Eryn
Eryn is a step in a different direction - a bit of an experiment, you might say.
Eryn was originally meant to be Brie's busty brunette companion, and she is a bit of that. Eryn is a completely fictional and AI-generated likeness of a busty woman with straight brown hair and brown eyes by default, and her facial structure is trained with hand-crafted micro-features in her face, carefully preserved in the training data. As a brunette version of Brie, she should be perfectly adequate, though her appearance should seem less "accidental".
She can also be posed with other people, unlike Brie, and her figure is more consistent relative to her body, which I'm continuing to improve over time.
But wait. There's more.
The training data includes images of the likeness in fantasy settings with vivid details, alternative forms (i.e., objects shaped like her), and diverse art styles. The result is that kxeryn has been trained as an integrated adjective that will cooperate if you ask it to put Eryn's form and likeness on any arbitrary creature or object. You can also call her a "model" (a kxeryn model) to have her show off the outfit she's wearing, or include her in a charcoal drawing, octane render, claymation still, comic book art, a crayon drawing, a ukiyo-e woodblock carving, pinup art, or whatever you like; mix and match is expected to work too.
Issues / What's next
I'll be continuing to improve this dataset and releasing improved versions as time and resources permit.
Areas of focus:
Always looking for prompts or angles that produce a wrong figure and finding ways to fix that.
Fictional characters tend to require extra specific outfit prompts to get modest outfits. They sure do like wearing collars, too.
Expand support for non-human objects shaped like Eryn.
The training data already includes an eryn-shaped decanter, a skyscraper, a lighthouse, a glider mid-flight, and a bust statue. For humanoids there is a sand giant, a tiny butterfly, a honey nymph, and a porcelain golem.
Description
A checkpoint from earlier in the same training run as v2. I've got better results from this checkpoint for some prompts.













