Haeun Flux/XL/1.5版本1.5 (ID: 976253)

This LoRA model was created to generate images of an original character I created for artistic reference. It's just my way of eliminating "same-face" in my illustrations!

Flux:

Flux generations were in ComfyUI with a strength of 1.

SDXL:
For SDXL I used realvisxlV40 Lightning

DPM++ SDE Karras at 4 steps with a CFG of 1.5 with a LoRA strength of 1

SD 1.5

My typical settings are DPM++ SDE Karras at 13-15 steps with a CFG of 5 to 7 with a LoRA strength of 0.8 to 1.

The Process:

This was a fun experiment. I normally used FaceApp to make new people, but FaceApp removed its face transfer tool, so I had to develop something else. I went to ComfyUI because I knew there was a face transfer tool. Even with FaceApp, I used images that were a bit more painted, but I wanted to see how far I could take it. I took an image I made for a previous Inktober and loaded it in. It worked! I took a previous dataset and placed her face over the others, upscaling and adding a bit of noise in the process.

The photos weren't perfect, so I did a little editing in Affinity Photo. I had to stretch her lips a bit and remove her double eyelid. I also had to play with a few smiling images because the teeth came out poorly.

I trained her in SD 1.5. I did this because the transfer tool didn't do well with full-body shots, and I wanted to fix her smile. That's when I had a happy accident! Because of how I had to stretch the teeth in the smile, the AI tried to "correct" it to what it knew. This gave her a nice and unique smile!

I took a few of the AI images and put a Gussian Nose filter over a few of them, only about 2-10%, and only on 5 images. Then I trained her again.

Thanks for taking the time to read this! I hope you enjoy using her for your generations!

描述:

训练词语: