All images were generated using abyssorangemixs-2-Sfw
Description
-first version
FAQ
Comments (9)
Gracias!
Come on vamanos, everybody let's go...!
Your lora is really good, did you have any tips or guide that you follow?
I've been trying to train my own lora but the face and eyes look very weird (kinda ugly or disfigured) and my lora with weight 0.7-1 always break the image (maybe overcooked?)
I'm not very knowledgeable about this, I only started recently
I don't do anything particular, I take 25-30 image on google, most art image from the show or official artwork and some art fan art (very few) and try to only take good quality images. Try not having a lot going on in the background, you can remove the background easily in a batch with the following tool.
First, I enlarge them if they are not square (only if the background is a solid color, like grey or white, or remove the background of the image to enlarge them) using: https: // imageresizer.com
secondly , if they are not 512x512, I crop them using: www.birme.net Next, I create the dataset using:
https://colab.research.google.com/github/hollowstrawberry/kohya-colab/blob/main/Dataset_Maker.ipynb
dataset_Maker Parameters:
tag_threshold: to 45
Note: some time I manually edit the tags in the .txt files in the drive to add tags that I think are missing, don't add too much
Lora_Training Parameters:
Train the lora using the link to the Lora training on the same page.
You don't need to change a lot except The followings parameter to the recommended value
network_dim
network_alpha
For the eyes, I don't your specific problem, but you should take a bunch of portrait of your character in good quality and again if there is a lot of effect in the background try removing it and replace it with a solid color.
or just use png image.
@TheGooder thank you for your reply, seriously that is very long text you type so i really wanna say thank you.
As for the lora training, i will try to change the background and hope for better results. I already train for 11 times already and for some reason the lora with the 30 image dataset is better 100. I train with 2000 - 4000 steps and 2-5 repeats but still can't find the best results.
@chronus
Yeah, I don't know why training 30-40 is better than 100. I never tried anyway, you should try to ask some other Lora maker on the website to see if they know your problems
Maybe the number of repetitions is the key. The more repetition the more the model can understand the images and replicate them. Again, I don't really understand what I am talking about.
also, i am very surprise you can fin 2000-4000 image of you character, do you take them yourself?
@TheGooder no, i only found around 100 image or something mostly on pixiv. The 2000-4000 i'm talking about is number of steps not number of image.
As for the training progress, i still don't know what I did wrong. How many steps/epoch and learning rate did you set for your training?
@chronus
Ha ok, sorry I didn't read it correctly.
most of the time I use 10 epochs (2000 step) with 24-40 image
Learning_rate:
5e-4
I don't really change the settings because I am not very knowledgeable about them.
also, you can try using another model for your trainings.
i don´t know if this counts as indecent information, but after i got frustrated about my generated images (anime girl, got bigger (adult or teen), breasts and so on), i tried to "cheat" on body sizes. and the absolutely perfect combination of prompts were "chibi toddler dora" to trigger the EXACT dora in 512x512 images.
i will upload example images and what happens when you force dora to stop laughing (she started sweating like crazy in the generated image) or force her to not be in the focus (she gets angry and stops breathing (blue skin).
yeah this version is not very flexible and for your problem, I don't know what is wrong with the model but I hope the v2 will fix it