Question about tagging character in different outfits with limited dataset
Hey š, thank you for sharing this tutorial you made!
I am currently trying to train a character LoRA with two outfits, but the bleedthrough between both concepts is severeš.
In your guide you say the tags should not be the same across concepts, however my character has both outfits in the same exact pose etc.
Should I better delete these images for one outfit?
My dataset is already very limited, I am using cropped/zoomed in versions and flip augmentation to augment the datasets (5+2 w/o cropped, 15+6 w/ cropped).
I'd really appreciate your help.
No, you donāt need to delete them.
When I say there should not be the same across concepts, it is only for the costumes you are training, the pose does not matter. All you need to do is to distinguish the tags of costumes, for example, if both costumes have dress, you can use something like āred dressā vs āblack dressā, or āsleeveless dressā vs. āhalter dressā etc. just makes sure there is no costume tags across two costumes.
Would it be advisable to prune the tags which make up the outfits for the outfits to be embedded into the activation tags?