hamadichihaoui / BIRD

This is the official implementation of "Blind Image Restoration via Fast Diffusion Inversion"
243 stars 18 forks source link

Why DDIM can ensure the generated images within the manifold of realistic images? #12

Open yxuhan opened 4 months ago

yxuhan commented 4 months ago

Thanks for the paper and the code!

In a recent paper, it demonstrates that using DDIM inversion, one can embed a dog face into a Diffusion Model trained solely on FFHQ (a famous high-quality face image dataset). See https://arxiv.org/pdf/2310.09213v1.

In my experiments, I also find one can embed a very blurry face image into a Diffusion Model trained on FFHQ via DDIM inversion. So I wonder why DDIM can ensure the generated images within the manifold of realistic images in the proposed method?

Thanks in advance.

hamadichihaoui commented 5 days ago

Hi, Sorry for the very late reply. In our case we are inverting the blurry image assuming that there is a blur operator that generates it. which is a strong regularization that makes the inversion favors a clean image over outputting the same blurry image. In the paper you mentioned, as far as I understood the authors invert the blurry image as if it is a clean image (they don't take into consideration blur model during the inversion)