This model is a lycoris for SDXL. It was trained on the sdxxxl_v30 checkpoint.

The model was trained with 42 pictures found online. Training with variable resolution did not work.

The images were scaled to a vertical resolution of 1024. Then auto split, cropped, and remerged.

Creating images with a resolution of 2048x1024.

The captions for training were of the form,

"sixviews, a women standing ?????, [flat|small|medium|large] breasts,[color] [type] hair [with highlights]"

Any additional captions were added at the end and included details that stood out from the picture.

Examples: "jewelery,moles,tattoo,tramp stamp, watermark"

Recommended Settings

The model works well with a,

• 0.8-1.0 strength

• 4 CFG

• 30 steps

• resolution 2048x1024

• DPM++ 3M SDE Karras

Other setting may work as well these gave good results.

Strengths

The consistency is good (not perfect), for the body, face, breasts, and tattoos.

The model achives the correct poses ~3% of the time (tested 1/40 and 3/100).

The model achives close to the correct poses an additional ~5% of the time.

Lower horizontal resolution can yield a subset of the poses.

The model is able to create multiple consistent views of the subject.

Weaknesses

There can be too many or too poses.

Poses can duplicate.

The model is slow because the images are double the standard resolution of SDXL.

Other

Prompting "a blue haired women standing ?????" has better consistency than "a women standing ?????, blue hair".

The tattoos (and other details) can be surprisingly consistent between the different poses.

Works with other subjects (non-????? female).

Version 5

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

There will likely not be another version of the model. Versions that have a subset of the poses are more likely.