Finana Ryugu | Character Sets #005

Composition is something I will put more thought in to in future, a big issue with that is the flexibility of the LoRA, the more dynamic you try to prompt the poses the more likely you'll start to generate bad anatomy and extra limbs etc. I do try to minimize artifacts especially hands as its something you can pick out very quickly, but other things like incorrect clothing and wrong details I don't really worry about unless its something that doesn't really fit the style. For example Finana's bodysuit wasn't really properly trained in to this dataset and any attempts to make it look more like a leotard/bodystocking turned it in to a swimsuit with features of her dress. When I make my own LoRA I focus a lot on face/hair details in my datasets so the flexibility with outfits is high, but when you want to train detailed elements of an outfit you lose flexibility with all outfits, in my experience at least. I can use things like controlnet to create poses but I tend to not use it as much as it has a heavy influence on the overall generation, at least in the times I've used it. Maybe I can look in to other methods with controlnet to generate some more varied poses and angles, at least use it at a low weight just to influence the rough composition without altering the style or detail.

I will definitely prefer quality, or in fact, in case of AI with less artifacts, as they are really making attention and ruining works. But I wish there would more variety to the poses. Right now all of them frontal with either standing or sitting. If you look at traditional artists, many satisfaction that they bringing is poses and camera work. For comparison take a look at this work: https://www.pixiv.net/en/artworks/110736543 She is tilted here 45° rather than being straight up and if you look at adjacent work: https://www.pixiv.net/en/artworks/110581340 there is pose from side on her 4. Some other works are horizontal (landscape orientation) with a character in lying pose. Anyway, stuff like that make variety more meaningful. Can't say how hard it would be to generate something like that, but if you looking for any improvement, just look how around traditional artists draw their works and then compare.

XCanG