Finetuned with WD14 captions, longer prompts seem to do better in general, but style and quality can wary a lot from one prompt to the next. An experiment to see how WD14 captions work with SDXL. The LoRa file can also be merged with the base model to create a checkpoint. Recommended negative keywords: blur, noise, photo. Adding "sharp, clear" as positive keywords can also work. I sometimes use a sampler that ends with "PeX", that stands for polyexponential noise schedule. It can be activated in A1111 in the settings.

Trigger tokens: sdgen / sdgen 1girl / sdgen anime style (or some combination thereof)

Clip Skip recommended.