This is a LoRA model for Lily White from Touhou. After much, much testing and trying to get things working, I finally decided to start sharing LoRAs when I make them. Giving back to the community, or something? Mostly made in an attempt to learn how to train LoRAs, maybe I'll come back and improve on this at some point.

Key Notes:

• Use "lily_white", "fairy_wings", and maybe "[color] eyes" in your positive prompt for best results.

• Default should work with her traditional clothes without additional modifiers. Add clothing prompts, and maybe turn down the prompt strength a bit, if you want her wearing something else.

• Should work on any SD1.5-based model.

• The example images were upscaled with the following settings on the A1111 Hires. fix: Upscaler: R-ESRGAN 4x+ Anime 6B, Hi-Res Steps: 90, Denoising Strength: 0.2, Upscale by: 2.0

Detailed Notes:

• Trigger word - "lily_white" - without this, you might get a tiny difference, but the LoRA won't really do a whole lot it in as a prompt. I'll be entirely honest - I'm not sure, exactly, how much much using that as a trigger word in my dataset txt files messes with the results, given that it does sometimes work on its own as a prompt for a lot of models. If anyone has any good reasons for me specifically not to use that sort of thing, and instead use something like "l_w_LoRA" or something of the like, please let me know. Thank you!

• Training set included multiple eye colors - I've read that her eyes are Green, but each to their own. Put tags like 'green eyes', 'blue eyes', etc. if it isn't coming out how you like. The eye colors in the training set included mostly blue eyes, though some each of red and green, so those three are the most likely eye colors to show up.

• Her 'main outfit' was in basically every image in the training set, so it'll be the default - adding in additional clothing prompts usually overrides it at 1.0 prompt strength. However, sometimes it doesn't; putting the prompt weight around 0.7-0.8 has had pretty good results for changing the outfit while keeping everything else mostly the same. That said, there have been a few instances where it didn't add her main outfit at usual strength, without modifiers, especially when changing hair styles or specific parts of the outfit - I'd recommend "white dress, long sleeves, white hat, red bow" in your positive prompts, if you're having problems.

• Most of the training set had her blushing, so that's something of a default, unfortunately. That said, throwing "blush" into the negative prompt does help with it, if it's becoming a problem. I've seen it mess with some emotions, when you try to prompt for them. Sorry!

• Her wings should come through most of the time, but if they dont, adding 'wings' as a prompt usually fixes it. I swear it's in the data set, despite it not showing up much from my example images!

• I'd recommend having "detached sleeves" in the negative prompt, unless you like that. I've put probably too much effort trying to weed out the whole "detached sleeves" thing - where a lot of images using this LoRA come out with a hole in the outfit between her shoulders and arms. Unfortunately...I couldn't really fix it. Weird it's cropping up so much, given none of the images in the training set had it...but oh well.

• The vast majority of my testing has been at 512x512 resolution, other than the 540x860 resolution I used for making example images. I can't promise it'll work amazing on other resolutions. That said, the little testing I did at 540x860 looked pretty alright, so I'm sure it'll probably work for most reasonable resolutions.

• Haven't tested this with more than a few anime checkpoint models, as I mostly use the BOM sfw; from a little testing, it seems to work well on most of the other orangemix checkpoints and on Anythingv4.0, though as it was trained on SD1.5, it should work on anything based on that. Your results may vary. As an anime character, I've no idea how this will work with realistic models, but I'm sure it would be amusing/terrifying, heh.

Well, I could probably go on with notes, but it's already getting long. Luckily, CivitAI has the "show more/show less". If anyone has any recommendations for tags or anything else I should add here in the notes, please let me know! I'm not really sure what else really needs to go here.

Thanks, and let me know what you think!