So Juliette is a massive pain to get consistent. The floof on her wrists and ankles is pure torture. She was the direct reason I started looking into LoRA generation XD
A LoRA is basically a specialization that you stick on an existing model, which makes the model do something specific even better. A character LoRA of Juliette, for instance, would know how Juliette is supposed to look like, so 'force' the model to generate more Juliette-like images.
For a character LoRA you need about 20-50 high quality images in various poses, with various expressions, with various backgrounds. Because Juliette was the first, I hadn't thought of varying the background yet, but that's easy enough to fix.
The combination of Flux Kontext with post processing with Novafurry is very powerful: It can generate extremely consistent characters from a single image. I've attached a workflow I've adapted that takes in an image, then spits out 12 images with various background, poses and expressions. The reason why is more in-depth explained in this submission: https://inkbunny.net/s/3706067.
Juliette's dataset is SFW for now, but judging from my experiments with Sarah it should be no issue to generate the nude part of the dataset later on.
