Here is V1 of my WildSwanku style for Illustrious XL models! This is using the same dataset as my WildSwanku Pony V2 model. I call it toon-ish because it's more anime than western cartoon, at least with the way I use it. It can make subtle nudges toward my toon-ish style if used with no special prompt terms. If you want outputs to look as distinct as mine, check the recommended settings below!
I am very impressed with Illustrious XL's ability to understand complex prompts, and it is better than Pony for many concepts I enjoy. It has become my main model, and I use the Holy Mix merge in particular.
Recommended settings:
Positive prompt: masterpiece, best quality, toon style[0.8 weight], flat shading, {prompt}
Negative prompt: worst quality, low quality, lowres, jpeg artifacts, bad anatomy, bad hands, watermark, big eyes, {prompt}
LoRA weight: 0.75
Scheduler: DPM++ 3M Karras
Steps: 30
CFG Scale: 7.5
Prompt notes:
Weight terms in whatever way is appropriate for your platform
"toon style" and "flat shading" get it much closer to my original style, but can introduce strong cartoon influence. This can be mostly countered by using "big eyes" in the negative prompt.
Testing Update #3:
I found a way to get much closer to my original style!
Along with the default prompt from Holy Mix/Illustrious, I add "toon style" and "flat shading". I experiment with weight of "toon style" between 0.8 and 1, and LoRA weight between 0.75 and 1. Example images will be posted soon and I may replace the original post as this is much closer to the intended style.
Testing Update #2:
I don't think the Illustrious default settings (24 steps, Euler A, CFG 6) are doing us any favors with this LoRA. I have been trying DPM++ 3M Karras and DPM++ 2M SDE Karras with 30 steps and 7.5 CFG and getting good results for a single subject. Multi-subject still comes with issues, particularly on the eyes, but they're easily fixed with inpainting. Bad eyes when they aren't in focus/up close are a common problem across many models.
Testing Update #1:
Illustrious responds well to the tags "toon style" and "flat shading". It can almost reach my style simply by using those tags. So, using those tags and lowering the weight of the LoRA works well and helps alleviate some of the quality issues. Apparently the on-site trainer did not keep any versions of the LoRA except the one I published, so all the earlier epochs are gone. I may retrain this model at some point to see if less epochs worked better.