An alternative version of (https://civarchive.com/models/361165/insertion-put-stuff-into-holes-sdxl).
Trigger word:
insertion assDescription
FAQ
Comments (7)
why did iyo udecide to split them?
I wanted to see if the results would be more reliable and stable if I trained only on pictures of each subject individually.
There was a lot of concept bleed and less control when I had them combined.
The results are a bit mixed, but alright for sure.
I think the split ones produces targeted insertions better, but it's still not 100%. Even though the training data is accurate and exclusively to one region, I think there is concept/token bleeding from the words pussy / ass in general. It's likely the training sets include lots of pictures with both pussy and ass.
@6tZ so in the conclusion is that it performs better when split in separate loRAs?
Iso this model makes object more truthfull to description?
I once tried to make a ginat loRA with 10 tags.
It worked very badly. Should have split it.
@barlogspam176 Yes, that's my conclusion.
With my 4 concepts, they also have a different number of images. But since they all train together, this means that they will train with the same number of repeats. So a concept with multiple times the number of images as another concept would have a stronger weight. I think.
Pony does this without a lora . The Lora it self does not work well in sdxl models. Keep up the work!