Trained on hand-captioned 131 image dataset.
Recognizes several styles: brief underwear, speedo, thong underwear, boxer brief underwear
Recognizes poses: in profile or looking to the side helps with generating a visible bulge.
Trained using full captions, not tags. No trigger words are required, but for best (horniest) results, use thong and bulge in your prompt.
This version still leaks concepts into the rest of the generation, and degrades image quality significantly.
