home models images videos posts articles bounties tools challenges events updates shop

Ema Yasuhara (SHIROBAKO) | 11 Costumes | Character LoRA | IllustriousXL

Name: Ema Yasuhara (SHIROBAKO) | 11 Costumes | Character LoRA | IllustriousXL
Rating: 0 (0 reviews)
Author: TonySSS

329

228

Updated: Nov 13, 2024

character

anime woman lycoris shirobako sdxl

Verified: 7 months ago

SafeTensor

Details

Type	LoRA
Stats	205 7 1
Reviews	Positive (34)
Published	Nov 12, 2024
Base Model	Illustrious
Training	Steps: 2,700 Epochs: 21
Usage Tips	Clip Skip: 2
Trigger Words	ema yasuhara, black hair ey1, beige jacket, open jacket, pockets, grey buttons, long sleeves, white dress shirt, collared shirt, shirt tucked in, green necktie, pleated skirt, two-tone skirt, vertical-striped skirt, striped skirt, blue skirt, white skirt, black socks, black kneehighs, ankle boots, fold-over boots, two-tone boots, black boots, white boots, vertical-striped boots ey2, pink shirt, long sleeves, buttons, grey collar, shirt tucked in, blue skirt, high-waist skirt, suspender skirt, suspenders, black socks, black kneehighs, ankle boots, high heel boots, two-tone boots, light brown boots, black boots ey3, beige jacket, long sleeves, grey buttons, open jacket, pockets, two-tone dress shirt, black dress shirt, purple dress shirt, horizontal-striped shirt, vertical-striped shirt, red cross tie, shirt tucked in, collared shirt, red skirt, pleated skirt, two-tone socks, two-tone kneehighs, black socks, black kneehighs, purple socks, purple kneehighs, horizontal-striped socks, horizontal-striped kneehighs, ankle boots, fold-over boots, two-tone boots, black boots, white boots, vertical-striped boots ey4, black ribbed sweater, raglan sleeves, v-neck, orange flower, white dress shirt, collared shirt, red skirt, pleated skirt, black buttons, brown socks, brown kneehighs, two-tone boots, ankle boots, fold-over boots, black boots, white boots, vertical-striped boots ey5, two-tone cardigan, yellow cardigan, white cardigan, buttons, long sleeves, beige skirt, pleated skirt, black socks, black kneehighs, ankle boots, high heel boots, two-tone boots, light brown boots, black boots, blue dress shirt, collared shirt ey6, beige cardigan, open cardigan, raglan sleeves, buttons, grey shirt, collared shirt, white collar, green necktie, black skirt, black socks, black kneehighs, ankle boots, two-tone boots, white boots, brown boots, bow footwear ey7, multicolored dress, light blue dress, blue dress, white dress, vertical-striped dress, pink belt, short sleeves, pink collar, brown sandals, toeless footwear ey8, school uniform, blue blazer, grey buttons, long sleeves, pockets, white dress shirt, collared shirt, orange bow, grey skirt, pleated skirt, black socks, black kneehighs, brown loafers ey9, school uniform, grey vest, grey buttons, pockets, orange bow, white dress shirt, long sleeves, collared shirt, grey skirt, pleated skirt, black socks, black kneehighs, brown loafers +2 more
Hash	AutoV2 4B3A2BF0DD

1 File

About this version

default creator card background decoration

TonySSS

I came back from the dead after I discovered a wonderful SDXL model called IllustriousXL from a Discord server called Jerry's Waifu Lab made by AngelBottomless, so I decided to use it for training waifus hoping I would get better results than a 1.5 model. At first I tried using OneTrainer but found out that it has an issue with pivotal training when using a SDXL model where I cannot use a lot of additional embeddings otherwise I would get an error in WebUI saying "KModel version mismatch", effectively rendering the LoRA useless. So I was forced to go back using the good old Kohya SS GUI by Derrian I used years ago, using a config file from someone called Qelt who told me "You don't need a ****-ton of images anymore, 10~30 will do". I must say that it gives out the same accuracy from OneTrainer with Prodigy optimizer, only this time it uses Came optimizer with Rex scheduler, sadly Kohya SS didn't add pivotal training support at all, so I had to go back at manually tagging each 11 costumes of hers myself, at least I'm glad this website added its own dataset tagger which came really handy compared to WebUI tagger IMO being more compact and organized.

I'm quite pleased with the results, even tho some costumes' little details were completely ignored or not accurate at all, almost as much as with OneTrainer and pivotal training. If somebody here knows how to accurately capture costumes down to the smallest details from dataset, I'll be glad to learn everything about it.
I needed more than 800 steps in order to get all of her costumes more accurately to dataset images, with 2000~3000 being the sweet spot.

You can find a 1.5 model of this character I trained long ago inside my profile, but from here on I won't be training 1.5 LoRAs ever again ever since I discovered IllustriousXL, this ****'s even better than Aurora ONE as it has enough data for loads of characters and artists (doesn't mean I'll stop training artstyles though)!

Prompts for each costume of hers can be found on the right as I'm not bothered to post them all here anymore due to the amount of descriptive tags I had to add for each costume here for dataset.