This is probably my favorite anime in terms of art and style it's absolutely stunning so I made this model to capture the style. The model was trained on 564 screen caps of the show in blu-ray edition to make the quality as high as possible. As always have fun generating!
If you want to support me financially, get some exclusive models or request a custom LoRa you can do so here: KoFi
Usage:
Trigger word: hunt3rd
Strength: I use 0.8-1 but feel free to experiment.
Description
FAQ
Comments (16)
this and the berserk lora are probably one of the most authentic ilxl style loras, well done.
would you be down to try a somewhat similar style for Escaflowne? I can assist you with some buzz.
@yourneighbourRiz Thank you very much means a lot! Good idea! I'm a sucker for old anime and I actually have Escaflowne on a hard drive just have to find it, I mostly train on my own hardware but buzz is greatly appreciated I'll start making the dataset today if I have time otherwise tomorrow for sure!
@NeoWave dope, didn't think many people knew it so it's cool that you actually have the movie as well lol. (The show is also great, but I think the aesthetic of the movie is just.. unmatched/unobtainable by today's standards). Oh not sure how you're tagging but if there is a way to make melefs (giant mechs) or machinery also somewhat promptable it would honestly be greatly appreciated. But I understand if that's too much to ask, take your time dude 👍
@yourneighbourRiz Here I uploaded the model it was trained on the movie and for prompting the mechs just use mecha, armor, and whatever you want to add should work but the scenes with the mechs are mostly cropped in the movie so there may be some issues.
@NeoWave awesome my guy, splendid work. The scenery shots you do are so nice
if you need more buzz it might need a bit since I'm accumulating them daily, but to be fair looking at your lora you're far out-training the website trainer since I assume you're using a lot of steps which the site trainer can't use within a reasonable amount of buzz
@yourneighbourRiz Thanks for the compliment. Buzz is appreciated thanks, but you don't need to if you don't have any it's okay. I do use a lot of steps yes because the datasets are large so it takes a while to stick, and the highest batch size I go is 2 anything higher doesn't work that well at least in my tests
@NeoWave indeed. I think illustrious needs far more steps for styles in general since the models are overloaded with their own range of style data. I can go like 2000steps on the site trainer which seemed like the bare minimum for a style to stick, and that one already requires extra buzz lol.. so it's really cool that you can and want to do these huge authentic styles. Cheers
@NeoWave ah kinda wanted to ask, do you recommend a specific range and model for wdtagger? I use between 2.5-3.5 but idk, perhaps it's preferable to tag way more for styles
@yourneighbourRiz I use JoyCaptionAlpha2 for tagging my images but there is still a lot of manual work with tags as always. You can get it on huggingface but it is a little demanding on the hardware side so you need a decent gpu. If you want something that's quick and gets the job done somewhat I'd recommend SmilingWolf/wd-swinv2-tagger-v3 at 0.38-0.40 probability.
@NeoWave ah okay, guess going past 3.5 might be worth a try. Much appreciated
😍 👍🏻
Flux Version?? Pretty Please :)
@QGC I'll think about it. Flux versions cost a lot of compute, and it doesn't handle large datasets like this that well. Also I'll have to retag all 564 images for Flux which would take me probably a couple of days of just tagging images.
Wow... seems like a lot of work :( I was thinking this style would mix very well with Aeon Flux (as those animations had some of the same artist working on them). Anyway thanks for reply.
@QGC This is the main reason I don't do much Flux loras, It takes a lot more time and resources, I'd be glad to do more of them but I can only manage a few, it usually takes me more than a week to make one so I'm really picky on what stuff I make a Flux lora. Flux also doesn't really like to learn an animeish style for some reason. But maybe I'll do a flux version in the future, nothing's off the table.
Awesome style!
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.