Raindrop 2.0 vs. 1.0
A quick comparison between v1.0 and v2.0 of the Raindrop finetune of Illustrious-XL.
Four pairs of txt2img sketches, using the same prompt, settings, and seed. The prompts are what I would use with v1.0.
In v2.0, at least with my prompts:
The style is closer to western comic books than anime. This is a nice alternative style to have.
Characters appear older. This is at least partly due to the more realistic style.
Backgrounds tend to burn, even at CFG 3. (Weird, didn't happen in the model showcase - more on this below.)
I'll also take the opportunity to write down these points on the Illustrious-XL series (such as Raindrop) vs. the previous popular PonyXL series (such as AutismMix):
To get a cute next-door face, you no longer have to call your characters "ugly", and if you do, you'll get exactly what you prompted for. But it still sometimes helps to have "pretty, beautiful, attractive" in the negative, to combat over-averaging. (EDIT: maybe I spoke too soon; see at the bottom.) (The AI models seem to believe the old adage that "average(d) is beautiful" - the next-door look, in contrast, is distinctive. Compare the classic US air force study from the 1950s on pilot seats in airplanes. No actual pilot has average proportions [1]. Likewise, no one has the average face.)
Great backgrounds, like AOM3 back in the SD 1.5 days.
Great hands, most of the time. In simple scenes, usually no need to inpaint them, and even at worst, needs 1-2 inpaint rerolls at denoise 0.5.
Very easy to inpaint with. No need for a separate inpainting checkpoint or ControlNet. Can go up to denoise 0.9 (!) without destroying the image (given enough surrounding context in Only masked inpaint mode). Discontinuities start to appear around 0.75, but these can usually be fixed by another round of inpainting over the seams at denoise 0.5.
Something that hasn't changed: generalization capabilities are lacking. It's still easy to find combinations of concepts that are out-of-distribution for the AI, such as a character cosplaying a specific planet. Perhaps this kind of thing needs a bigger model.
The issue with the backgrounds that I ran into with Raindrop v2.0 is weird. As mentioned, there are no issues with the backgrounds in the v2.0 showcase. Also in my own usage, I've used the previous v1.0 at CFG 5 with no issues. Possible reasons that come to mind:
Something in my prompting that v2.0 doesn't like? It's the same base model, so the quality triggers shouldn't have changed.
v2.0 works better at CFG 7, which was used in the showcase? Doesn't seem likely, because higher CFG values typically burn images more easily.
Sampling? The showcase used Euler A with 30 steps, whereas I've used DPM++ 2M with the SGM Uniform scheduler, with 20 steps. This combination has produced the best results with all other SDXL finetunes I've tried, including the previous Raindrop v1.0.
Anyway, I'm posting this as-is for now - I might return to this later.
EDIT: As for the backgrounds, seems to have been a cursed session. Hasn't happened again.
EDIT: To get a next-door look, calling a character "ugly" still works in v2.0, but the strength for this needs to be lowered to 0.8 or so.