LORA that gives Hunter X Hunter 2011 anime style along with some charcaters, it does mainly gon_freecss, killua_zoldyck, Hisoka, Kurapika, it possibly could do Leorio and Netero with some good prompting but i didn't train it on them much.
Although good prompting (describe appearance, clothes etc) is best even for the best trained character.
it was trained on Anything 4.5, but it should work relatively well on most anime models based on my experiments, my preferred weight is 0.5-0.6, might differ between models.
Have fun and please share your creations here !
maybe one day when SD animation capabilities are good enough, one of you will continue the anime haha
PS: The trigger word isn't entirely necessary, and sometimes you might actually get better results without it, so experiment and see how it goes.
Also, i found if you're doing many characters it's best to do low steps (12-18) then highres fix, otherwise the features will get jumbled up and you'll get clones. i don't know if this is specific of this lora,, or it's an sd thing in general.
Description
FAQ
Comments (18)
This is very good, the author can update the Naruto anime lora?
thanks, what do you mean the author can update the Naruto anime lora? you want the naruto one as a lora ? it's not actually mine, it was by lambdalabs i only turned the diffusers to a safetensors, although check with @Kappa_Neuro if he's interested in extracting the LORA from the model.
@Aili Yes, of course
@Kappa_Neuro you're a legend
@Aili No, I was hoping that the author you update Naruto's role
@tiger078 i have to be honest with you mate, i have no idea what you're saying
Very cool! How do you train multiple characters and styles in one LORA?
in this one i manually captioned the dataset so it did take time, although i think you can train multiple concepts at once, which is probably the easier method, but i haven't tried it yet.
@Aili May I ask that how many images you were using for each character?
I don't understand the tag is < lora: hxh2011:1> or <lora:hunterXHunterLORAStyle_1:1>?
in my examples the model name is simply set as hxh2011, when uploaded it gets renamed to the post title, use whatever auto puts in (which is the file name you have, in your case probably the later), but the trigger word in the prompt is hxh2011
it captures the face really well, but it detects Killua as a girl when i make him topless, any idea how to fix this?
(1girl, female, breasts, woman:1.5) in negative, (1boy, male:1.3) in positive maybe ? and try topless_male instead of just topless.
never tried this myself, but could work.
I gotta try out this style, love the show (will probably do a Kurapika LoRA)! Also can you share how many images you used in the dataset. I read somewhere that 800 was good for style, but idk if that was overkill or not. Also like 3000-4000 steps?
i can't quite remember how many images, but i think 250 images maybe, for the steps, this is old and it was overtrained a bit that's why i say the weight should be low, i think this was something like 50 epochs, but generally speaking, 15-30 epochs usually works well, I'm not sure how many steps that is.
edit: it was actually only 84 images, the 250 was another lora i didn't publish.
the point is you don't need 800, although it might help. i do 10 repeats.
@Aili Gotcha thanks
@novelProphet i just want to mention that the data was manually captioned, i auto captioned then edited the captions and added character names, thats how it works for characters as well as style.
@Aili Yeah same. I think I should have enough images of the official style from the 6 Young Justice ones I’ve already captioned, but will prob grab more.
One thing I’ll likely need to do is re-add the physical attributes (hair/eye/skin color) for characters. Since I wanted those to be directly associated with the character name when making the individual LoRAs.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.