Undistills MiniMax-H3-Image, making it act more similar to a base model. Designed mainly for training. Unlike Ostris's V1 adapter, this one is trained using a continuous timestep distribution, rather than a discrete timestep schedule. You should train on top of it like any other base model, without additional tricks or special configuration. Works well for training in diffusion-pipe. AI-Toolkit might be hardcoded for Ostris's adapter and the discrete timestep schedule, so it may not work properly there.

Trained on 10k diverse images with Gemma-4-31b captions for 15 epochs. Equal mix of 512 and 1024 resolution. The model was trained without architecture modifications: it still selects a PDD head based on the timestep, following the ComfyUI implementation, but it is trained to accept any continuous timestep between 0 and 1.

Weights are in ComfyUI format.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support