Flux 2 Klein 9B Update:

Okay, well, forget ZIT then. Because whilst Klein 9B's results aren't broadly as realistic or compositionally impressive, it's so much easier to prompt with. The LoRA can just act as a character LoRA, and not having to make a style LoRA tricked into thinking it's a character one -- as in ZIT's case.

If you want to present Sakuyamon more creatively (i.e, uniforms or nude), lower the weight down to 0.6 ~ 0.7 and prompt the text with "woman".

As it turns out Klein can natively detect what the features in a LoRA, similar to Illustrious -- even Digimon. This is something ZIT could not do well from my time training with it, which is why features of a character had to be forcibly called upon via trigger words.

I should really just integrate the various model LoRAs into one profile like a normal person going forward.

Important Z-Image Turbo guidance:

General purpose, start with 0.75 weight and lower it to get poses you want. 0.85 ~ 0.9 is the highest you can generally go before overfitting starts to garble the image.

Include the descriptive features added in the activation texts and remove features where necessary (i.e. "midriff" if she's viewed from behind). I tried, but Z-Image Turbo isn't consistent or reliable at utilising a singular trigger word to create the character.

For simpler LoRA control, use Smooth Mix Ultimate (ZIT) and you can broadly generate the character with the activation word Sakuyamon X at weight 0.85 ~ 1.0, followed by your prompt instructions.

The LoRA is going to struggle with certain poses due to how Z-Image Turbo is trained versus SDXL. Let me know what you want more reliably recreated and I can account for this in possible updates.