Hello! This is my 3rd Lora and of my favorite character in Magic The Gathering. As far as everything goes she is extremely consistent and reliable. The only errors are occasionally in the eyes, and creating the spear.
All my art is using it at .6-1.0
Description
FAQ
Comments (6)
This could really use class images. The trigger words listed are not enough to properly generate the character, and require additional tags listed in the preview images in order to get something even relatively cohesive. Said trigger words such as "long hair" and "white hair" shouldn't be used whatsoever to help generate the character. It feels like this was trained on a relatively small dataset, and uses captions that would be more appropriate in a large dataset. You might as well not even include the Avacyn trigger name due to how little weight it has in this model. Not a fan of this currently, but I look forward to this if it is updated.
You are correct it was only trained on about 20 images. There unfortunately is not enough images of her to get a large variety. My tags are auto generated from the dataset maker I use so I just grab a few. And throw them in. My images however have been pretty consistent with the few tags I do use however.
@RedRidingHood You would probably be able to generate a class image without any of the extra details of the additional tags, which would allow your trigger name to override the pretraining it has. If you removed words like 1girl, long hair, white hair, and the like from the captions, those details would then be adopted with the trigger name, allowing you to not only use fewer trigger words, but get a more accurate result as well. Since you're having trouble with the eyes, the best thing you can do is create small crop closeups of said eyes and caption them appropriately in order to help train them better. Using a dataset maker is fine, but using them unedited by default is a bad idea, because not only can they be incorrect, but words like the 3 I mentioned earlier rob the bot of details that are inherent to the character that it needs to learn. Using a classifier token even without any class images will greatly improve the trained end result, regardless of the source ckpt.
@AxizP I don't really understand how to do that but I will look into it. I can't train in SD as I run on Amd. Thankyou!
@RedRidingHood If you're using Kohya_ss you can just label your character concept folder X_Avacyn 1girl, which declares the concept token classifier. In the class images folder directory, you make a folder called X_1girl, the number of repeats being really low cause of a small dataset, something like 1 or 2, then you just put in 1 or 2 class images of the character with either closeups or just details that are purely inherent to the character, like without wings or something.
@RedRidingHood Oh and I don't know precisely how it would work for you if you are using colab, but I am pretty sure there's a way to use class regularization images in a very similar way. I can't help further there though since I've never used it.