This LoRA replicates the artstyle of the artist Cherrygig - https://x.com/gigcherry

Details

All example images are made purely in txt2img, then upscaled using 'Ultimate Upscale' extension with a denoising strength of 0.2 in img2img tab.

I use this upscaler: https://openmodeldb.info/models/4x-PBRify-UpscalerDAT2-V1

DPM++ SDE sampler, with Karras scheduler

20 steps usually, and 5 cfg.

Minimal inpainting is done only to the faces, with 0.4 denoise, and sometimes no inpainting at all.

Tips

(Experiment with adding or removing the keyword 'cherrygig' I no longer understand exactly how it affects things)

Be verbose when prompting for best results, it can work without it, however to reliably achieve the best style results, verbosity is better.

The dataset was captioned in the following manner listed below, prompting at a minimum these concepts will give you reliable results, also see the example image prompts for demonstration of this.

In parenthesis are either new tokens, or heavily used tokens in the training data.

Concepts

looking, (women looking at each other)

hair style, (lifted ponytail)

hair color,

eye color, (goat eyes)

lip color, (plain lips)

expression, (crazy eyes, ????, awe, smiling, smirk, parted lips)

clothing,

details, - meaning anything not in the surrounding categories

men, - (man's face, faceless male, multiple men)

background (space, stone wall, wooden wall, toilet stall)