LoRA trained on 58 of my favourite images that I've made using stable diffusion, and hand captioned (dataset will be posted for download). Trained at 768x768, and I've taken the version after 2320 steps. I think it adds more details to the subject, and makes more interesting backgrounds. There does appear to be a bias towards starry skies/nighttime images which may require prompt weights to counteract, but I like it.Sample images made using a variety of models, varying from photorealism to full anime - seems to work on any model.

Files with SHA256: 8de8ab7f31060e00732db928c2a2787a139493ea5d80e11357d284bb204e8bb0