Please, check the description, there is a little bit of useful information.
Inеtroduction:
Tsurime eyes - a particular style of drawing eyes in anime where the outer corners point upward, ending in a distinct point, used in contrast to the traditional Moe Moe soft drooping style,
Tsurime symbolizes strong will, arrogance, or pride in a character. Most noticeable in female designs (since it plays against the stereotype) and is commonly an attribute of tsunderes, kuuderes and particularly badass action girls.
Compatibility (Pony):
Humans: great
Anthros: great
Ponies: good
Multiple characters: good
Character LoRAs; good
Unique characters: varies
Recommendations:
1.0 emphasis
Clip Skip 2 - SD 1.5
Clip Skip 1 - PonyDiffusion
AutismMix - Pony for generations
tsurimeas an activation tag
Credits:
Trained on NAI Anime Full.
Dataset of 1000 pictures (v1). 10 Epochs. Clip Skip - 2.
Trained on PonyDiffusion v6 XL
Dataset of 150 pictures (v1), 20 epochs, Clip Skip - 1
Really big thanks to:
Me, for not being lazy
Some good folks from Discord server
You, for reading, downloading, liking and reviewing this LoRA
Description
Initial release
FAQ
Comments (19)
Looks great. Was the dataset scraped from booru sites? How much filtering did you do?
Hi. 1000 handpicked images from danbooru. ~600 pages of searching.
i've just used tsurime -animated for filtering out unnecessary animated images.
@TroubleDarkness Thank you! I'm trying to do another eye lora and wasn't sure how picky I needed to be to be effective on such a small part of the image. 600 pages is 12000 images so you manually filtered out like 90%, so quite a bit lol.
For what it's worth, gelbooru.com has support for searching more than 2 tags at a time (can do like -speech_bubble -translated -translation_request for text), but the tag quality is lower.
@ManaMomo yeah, i know about gelbooru. But there is a limitation for browsing depth. ~460 pages, i think, or some more. Actually, if you're doing concept LoRA, ~100 images would be enough. ~300 - great. More images - even greater.
For example, good settings for 300 images in dataset - lr = 5e-4, unet = 5e-4, text encoder = 1/2 unet, 10 epochs. For 1000 images in dataset - lr = 1e-4, unet = 1e-4, text encoder = 1/2 unet, 10 epochs. If you're doing style LoRA - lr = 1e-4 / 2e-4, unet = 1e-4, 2e-4, te = 1/2 unet for 300 images, 20 epochs. I'm using this settings and pretty happy with that.
@TroubleDarkness I was at 100, but I think the quality wasn't good enough because weird stuff was bleeding through before the eye effect was strong enough. I'm retrying now with almost 400 after trimming images with strong styles.
For gelbooru, because you can do so many tags, I've been searching cowboy_shot, full_body, portrait, close_up, and negative-all-those separately, so none of the searches go that deep. Haven't gone through 12000 images to see if it becomes a problem though lol
Thanks, I'll give those settings a try if I can't get DAdapAdam to work out. It's weird. I was getting okayish results but it wasn't training the trigger word at all. All the effect was controlled by the lora weight. I'm guessing that's the difference between text encoder and UNET learning?
What model are you using? The sample images all look amazing!
Hi, it's still unreleased anime version of DIvineEleganceMix. I'm preparing page for this right now.
@TroubleDarkness Please let me know when it's out! I'd love to use it :D I already love DivineElegance :)
@madaraxuchiha88 Hi, DivineAnime mix is out https://civitai.com/models/95587/divineanimemix
Closed eyes concept would be cool, I believe. Many models are not good with accurate closed eyes prompt. Can you try it?
Hi, actually, there is way more concepts, that models can't understand - tsurime, tareme, closed eyes, half-closed eyes, and even more. There is always would be a place for this kind of LoRA. If i'll have more time - i'll do it something for that.
WORK WELL WITHOUT ANY UNDESIRED ADDITIONAL EFFECTS. GREAT WORK!
For some reason, got notification about your commentary just right now. Thanks for commenting! Good to know it works well you ya'.
I need this for pony and XL models, are you going to release a version for those models?? Please say yes :(
Hi, i thought it was already implemented in PonyDiffusion model. At least, in basic one. Also, i'm pretty sure most of its concepts should work on AutismMix. But if you're saying, it's not working there, i'll look into it. It's actually pretty easy to build a tsurime eyes LoRA for PonyDiffusion.
Also, there is a lot of LoRAs on CivitAI. Did you checked the tsurime eyes LoRA availability? UPD: Well, i've checked. There is no tsurime eyes LoRA for PonyDiffusion. At least i've not seen any while scrolling through "tsurime" request.
Hello again, i've trained Tsurime eyes LoRA for PonyDiffusion. Needs a little bit of testing and previews, of course. On first ~100 gens it works good, so, i'll release it today (UTC+3). As only all my commissions will be completed, i'll retrain it with better tagging to remove any possible issues from autotagging.
I mostly use Animesh for what I need that is my own OC and Animesh creator said that he is giving priority to XL model because he said Civitai only let user use certain number of models creators and not all of them he said he has all his models for generation, but civitai don't, so for XL there is not Tsurime model that I can use with Animesh XL, thanks man I'm going to test it and see
@Bram005 Just to let you know, I've trained a LoRA for PonyDiffusion-based models. It would be incompatible with any regular XL-based model, even if it's an anime finetune. And of course, there is no chance i will ever train a LoRA for such models because it is the same dead end as SD 1.5 models.
@Bram005 Released tsurime eyes LoRA for PonyDiffusion-based models.