Accelerate the unquantized MiniMax-H3 model from 28 steps to just 4-8 steps. High-fidelity motion and synchronized stereo audio in a fraction of the time.
MiniMax-H3 generates beautiful videos with synchronized stereo soundtracks, but the base model is computationally heavy and requires 28 sampling steps by default.
Using the optimized Turbo LoRAs, you can achieve equivalent composition quality, motion consistency, and audio stability in only 4 to 8 steps.
The Turbo LoRA integrates directly into ComfyUI. Click through the pipeline steps below to see exactly how to wire the nodes and configure their settings:
Load the MiniMax-H3 base checkpoint
Apply the dynamic SVD-resized Turbo LoRA
Euler + Beta scheduler with specialized sigma shifts
Separate video and audio streams into a stitched MP4
Load the main ComfyUI-compatible base weights of the MiniMax-H3 model. In ComfyUI, this is loaded into the workspace using the `Load Diffusion Model` node.
| Parameter | Recommended Value |
|---|---|
| Model Checkpoint | Comfy-Org/MiniMax-H3 base model |
| Precision | bfloat16 (unquantized for best quality) |
These weights have been converted to the ComfyUI namespace and pruned/compressed using singular value decomposition (SVD). Click any file name to download it directly:
| LoRA Filename | Type | Avg Rank | Size | Frobenius Energy Retained |
|---|---|---|---|---|
| minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors | Pruned EMA (Recommended) | 128 | 354 MiB | - |
| minimax_h3_fl2v_turbo_8step_v1.0_comfyui_resized_avg_rank_21_bf16.safetensors | LightX2V FL2V (8-Step) | 21.7 | 311.9 MiB | 97.72% |
| minimax_h3_fl2v_turbo_4step_v1.1_768p_comfyui_resized_avg_rank_28_bf16.safetensors | LightX2V FL2V (4-Step) | 28.5 | 375.6 MiB | 93.59% |
| minimax_h3_ref2v_turbo_4step_v0.1_comfyui_resized_avg_rank_21_bf16.safetensors | LightX2V Ref2V (4-Step) | 21.7 | 311.8 MiB | 98.33% |
To run these LoRAs, place the downloaded .safetensors files in your local ComfyUI setup:
ComfyUI/models/loras/
Make sure you also download the example workflow file fl_minimax_h3_turbo_lora_example_workflow.json and load it in ComfyUI to populate all required sampler nodes and connection routes.