search

kongzhecn / OMG

[ECCV 2024] OMG: Occlusion-friendly Personalized Multi-concept Generation In Diffusion Models
https://kongzhecn.github.io/omg-project/
559 stars 38 forks source link

Fixed keywords ? #7

Closed Adenialzz closed 3 months ago

Adenialzz commented 3 months ago

hi, thanks for you release this great work. I have a question here.

https://github.com/kongzhecn/OMG/blob/master/inference_lora.py#L276-L283

is this fixed keywords man and woman ? what does this mean ?

Looking forward to your reply.

Adenialzz commented 3 months ago

maybe this two keywords just used to make difference masks from stage 1 results? and stage 2 put lora on the different masks to generate different identities? is my understanding correct?

kongzhecn commented 3 months ago

Yes. The overall architecture of the proposed method contains two stages during sampling. The first stage generates an image containing a man and a woman based on the input text prompt, without considering personalization. After the image segment, the ID of the man and woman can be injected through the mask.

Adenialzz commented 3 months ago

OK, I see. Great work.