search

yuanzhi-zhu / DiffPIR

"Denoising Diffusion Models for Plug-and-Play Image Restoration", Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, Luc Van Gool.
https://yuanzhi-zhu.github.io/DiffPIR/
MIT License
388 stars 31 forks source link

How about unknown complex degradation type? #12

Closed green-gooes closed 1 year ago

green-gooes commented 1 year ago

Hi, thanks for the great work first :)

There are 3 types of degradations mentioned in the paper, which have explicit expressions as suggested by Eq. 25, 27, 29. In real world, the degradation might be complex and unclear, I wonder how to deal with such degradation? could you please give some insights? thanks in advance!

yuanzhi-zhu commented 1 year ago

Hi @green-gooes ,

Thank you for your interest, the existing several training-free diffusion-based IR methods such as DDRM, DPS, DDNM and ours all require known degradation operators.

For real-world IR tasks with only paired data, you can either estimate the degradation kernel or train an end-to-end model with synthetic (augmented) data.

green-gooes commented 1 year ago

Thanks for the reply! I will try the second solution.