Ryuichi @ryuichi@social.cleverlibre.org

Lora, called flat and flat2, intentionally learns low-detail images, which apparently improves the detail of generated images when Lora strength is specified with a negative value. For example, it can be specified as <lora:flat2:-1>. It's a interesting idea.

https://huggingface.co/2vXpSwA7/iroiro-lora/tree/main/test2

"I've seen a lot of people in kimonos lately. I decided to try one on. How does it look?"

"最近、着物姿の人をよく見かけるようになった。試着してみることにしました。どんな感じなんだろう？"

#aiart #StableDiffusion #AIイラスト #catgirls

#AIGenerated #catgirl

#AIGenerated #stablediffusion

@ryuichi@qoto.org https://comfyanonymous.github.io/ComfyUI_examples/faq/ explains it better then i can.

The main benefit for me is control over upscale, instead of doing one big hiresfix step, I have a custom workflow that's a couple of small upscales instead of one big. The prompting is mostly the same, I never really used weights that much so the different behaviour isn't really an issue. Heres an example of my workflow, its somewhat overloaded, but i have plenty of flexibility with just moving couple of outputs around.