Hugely inspired by Master Corneo and all of his great concept embeddings (his work is amazing, check it out!). I wanted to learn how to make concept embeddings, so I tried my hand at the "back view blowjob."

No trigger words needed, just include the name of the embedding. It is definitely a touchy embedding. It can get overpowered, be overpowered, and it can generate some real mutants, but one in five or so generations is pretty darn good. With cleanup in img2img and inpainting, it can make real gems.

All sample images are straight out of img2img to show what it's actually outputting, and should be replicable. I used clip skip of 1, and otherwise pretty standard settings. hands_in_another's_hair and hands_on_another's_head can target for that. lying can get the girl off her knees. This model was trained exclusively on images where the girl's face is not showing or is barely showing, so face and looking_at_viewer make good negatives. I trained it on anime and photography, but I've only tested it on making anime so far.

I'd hope to improve the embedding in the future by better curating and tagging the dataset in the future.