This model is a lycoris for SDXL. It was trained on the sdxxxl_v10 checkpoint.

The model was trained with 17 pictures found online. The pictures were cropped to remove non-subject matter.

The captions for training were of the form,

"a women [sitting|lying|reclining] on a [couch|bed|chair] fisting her pussy, fisting to [wrist|knuckles]"

Any additional captions were added at the end and included detailed that stood out from the picture. Examples: "naked, wearing shirt, outside, glasses"

Settings

The model works well with,

• 0.7-0.9 strength

• 3 CFG

• 40 steps

• resolution of approximately 700x1100 (width x height)

• DPM++ 2S a Karras

Other setting may work as well these were found to have the best results, but the testing was not exhaustive.

The model responds well to different prompts for clothing.

The model can change the position and the location.

Lying and reclining have a higher success rate compared to sitting.

Weaknesses

The model struggles with drawing entire legs as they weren't always in the training data.

Sometimes it will generate hands instead of feet, or try to fist with two hands instead of one.

As there were no fingers in the training data, hands are nubby when not generated at the correct location.

The model can generate at other resolution but the success rate is lower because of the issues drawing legs correctly.

Teeth are frequently of poor quality, a negative prompt of "open mouth" is recommended

Version 2

A new version will be released if the model can be improved.

The largest issue right now is good quality high resolution training data.

The resolution of the training data was approximately 700x1100 (or 1100x700), higher resolution would yield a model that produced sharper images.