Recommended weight 0.6-0.8
Works fine with ReV Animated and AOM3
Sometimes images can be generated without some details, so I recommend using tokens of what you want to see such as tail, hooves, horns, glowing eyes, blue skin.
Description
learning
Treined on AOM3
Tool: Kohya's GUI
70img * 100steps
Dataset https://mega.nz/folder/wEEWwZbT#rN-USn_su4ZdMrV9pxaFNw
FAQ
Comments (9)
Hey! This lora looks fantastic! Very coherent character. How did you train it? How you made the tail to appear so consistant? Did you use images with the character from the back view? I'd gladly see the initial arts that you used for training if it's not a secret.
Ok, later i will post the dataset somewhere and post the link in these comments
@Chik thanks! Much appreciated!
@Chik Wow! Very helpful! So you took any possible image from every possible angle. I thought you need to use strict portrait ones. An you cut them out on some images and left with background on others. Could you please explain your thoughts on how you choose images and descriptions?
@drakmour you can use this guide https://rentry.org/2chAI_LoRA_Dreambooth_guide_english
Usually I select 50-75 images, do not recommend more than 75-80, because lora can overtrain.
Images should be free of unnecessary details that you would not like to see on the art. I also try to remove the background on images if possible, but I'm too lazy to remove with Photoshop, so I use skytnt/anime-remove-background.
Usually I just choose images that I like from different angles and even different styles.
I usually download images along with descriptions from boru boards using the grabber program. Then I edit the description to include only the most important tokens.
@Chik Ok, thanks for the explanation!
ααнн ѕмαrт.
ι dιdɴт eveɴ тнιɴĸ oғ тнαт.
ι wαѕ woɴderιɴɢ αвoυт тнe eyeѕ ιɴ ѕαмple ιмɢѕ
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.