I did not like how Angela is generated by many models and I decided to make my own LoRA. Perhaps it overfits a bit, so try weights from 0.6 to 1. I also noticed that using only the trigger word lora_sarafyan_girl without additional angela sarafyan im promt gives the best results, so the checkpoint model does not use own knowledge for the subject. The bigger v0.1 model gives more photo realistic results, the smaller v0.1t model tends to be more cartoonish.

Angela Sarafyan

Same dataset as in v0.1 but trained for 2 epochs on 32 ranks. That keeps the model smaller. The inference results are a bit cartoonish. If you prefer photorealism use v0.1. Enjoy.

Files with SHA256: 09452b711dc7464a0d06181a9ea524d586fd10cee747f4727ee4957a1b40e0c0