TL;DR - Trigger words are pussy, clitoris, clitoral hood, and cleft of venus. An explanation of these tags are available on the danbooru website, as most anime models are trained on booru tags.

If you appreciate the effort, send me a couple bucks: https://ko-fi.com/aisavant

Full Description

AI is bad at things like hands because of the variety of contexts and shapes they are in.

This is true for female genitals as well. Not to mention, most of the anime models out right now have been trained on censored posts which I believe skews the ability to draw somewhat realistic vaginas in their current state.

This is the first version worth releasing, but I hope to push further improvements.

Current status:

Trained on about 400 hand picked images

Carefully categorized and tagged training images to teach the lora on different parts of vagina anatomy. Current learned words are: pussy, clitoris, clitoral hood, and cleft of venus. An explanation of these tags are available on the danbooru website. While things like "cleft of venus" are not their anatomical name, most anime models are trained on danbooru tags, so it's best to stick to using them when training and using lora.

Locan trained at 896,896 resolution, .00025 learning rate, 5000 steps.

Results seem to outperform popular anime models by themselves, weather that's standing, sitting, solo, or group activities.

Long term trial and error to do list:

• Expand data set with even more images

• Try extremely low learn rate with extremely high epoch

compare and contrast training with different experimental lora features

Tried a retrain on Loha, experimental training features (Gamma SNR), as well as varied learning rate / epoch settings and haven't come up with any substantially better results. In fact I think this version is still the best I've produced.

Please let me know if there are any different training advancements that I should try to improve the LoCon.