This lora was trained on 50-230 images from the classic beat-em-up MMO Dungeon & Fighter (Dungeon Fighter Online). Most of the data was taken from the most recent artwork, but some of the classic images are used as well. Please note, this is a style lora. You will have great difficulties trying to prompt the classes by name alone.
A great deal of the images from the data are very dynamic, so the generated images tend to take a lot of liberties as you gen. Full body gens are particularly unsavory. They look better when upscaled, like most things, but this lora sure loves to create pretty nonsense.
Nonetheless, it's 1girl,standing game is solid. Anime boobies, shiny, muscular man, it's all there. Enjoy!
Version 2.0 was trained on Noob 1.0 V-Pred.
Version 1.0 was trained on Based Pony.
To get the results you see in the images above, these are the settings I used.
Description
FAQ
Comments (5)
Dude this is looking so awesome!! Like a style i was looking for! Thanks for making it.
Also, you shoud add cropped images of some bodyparts like heads, faces, upped body , lover body, hands, as i saw it makes model work better.
and its also helps to make dataset bigger. If cropped images gets too small or blurry you can use opensource upscale program like this one https://github.com/upscayl/upscayl/releases/download/v2.10.0/upscayl-2.10.0-win.exe , remacri or balanced presets are best for artworks.
and your Vanillaware
model is awesome too! Hope i can help you make your models even better!
Can i have your dataset and train settings? Your lora performance is GREAT!
For training I use these settings: https://civitai.com/models/281404/lora-training-guide-anime-sdxl?modelVersionId=316795
In this case, I set the repeat to 1 with epoch of 60, saving at every 10 epoch. I use Kohya for training. The file link above doesn't set the caption extension to ".txt" as it should, just as a word of warning.
As for the data, most of the images were gathered from: https://wiki.dfo.world/view/Main_Page
I change the repeat, epoch count and data size with each new lora so it's always a case of trail and error. I look out for things like prompt comprehension, over/under baking and how well the generated images look like the source images. I hope this helps.