Naruto Lora Series: Uzumaki Naruto
火影忍者Lora系列：漩涡鸣人
Weight: 0.8（v1.0）、0.5（v2.0）
推荐权重：0.8（v1.0）、0.5（v2.0）
Description
FAQ
Comments (8)
I really like how your lora gets a pretty faithful naruto into the picture with just the word "Naruto".
I notice that his outfit can kinda go all over the place based on your examples. I looked at the lora metadata and it looks like you trained this all in one folder.
If you wanted to improve the results without too much more work, you could try separating out your Naruto pictures based on his outfit; try two separate folders where one is Naruto from the original series and the other is Naruto from Shippuden. Change your trigger words slightly for the Shippuden version, and with any luck you'll have a lora that can do both without mixing between the two.
I might be mistaken, maybe you only used pictures of Naruto from the original series and the outfit stuff has to do with your unmodified BLIP tags. If so, that's fine. The results are really strong as they are.
Thank you very much for your advice. Perhaps because I am not a very loyal fan of Naruto, I am only using it as a training object, so there may still be room for improvement in this Lora, as you suggested. Thank you very much for helping me.
Separating the folder doesn't do anything when using captions (only allows you to use a different number of repetitions), you just need to caption them accordingly (i.e. use a tag for each kind of thing you want it to distinguish between)
@eeeeeqeq Just speaking from experience on what I've had success with, and also the lora I have seen with different outfits often use different folders (though this may be based on a need to change number of repetitions). I'm by no means an expert, here, and if you have any stuff I could read I'd appreciate the info.
@Knots https://github.com/kohya-ss/sd-scripts/blob/main/docs/train_README-ja.md
"If you put a file with the same file name as the image and the extension .caption (you can change it in the option) in the training image and regularization image folders, the caption will be read from that file and learned as a prompt.
* The folder name (identifier class) will no longer be used for training those images."
@eeeeeqeq Having watched the kohya_ss console, I'm under the impression that having them in separate folders does something. It does appear as though it may influence, at least, the order in which the images are trained on? I will rtfm.
@Knots Probably, but the order shouldn't be relevant anyway (all images will be shuffled), what you can do by putting them in separate folders is use a different number of repetitions, which allows to balance training by using a different number of repetitions for specific groups of images, such as a low quality and high quality folder (with low and high number of repeats respectively)
My favorite ninja technique? The sexy jutsu, of course.