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
337 stars 25 forks source link

Results of DPIR baseline #29

Closed ChongWang1024 closed 3 months ago

ChongWang1024 commented 3 months ago

Hi,

Thank you so much for sharing such an interesting work. I have a question about the results of DPIR on FFHQ and ImageNet (since these datasets are not evaluated in the original DPIR paper). Do you train the DPIR denoiser on FFHQ and ImageNet separately or directly use the original pretrained DPIR?

Thanks in advance!

yuanzhi-zhu commented 3 months ago

Hi, As mentioned in the paper,

To ensure fairness, we employed the same pre-trained diffusion models and blur kernels for all methods in the comparison.

the same pre-trained diffusion models are used as denoisers to evaluate DPIR.

ChongWang1024 commented 3 months ago

Hi, Thanks for your prompt reply. So you use the same network trained using diffusion training, and directly plugging into the DPIR iteration to produce the result?

yuanzhi-zhu commented 3 months ago

Exactly, as we have argued in the paper that diffusion models are trained as denoisers (and we can use them as denoisers!)

ChongWang1024 commented 3 months ago

Interesting! this has addressed my question.

yuanzhi-zhu commented 3 months ago

@ChongWang1024 you may want to have a look at this slide to better (than the paper) understand this work (and related works)

ChongWang1024 commented 3 months ago

@ChongWang1024 you may want to have a look at this slide to better (than the paper) understand this work (and related works)

Thanks, these slides help a lot!