circlestone-labs / MiniMax-H3-Image-Training-Adapter

Overview

This adapter allows training MiniMax H3 LoRAs on images, while preserving the distillation and quality of the base model. It does two things:

1. Undistills the model
2. Biases the model towards an image-only distribution

It was created by simply training a LoRA on 10k diverse images with Gemma-4 31b captions, for about 10k steps, using a standard flow-matching loss.

Why

MiniMax H3 is a guidance distilled and/or RL'd model. Training it quickly breaks the distillation and lowers quality. Additionally, training it exclusively on images gradually degrades the video knowledge that the model has. This adapter has already undistilled the model, and also already caused the video degradation that image-only training would lead to. Therefore, when you do image-only training on top of this adapter, your LoRA (ideally) learns just the concept you are trying to teach, and not also the undistillation / video knowledge degradation.

How to use

Fuse this adapter into the underlying model weights and then train a LoRA on top of it. Do inference normally without the adapter. You can either manually merge the adapter into the MiniMax H3 weights ahead of time, or (more easy) use a training script that supports runtime adapter fusing, like diffusion-pipe ( example config ).

When using this adapter, DO NOT use a CFG-augmented training technique, "guidance preservation loss", or however else your training script of choice calls it. Just do normal training that you would use on a base model.

The adapter is designed first and foremost for image-only training, where it performs significantly better than the latest Ostris adapter as of this writing (2026-09-10). But it also seems to work pretty well for video training or mixed image-video training, though I haven't tested this much.

The adapter weights are in ComfyUI format (same as Ostris's adapter). It definitely works in diffusion-pipe, and it should work in AI-Toolkit also.