Today we're gonna create an OC LoRA with multiple forms!
This is probably common knowledge to those who make LoRAs with multiple outfits, but I'm just going to document my process and apply it to 1 pic OCs.
Note: This article will skip over a lot of information I've previously mentioned in my other articles. Basic LoRA knowledge is recommended, as I'm not good enough to explain such things myself.
Now, Osayo here has two different looks, Human and Kami. Human has an alternate outfit with the blazer off, too.
The Kami form is not just Osayo in different clothes, but with different hair ("starry hair")
I believe the Kami form will be the most difficult, so I'll put slightly more focus on it, then blouse, then blazer.
BASE IMAGES
Prompt:
perfect quality, best quality, masterpiece, 1girl, solo, beige background, full body, mature female, standing,
blunt bangs, black hair, very long hair, sidelocks,
red eyes, tsurime, eyelashes, expressionless, lips, white pupils,
skinny, small breasts,
dark green suit,
office lady, pants, office lady, blouse, high heels,
id card, red lanyard,
With Kami, I prompted:
perfect quality, best quality, masterpiece, 1girl, solo, beige background, full body, mature female, standing,
blunt bangs, black hair, very long hair, sidelocks,
red eyes, tsurime, eyelashes, expressionless, lips, white pupils,
skinny, small breasts,
green kimono,
japanese clothes, layered kimono, red shawl, okobo, lace, white lace, lace trim,
My process here was to generate multiple pictures until I got an outfit I liked, then cut and paste the head from Human picture onto it. This is to make sure they look like the same person.
I saved over the original, oops.
Then, I just edited her hair into starry hair using a picture I got off Google.
I just did this in Clip Studio Paint, then did img2img to refine it.
img2vid, lots of img2img
I used Kling here on Civitai (https://civitai.com/generate) to generate three videos, one of each picture. After that, I screenshotted frames off the video and used img2img to upscale them.
I ended up with 18 images in my "7_osayo" folder
"4_artstyle" contained 94 images.
That's a lot.
Tagging
Usually I just slap tags on, but here I'm gonna be a little more careful.
Human Osayo's activation word will be "osayosenjou"
Kami Osayo's activation word will be "osayonokami"
I make sure all Kami pictures have "osayonokami, green kimono, red shawl, starry hair, lace" first in the captions, as I'll use those to help prompt that form.
Human pictures have "osayosenjou, blouse, lanyard, id card" first in captions.
Training
While I usually include a regulations folder with non-AI pictures of anime girls, this time I excluded it just because of my amount of images. I'm unsure if it did anything, but the LoRA turned out well!
(Excuse the difference in style, I used an artist tag)
Both forms generate! Although the Human forms skin tone is slightly lighter.
That's okay, I just edited the base images in Clip Studio Paint and darkened her skin, as I'm making a second pass using images from the first LoRA.
This is to refine the second LoRA and make it more flexible.
I set up some wildcards and generate an amount of pictures of both forms in various poses, lowering the strength of the LoRA occasionally. Human form could generate poses very well, but Kami form tended to have her hands in a similar pose.
Using these images, I pop them back into the oven.
Training (Again!)
{ "kami": { "n_repeats": 7, "img_count": 17 }, "human": { "n_repeats": 7, "img_count": 10 }, "3_artstyle": { "n_repeats": 3, "img_count": 94 }, "1_girl": { "n_repeats": 1, "img_count": 90 }
This time, I did use my regulations folder ("1_girl")
On epoch 4/10, the LoRA was satisfactory to me!
You can download Osayo's LoRA here:
https://civitai.com/models/1699219/illustriousoc-collection-6