User LoRA
XL
User LoRA
ema yasuhara
https://civitai.com/models/827419/ema-yasuhara-shirobako-or-11-costumes-or-character-lora-or-sdxl
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.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.