Subtitle: Some tips on recreating Annaliese in my original intent, and how to change the style to your taste (maybe...with a lot of caveats...if you're not expecting too much).
When I first created the character LoRA (https://civitai.com/models/2061758?modelVersionId=2333090), I was gutted that it (seemingly) came out horribly. Eventually I got out of that deep dark depressive feeling of failure and gave it a good hard look, both here on Civitai, and on other image generation sites where I've also uploaded it. So here's what I've learned so far.
It's a realistic to semirealistic style
You're never going to completely escape that, even if you use anime-based checkpoints and dump a bunch of anime style LoRAs on top of it. But it can be nudged toward a more cartoonish to 3D effect rather easily. The idea that it is primarily a realistic style should have been obvious to me, thinking back to over a year ago when I first generated the images that were used in training in realistic Pony checkpoints like DucHaiten Pony Real and Cyberrealistic Pony. Those original images were also heavily stylized, using models like Moriimee and SPO-SDXL-4K (that one in particular was a game-changer for me back in the day and I used it on everything). I guess my point is, being super new to the "creator" thing, I thought that somehow the AI would magically transform realistic training data into whatever style the end user wants, effortlessly. How wrong I was! All this to say, you're definitely not going to make an anime version of Annaliese, and I'm perfectly okay with that...sorry anime lovers! 🤭
Tip #1: Keep your CFG LOW, steps high-ish
Again, I guess I shouldn't be surprised by that, the training images were all generated with low CFG to begin with. I always loved the muted colors and contrast that a low CFG offers. That being said, I recommend keeping your CFG down to 1.5, maybe creep up to 2, but certainly don't approach CFG:4, it gets really weird at that point...distorted blurry eyes, way over-saturated and dark colors, a braided ponytail that splits into two for some reason...odd stuff happens lol. CFG:1.5 and around 40 steps with a realistic to semirealistic model will pretty much automatically recreate my original intent for the character.
Tip #2: Lower the strength
The very first thing I noticed when I first created this LoRA is that at full strength it does not play well with very many models (I don't even remember which models I did get it to work with and can't be arsed to look, but it wasn't many). Lowering the LoRA strength to 0.7 really is the sweet spot. At full strength, you'll get a lot of the same problems as a higher CFG that I mentioned under Tip #1, and it absolutely failed to generate anything usable in most checkpoints. At 0.7 it opens itself up to being usable across a much wider range of checkpoints.
Subtip #2a: Euler A or DPM++ 2M Karras
Take your pick. I like them both in different ways.
Tip #3: Prompting
If you want any chance of getting something other than realistic/semirealistic, prompting is key. Even using an anime or cartoon checkpoint, even using anime or cartoon style LoRAs, you're going to have to specify in your prompt. western_cartoon or anime along with maybe some artist:x tags in the positive prompt and realistic, semirealistic in the negative will help...some at least.
Tip #4: Just play with it
If there's enough interest, I'll keep going with my plans for a V2 that will (hopefully) be much more user friendly. But in the mean time, hopefully the tips above will help you get some use out of this more contrary version. Experiment. Use it across multiple checkpoints, play with adding style LoRAs and see what you can come up with. Who knows, if there's enough interest, I might have an Annaliese bounty as I did with my other character, Ophelia. With the very generous tips/contributions I've received from some members lately (y'all know who you are), an Annaliese bounty will likely have a bigger payout than my last bounty.