Yep, that's a fetish.Flux doesn't have all the base concepts that made training this model on Pony easy. If you can't get what you want out of this one, give the PonyXL or SDXL version a shot: <a target="_blank" rel="ugc" href="https://civarchive.com/models/590075/pants-wetting-panty-peeing-omorashi-concept">https://civarchive.com/models/590075/pants-wetting-panty-peeing-omorashi-concept</a>2024-08-24: v0.2 improves prompt adherence and fixes some image clarity problems. Anatomy can still be broken in some cases but I believe it is less prevalent than v0.1. It might have a harder time generating non-photorealistic images; try "sketch" or "doodle" in your prompt (wasn't in my dataset but seems to help) for flatter images. Higher guidance and/or CFG may also help.<h3 id="prompting-xcjj17te9">Prompting</h3>See the sample images for prompt examples. This is not designed to be prompted with booru tags. The images are all manually captioned, but in a style similar to VLMs. You shouldn't need to prompt exactly like the captions, but here is an example caption from an image in the dataset:"Digital artwork of a woman walking on a brick path next to a wooden fence and brick buildings while she pees in her blue jeans. There is a large shiny wet patch on her butt and the back of her jeans as evidence of her accident. She wears a black shirt with sleeves rolled up around her forearms, and her head is out of frame but her red hair is partially visible. Her left hand is on her hip and her right hand is at her side. The image is not fully realistic but the use of lighting and reflection is detailed and precise."For testing, I use a random prompt generator and it spits out garbled prompts like this and they mostly turn out okay:"A woman is peeing her panties, creating a wet spot on her panties around her vulva. She is in a half-crouched pose at a grocery store. She is wearing panties, navy blue plaid skirt, and unbuttoned shirt. She has red hair. She is looking at the viewer. Her expression is conveyed with a open mouth. She has amber eyes. The image is a sketch."<h3 id="settings-036lteh7j">Settings</h3>Should work for Dev and Schnell, but Dev is preferred. 1.0 strength and 3.0 guidance (for dev) is preferred. Using around 1.5 to 2.0 real CFG can help a bit in some cases but is generally not necessary. I have not tested usage other than ComfyUI with fp8.<h3 id="training-7nztjw5st">Training</h3>v0.2 of this LoRA was trained with custom scripts on a 4090. It was done in a "perpetual stew" style by training from the previous LoRA weights while improving the captions and dataset composition between runs, so I couldn't tell you exactly how many steps, but I think around 50k by this point. Most runs had around 200 concept images pulled from specific categories to try to get an even balance between subconcepts like skirts, shorts, etc. There is a definite bias towards pants, though. Image sizes during training were between 0.3 and 2.0 MP preferring their original size, but I downscaled images with significant artifacting. Most training runs included around regularization image unrelated to the concept for each concept image.

Files with SHA256: a4cf5d440defcd69bdf0842a65ed44c658d5a072648084ec4f8725c0d5f2271d