While training on SDXL (or PDXL) takes more time, it seems to have higher chance (compared to sd1.5) to succeed at making something useful with very limited dataset. So I tried with a light novel waifu who has only a few official illustrations.
Description
Made with artificial dataset created with initial version.
darkmel for character, darkmel-armor for the default outfit. Also, headgear, hairband, halo, wings can be activated or not. Works well with 0.6-0.8 strength, I recommend using 0.8 for txt2img and 0.6 for upscaling the most.
Some details (shape of hairband, feathers on pauldron, halo being black) might require higher strength to get consistently.