>>26536
Not gonna lie the hat being missing is probably my fault.
I trained the dataset to consider the clothes *separate* from the character, by tagging the clothes in the image so that the lora trained less on it.
The bonnet would've appeared more consistently had I tagged it in my generation prompts. Similarly, I was getting red hair and white sclera until I forced it to be correct with my prompts. The winged arms were also missing until I tagged it.
Custom loras being easy to make? Depends on your personal patience and autism level. I can probably make a dataset for one in at most an hour or two, but 15-30 minutes if I don't really care about scrutinizing the tags. After that, it depends on how good your GPU is. You could maybe pull by on an 11gb VRAM gpu and low quality LORA settings, but it would take several hours.
These are generally the programs I use.
LORA Training Reference Guide:
https://rentry.co/qab9g
Kohya SS GUI (Lets you train LORAs with a GUI interface and config options):
https://github.com/bmaltais/kohya_ss
Taggui (Lets you prune or batch add tags to images, remove duplicate tags, and auto-caption images reliably with something like waifu diffusion. You shouldn't rely on it fully, but it does most of the work):
https://github.com/jhc13/taggui
Message too long. Click
here
to view full text.