santa hat
deerdeer nosedeer glow
Sign In

Advice on Colab LORA training for multiple outfits

So I've gone through the various guides found here and they have proven useful no doubt. But when I'm training a particular character I'm still having problems with, I suppose the term could be 'bleed through', from the other outfits where elements of them appear on an outfit when I didn't prompt for them.

I've pruned the tags, refined the dataset, tried it as LORA, LOCON and LYCORIS and changed the training steps but it seems the only way to lower the 'bleed through' is by putting the other outfits in the negative prompt. Any advice on how to prevent this would be very welcome.

2 Answers

It should be possible with any method of training. The only thing that matters there it would be the dataset and the captions.

Are the elements that are bleeding through tagged or omitted? i.e. have you removed the tags and used a trigger word for an outfit?

Hi, thanks for replying.

So for each outfit I've given it a collective tag to seperately identify them from one another, one civilian/standard outfit and two superhero costumes. Both the superhero costumes have distinct domino/eyemasks and sometimes they bleed through as well as the arms/gloves (one having white gloves and the other having black gloves). Another thing is one outfit has faux animal ears which sometimes appears on the one that doesn't.

I've omitted all tags from my dataset which identify distinct elements. So for one outfit with white arms/gloves, I haven't included that as a tag.

They all share a common tag which is the tag for the character itself, it precedes all other tags in the text files.

Your answer