Retrained version of previous lora with cleaner dataset with addition of generated photos that looked correct. Trained on regular settings with AdamW8Bit optimiser and Clip Skip 2 .