This is, I’m quite certain, the only Madison Beer Lora for SD 1.5 trained exclusively on well-varied (in terms of pose, angle, etc) images that were significantly higher than 1024×1024 to begin with, properly cropped to the following resolutions (all 1024 on at least one dimension):
512×1024
768×1024
1024×1024
1024×768
1024×512
Checkpoints I have personally found this to work particularly good with include PicX Real 1.0 and Babes 3.1. “PsychomergeRealistic” as used in (some of) the example pictures is a personal checkpoint merge that I plan to also release on CivitAI at some point relatively soon.
This Lora was trained at Clip Skip 1. Clip Skip 2 has no meaning or purpose beyond being what Novel AI (for unclear reasons) trained their eventually leaked in-house model on. That model was then adopted for the training of various 2D anime focused models. All of which is to say, Clip Skip 2 was not at any point and still is not currently some kind of magic image optimization technique, it has literally no relevance whatsoever beyond exactly what I just described.
This Lora may be updated in the future with a larger data set, although I’m very pleased with how it turned out for this release in any case, to say the least.