Update 1:

Added a 18 vector per token version for better fidelity with certain models. One isn't necessarily better than the other. They are different. PxiBel2 will not (in general) generate a better or more detailed version of images made with PxiBel. It will create a different image.

It should generate a better likeness for checkpoints that weren't workinfg so well before. With both, it should be possible to get good results with most checkpoints. (Dreamlike Photoreal 2.0, Analog Diffusion 1.0 don't work well with either)

Description:

Turn of the century Euro Model ???? star Pixie Belle, AKANadia, aka a few other stage names.

Does well with a weight of 1.1 {e.g. (PxiBel:1.1) in Automatic1111 if you're new} but this depends on the checkpoint. Tested against 39 checkpoints with good results from 37 on weights between 0.7 and 1.2. Be aware that does well is fairly consistently generating a good likeness without a huge amount of retrying. After a number of training attempts by different methods or with different data, it feels like there are elements to her face Stable Diffusion doesn't want to recognize. Again, results will vary by model. Many of the more popular checkpoints tested tend to try to genericize her face. Weights between 1.1 and 1.2 help here.

Trained on prepared, hand curated, SFW images but she was a ???? star and the training set reflects that. NSFW checkpoints tend to want to have her shirt off more often than not. Average vector strength: 0.063 and average vector magnitude: 0.7695 so it should(?) be pretty flexible

My first "model" of any kind that seems ready for the public. Be gentle.