This flexible asari LoRA for SDXL is not strongly biased towards the specific Mass Effect character. That made gathering datasets difficult, but it paid off.

Default settings of LoRA weight 0.8 and CFG 5-7 are recommended.

It's trained on 80 images of varying aspect ratios, though more often portrait orientation than landscape. Many had to be upscaled, however.

The captions follow this format: "An asari woman in/wearing ... and ..." and then a list of captions.

• pale/dark/<blank> blue/purple/teal/green skin

• faint/black/white/red/yellow eyebrow/full/crest/lip markings

• comic style, realistic, digital painting, digital drawing

• fullbody, portrait, halfbody, closeup

Multiple colors in a prompt are always a recipe for disaster, and the flexibility means that a purple asari in a green dress might end up as a green asari in a purple dress.

Limitations:

• Realism is harder than a drawn/comic style.

• Skin color below the neck ends mostly right, but sometimes it fails, especially if you're going for realism.

• Like with bad hands, five or eight crests are common issues.

Settings comparisons for LoRA size and alpha can be found here: https://huggingface.co/vmaple/sdxl-asari-lora/tree/main