A lora based on my favourite gacha game Last Origin
works fine with anything between 0.8-1.0
Fixes
trained on a black background instead of white background to achieve a much better style
Description
Completely retrained on jpg images with a white background to avoid any png artefacts.
ALL IMAGES ARE RAW OUTPUTS USING HIRESFIX
R-ESRGAN_4X+ UPSCALED BY 2X DENOISE 0.5
Works best at 0.8-1.0 strength
Adding in thick outlines,thick lineart helps
fullbody + shoetype (fullbody,high heels)/(fullbody,trainers) helps with recreating the gacha look
FAQ
Comments (4)
I've been planning on making a LoRA like this for a different game, is there any reason why you chose white background over black background? Also, how much so did you prune the dataset tags?
a white background is just the default jpg look + it adds better contrast than black in my opinion and making it black would take so much more effort for 440 images . All I did was convert the pngs -> jpg. As for the tags I didnt prune them at all. All I did was add last origin, to the start of each txt file.
@mnemosynekotorin I see, thanks!
in my experience, if you want to rip the drawing from the background, white or grey makes the job much easier than having the background black.