA also published a single-file pruned conversion at H3_Turbo_ComfyUI.

Early community tests at LoRA strength 1.0 with Euler/Beta report sharp 4-step output; the audio stream is the current weak point of this early checkpoint.

Availability

• Original weights: larryvrh/MiniMax-H3-Turbo-Lora (Apache-2.0)
• Custom nodes: Larryvrh/ComfyUI-MiniMax-H3-Turbo
• Pruned-model conversion: drbaph/MiniMax-H3-Turbo-Lora-ComfyUI, QrusherZA/H3_Turbo_ComfyUI
• Catalog: Turbo LoRA entries added to the MiniMax H3 model page with the required custom node flagged