Kinui Rena - Blue Archive / 衣斐レナ - ブルーアーカイブ

Trigger token: rena \(blue archive\)

school uniform for official outfit. Individual parts of the outfit are not tagged, besides the skirt.

twintails was left in the dataset for reinforcement. I have not needed to use it in testing, but I'm sure adding it may help stabilize her hair. Not that I had many issues, with about 80% accuracy.

A handful of images included her loose socks, shoes. Getting accuracy on those was not the focus, but they do generate reasonably well.

Typical LoRA issues, because I'm not a AI engineer, of some fine details specifically the hair bows in the twintails, too many, too few, and incorrect placement of the white bow. handbag generates correctly, most of the time.

Prompts and checkpoint models should be included in the examples images, with several different checkpoints used for compatibility. If not I'll comment on them.

Generated with an upgraded/customized version of Forge NEO with Euler DDIM. Upscaled by 1.5 with 2x-AnimeSharpV3 0.55 denoising, Euler A DDIM, + 0.5 to base CFG, and + 1 step from base.

Trained on 36 images (no AI, this time) with Derrian Distro. Config is embedded into the model.

State of LoRA's rant:

What ever happened to simple LoRA's, why do I have to include a college worthy paragraph of tags to destroy my prompt. Am I missing something? Does XL not care for tokens anymore? I care, and I care about how much you need to put in your prompt.

'Need' is the focus. When you train a LoRA you are training to capture what it sees as a concept. When you tag everything about the concept, the concept disappears. Examples of this type of LoRA training is all over CivitAI, a character is trained, and EVERYTHING is a trigger word. Stop that. Train the concept, i.e. the character, and remove tags that describe the character, please. Eyes, hair type, hair color, hair details, every facet of a base attire...

I don't know if I missed something over the past 2 years, I could just be stuck in the past.