Sample prompts:

a high quality selfie photo of a sexy skinny naked Asian cumface woman with lots of white, thick, gooey cum all over and covering her face, cheeks, hair and forehead. The cum coats her face in a thick layer. Her eyes are open while holding a sign with one hand saying "Make Sticky Semen Selfies!" She's smiling and looking at the viewer.

a high quality photo of a young sexy skinny naked cumface woman with lots of white, thick, gooey cum all over and covering her face. The cum coats her face in a thick layer.

Sample Prompt:

a high quality photograph in the style of cumface of a sexy young naked Asian woman with a black leather collar that says "SLUT". She is looking up and her face and forehead is fully coated in tons of thick, dripping, white cum. Another woman is standing in a black leather dominatrix outfit and is behind her dumping tons of cum from a bowl on her face. Her whole head and body is coated in the cum dripping from above her. There is a ton of white cum coating the naked Asian woman's entire body and face in an excessive, sticky, messy layer. The naked Asian woman has her tongue out holding a sign saying "INCLUDED! CUM DUMPING!"

This is v3 of Cumziilla. This release doesn't seem to perform any better than v2, so you might want to try that out as well. It was trained on the full flux1-dev model. The model needs a large amount of prompting help to render cum properly. See example images for help with prompting - e.g. if you don't tell it cum is white, you're going to get lots of random colors. See example images for prompting advice.

Changes

Using Joy captions (highly descriptive) for training - don't notice much of a difference

Increased number of images 50%

Decreased steps to just 50 epochs vs 150 in v2. I have compared versions at different training durations and I don't notice much of a difference. The training loss doesn't change much, so I assume lower steps is the move here.

Trained on a new LORA config.

NOTE: this LORA performs poorly against the flux1-dev-fp8 quantization of Flux. Don't be surprised if results are off. It was trained on and performs the best on flux1-dev. Not sure how to help this in terms of training. One user suggested using Heun with 8 steps for the fp8 quantized Flux.

credits: NSFWrobot