This style lora is trained on images by KYHU, the old pen name of artist IAHFY. The dataset mainly consists of images of Korra and Asami, but it also works good for other characters. The images have monochrome or low detailed backgrounds. So this lora is best suited for generating characters without detailed backgrounds for Illustrious. Since Anima seems to be better at generating detailed background, you can try if it does better there, but I haven't really tested it. Works nice with additional style tags like "monochrome", "sepia" or "sketch". All versions are trained with the same dataset and captions. If you want to support the artist, you can find options here.
trigger word: kyhustyle
additional style tags: monochrome, sepia, lineart, sketch
character tags: asami sato, korra
Feel free to comment, if you have something to say about the lora and share your images on the model page. Sharing your images on the model page gives the creator a small blue buzz bonus. But more important I like to see, what people do with my LoRas.
If you like the LoRa please have a look at the showcase pics and leave a reaction.
Description
This is trained with anima-preview1 as Base model, so it won't work as good with preview2 or preview3 version. For the showcase images of this version I tried to reproduce the images of the illustrious version. It worked good for longer prompts, but shorter prompts didn't come out as well, since Anima isn't made for them. So I had to expand the prompts in that case. Sometimes they just came out different.
I used Euler A sampler adn Beta scheduler for the showcase images in combination with positive and negative quality tags, since it resembles the illustrious style the best in my test. You can try using other samplers like ER SDE though and omitting quality tags gives you a style closer to the original artist's style.
Also while testing I observed prompt adherence with Anima improves when you use the recommended prompt structure, i.e.: quality tags, rating tags, people number tags (1girl,1boy,1other,...), character tags, series tags, artist tags, all other tags. Overall it seems that Anima has better prompt adherence than illustrious, but may be bad at mixing concepts, hopefully that can be improved with a newer version of Anima.
Have not really tested this LoRa with other checkpoints than anima-preview, so style can look different, if you use another one.
FAQ
Comments (3)
So two days late, here is my first anima LoRa :D
@CitronLegacy
I'm 2 months late... 😅 This looks great!
I just tested out all your Anima loras and they are looking good. 💯
I'm curious to see how your Korra Ultimate lora will work on Anima but I think you said you did a crazy long training session for it? So maybe you'll need a different strategy for Anima?
@CitronLegacy Thank you for trying!
I guess they work, but the styles don't seem to do much with finetuned checkpoints/merges.
Also the compatibility between different preview versions is not great (Nami's outfit looks wrong, but that's probably also bc of the schedulers).
So I won't be trying to make Korra Ultimate for anima yet. Maybe when full anima is published. Yeah, I think I need to do several training runs then. Also I need to verify multi-outfit loras work as good with anima as with illustrious.
Are your checkpoints now available for generation? I need to try them a bit then :D