Recently, I made a Lora of a character (you can see and test it here: https://civitai.com/models/48810/emiri-uchi-watamote-ucchis-face). I tried to embed 2 concepts in it: the character itself and her emoji face.
For the character part, it worked pretty well. But for the face, it heavily linked to the character and I need to make some tricks to put it on another character without totally modify the picture (plus these tricks aren't consistent).
So I ask for those of you who has tried: How can I improve the Lora on this point ?
I used 26 images for the training of the concept, and since this character is not very famous, there aren't a lot of good quality images of her, so she makes about 2/3 of the training set. The rest of the directory is made of images of other character with similar face.
The parameters of both concepts are the same, except for the number of repeat. I used 15 for the face and 10 for the character (I already tried to also put 10 for the face concept, but the results were worse)
Thank you in advance for your answers.
1 Answer
(had to write this again, somehow it didn't save my previous response)
Basically you trained the model to associate the emoji with that character. You need to show the model otherwise. Ideally you would have more characters with the same (or very similar) emoji. If you have multiple characters with the same emoji, you can tag the emoji trigger word whenever it is visible, and not tag it when it's not visible.
I don't think you could improve that LoRA, but you should instead train another one separately.