Issues with the Aesthetic Model
Every image generated with aesthetic just looks like Pony XL, regardless of whether you put score tags or not.
The "anime screenshot" tag isn't adhered to at all unless you increase the weight to 2.
Quality and anatomy suffers significantly in some images.
Those issues don't happen with v1.0b, or at least to a lesser extent.
It would be nice if you could ship a "b version" for the next aesthetic models as well, or maybe just look into fixing that entirely.
v1.0b:
v1.1:
trash style
I’ve been sticking with Base. Tried the 1.1 version today, and noticed the outputs have a more pronounced AI aesthetic.
It's at least better than v1.0 since there's no noise issues anymore, but yeah I don't really see a good reason to use it over v1.0b or base. Though I'm pretty sure v1.1 is just an adjustment of the Loras merged into it, so v1.0b is the same finetune.