Fizgig MiniMax-H3-Image Training Adapter
A training adapter for MiniMax-H3-Image, made for Fizgig. Keep it on, frozen at strength 1.0, while you train a LoRA on MiniMax-H3-Image, and leave it off when you use the LoRA. Without it, a LoRA trained straight on MiniMax-H3-Image turns faces into caricatures; with it, the LoRA learns the person with photographic skin and the model's own quality.
Fizgig downloads and uses it for you. It's a standard LoRA file (kohya keys, rank 16, attention and MLP of all 50 blocks), so ComfyUI loads it too.
Training: real photos plus MiniMax-H3-Image's own renders, about 1 MP, the model's 6 distilled noise levels each through its own head, AdamW at 1e-4, EMA 0.998, 4,000 steps. Only the adapter is published here, not the training images.
Licence: this adapter is a Model Derivative of MiniMax-H3-Image and is distributed under the MiniMax-H3-Image Community License Agreement. Its use restrictions (Section V and Exhibit A) apply to you as a user of this file. See the NOTICE file.