Should work on any anime model that uses Danbooru tags depending on how stylized it is. Takes a lot more wrangling then I'd like, but this is the best I've gotten for this style at the moment so I figured I'd upload it at least. If anyone else want's to give it a go, please do. I'm kind of sick of trying to train and experiment with this style over and over again (25th and best attempt), and as much as I love Yuji Himukai's EO art and would love to be able to imitate it with SD, I'd rather experiment on other things for training that aren't as hyper stylized for now. Used only Etrain Odyssey portraits (NPC and Player) for training data at 512 resolution, and not even that many after purging many problematic images that made the end results worse. This causes some problems I believe, such as armor being a mish-mash of everything and being mostly a mess. You can get some decent armor at times, but not often enough to be consistent. Leave the weight at 1 to start and lower or raise it from there. Put White Background and Simple Background in the negatives, helps deal with background issues, but not always. It also doesn't like to do hands, so that will take a lot more manual work to get decent looking hands. Also wouldn't recommend latent upscale too much, as it seems to weaken the style heavily. Use img2img non latent upscaling instead. Raw outputs look pretty bad on their own. Anyways, short of it is, this LoRA is a mess but I figured I'd share it anyway just for the sake of it. Sorry for the ramble-like nature of this description, just kind of done with this for now. |