weight could be 0.6 - 0.8( makes it flexible for styles ) - trained for 768 or higher - to keep close to her old design try any of these - or put in Neg to reduce chance - SFW/NSFW
may need to add 3d, or realistic, in neg for certain 2d/anime styles due to having more 3d image training - or lower weight for less influence - higher the versions
V2 - TW was changed to - pl3stic, - also has stronger training
face/body
plastic, or pl3stic, brown eyes, black hair, ---- - short hair,
Outfit
white shirt, single forearm glove, or glove, headband, or black headband, - pants, or black pants, shoes,
headphones around neck, - or headphones in NEG
suggest nipples, in NEG for SFW
Description
FAQ
Comments (4)
I'm thinking the trigger word being a common English word is resulting in SD ignoring it and just using the rest of the prompt to infer what the model looks like. The pic I posted is the same prompt and seed against 3 different checkpoints and the lower row is the same prompt without the trigger word in it. This is probably why we see leetcoded character name triggers like "pl3st1c".
Id assumed plastic was a term that had no effect on the results since im not even sure what plastic would even spawn - Ill just change it sometime today, it would have to be retrained
@True_Might Thanks for the reply, I hate making your job harder
@ydoomenaud actually I think an image in the dataset may have sabotaged the dataset for both loras - so i will be unpublished in a few