How to train a LoRA for MiniMax H3?

last edited

LoRA training for MiniMax H3 works today through community trainers. Fizgig and AI-Toolkit both train for the MiniMax H3 Image model, and quantized weights fit a 12 GB card. The generic workflow is in LoRA basics.

Is there an official MiniMax H3 trainer?

No. The model card lists only inference stacks: SGLang, vLLM-Omni, diffusers, LightX2V and ComfyUI. Third parties supply it, compared in training tool comparisons. Adapters for the video model, MiniMax H3, do not carry over, since the image weights diverged.

How much VRAM does a MiniMax H3 Image LoRA need?

Fizgig measured on an RTX 5090 at rank 8:

base precisiontraining at 0.25 MPpeak with preview
int821.0 GB22.7 GB
4-bit nf411.3 GB12.8 GB
4-bit hqq17.0 GB (1 MP)17.3 GB

A 12 GB card also needs about 19 GB of free system RAM, since the text encoder streams from memory.

Which settings work for a MiniMax H3 Image LoRA?

The model is distilled to six fixed noise levels, and Fizgig trains each level through its own output head. The Fizgig preset named standard uses small 0.25 MP images. It starts the optimizer at 2.5e-4.

A frozen training adapter at strength 1.0 prevents caricature faces; it stays off at inference. If faces break, try an earlier epoch first. In ComfyUI, sampling uses euler, simple, 6 steps and CFG 1. Default dynamic loading applies a LoRA about 15% weaker; `--disable-dynamic-vram` restores full strength.

Can a MiniMax H3 LoRA be used commercially?

Only under the MiniMax-H3-Image license, which excludes the EU, UK, Korea and USA from its territory. Written permission is needed above 1 million USD revenue or for hosted services. A trained adapter is likely a derivative, so the same terms apply. Compare enterprise pricing.

Sources