QWEN image Lora trigger word

[Anothergazz]

I trained my first LORA for QWEN image on a person. But when I use that LORA all people of that gender look like the character in my LORA. I assume this is because there is no unique identifier for my character and hence it is associated with the gender.
How can I get a unique identifier for my character?
Can the text encoder bye trained to solve this ?
Any suggestions ?

[ryanzigler]

You need to build your unique identifier into your training image's captions. So if you have a male character named, Bob, you want to caption your images of Bob something like this:

• b0b man standing on a city street, looking off to the right. He wearing a white t-shirt shirt.
• `b0b man in a dimly lit sports bar. He is wearing a red baseball hat and red sweatshirt. Lights and multiple TVs are in the background."
• b0b man wearing a red and black flannel shirt. The background is a field at sunset.

By embedding the unique identifier in the caption, you can train Qwen that only b0b man looks like Bob rather than all men looking like Bob. The other thing to remember is anything you mention in the captions that describes your character are things that should NOT be a part of the character's core likeness. So if Bob always has a curly red afro for his hair style, his curly red afro should never be mentioned in the captions. But if Bob sometimes wears glasses and sometimes doesn't wear glasses, you would want to mention that he is or is not wearing glasses in your captions.

[Anothergazz]

Thanks for the feedback, I actually did this but it still made all the woman look like twins. I wonder if this is a case were some regularisation images would help.
I'm not certain that OneTrainer encodes a unique identifier into the Qwen model or Text encoder (does it even need too). All above my knowledge level.