| \r\n\t\t\t\t\t\t\t\t \r\n\t\t\t\t\t\t\t\t\t \r\n\t\t\t\t\t\t\t \r\n\t\t\t\t\t\t\t\t | \r\n\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\tManicMoon\r\n\t\t\t\t\t\t\t | \r\n\t\t\t\t\t\t
I'm sure it'll come in handy some day for some more pregnant hangouts ^^
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 Tamina, for instance, would know how Tamina is supposed to look like, so 'force' the model to generate more Tamina-like images.
For a character LoRA you need about 20-50 high quality images in various poses, with various expressions, with various backgrounds.
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.
Tamina'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.
That's the last dataset I'll share, otherwise it'll turn into spam, but I hope the three datasets showcase my hope for character consistency well!