I had a goal to make good Hajime Soroyama model from the start my (short) training spree, and its finally here , i was annoyed about poor quality images on the net and so i bought the art book ;) and scanned database myself. After few tries many different configs i think i trained 95% of what i want.
100 images, 85 epoch with split danbooru and description captions.
Grab it and generate some naught pin-ups ;-)
Description
Much bigger better quality dataset, different caption engine and bit of a QC on the captions ;-). overall more consistent with great fidelity.