I decided to release this experimental style LoKr that builds on an older dataset (here). I think it's different enough to warrant its own release. Again--just testing things. You get what you get. See below for suggested settings.

Trained on a few hundred hand-selected publicly available images of early-2000s emo/scene style selfies. It combines for the visual style of the photos taken at the time as well as the fashion/hair/jewelry aesthetic.

In comfyui, you load it up just like a LoRA. Set the strength to between 1.0-4.0 to increase the magnitude of the style.

Suggested settings:

• Important: Strength sweet spot is 2.0-2.5. You can go beyond that up to ~ 4.0, but Klein's poor anatomy starts to creep back in.

• Use er_sde/sgm_uniform for decent results. ~5 steps. CFG 1 (obviously).

• No trigger

For prompt, try something like:

Early 2000s style amateur point-and-shoot mirror pic of an emo 23 year old woman in a bedroom, leaning in, ponytail brunette blonde hair, messy eyeliner, grungy short shorts and black low cut crop top punk t-shirt that says "RLY", yelling, scrunchie on wrist, bracelets, small devil tattoo, jpg quality

Or, use the wildcard prompt added to the showcase pics with a wildcard text node.

Caveats:

• Flux anatomy

• Cameras that'll make you go "what are those!?"

• Mostly selfies, even when unintended

• Sometimes nsfw

Let me know how it goes for you and maybe I'll build on it. Or delete it. We'll see.