It steers the model towards the style of gpt image 2. Improving details and lighting a lot, though the image might end up looking slightly sloppy.This is trained without a reg dataset, so the trigger word is not needed, but helps ever so slightly.v1.1 has been trained with better (manual) LR scheduling, and a few more image. It's honestly near perfect, so I don't think there will be another version for preview3.

STYLE > gpt-image-2

Training Settings: batch size 4, 1024px, ~114 images per epoch, optimizer is SinkSGD_adv

Files with SHA256: 5a2bd0706a9b018ea99ec5e2f7afafad38a9f316950bc5429b3da7b6be087251