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

please give me some guidence for training a Super-Resolution model on my own dataset #14

Closed lmvgjp closed 11 months ago

lmvgjp commented 11 months ago

Hello,

I am a bit lost of how to start...still reading your great paper, so i still have many things to learn. The idea is to train the denoiser on my own data to get super-resolution but i am a bit confused with the code. How do i train a model for that? thanks a lot! greetings

yuanzhi-zhu commented 11 months ago

Hi @lmvgjp

Thank you for your interest.

The key idea is that diffusion models are trained as denoisers and we can use pre-trained diffusion models as plug-and-play image prior(denoisers). That's to say, our code base does not involve the training of diffusion models as we use the pre-trained ones provided by previous work.

See this slides for a better understanding of our work!

If you want to train a diffusion model from scratch, you can refer to https://github.com/openai/guided-diffusion or https://github.com/NVlabs/edm.

If you want to train diffusion models specifically for SR, you can try https://github.com/Janspiry/Image-Super-Resolution-via-Iterative-Refinement

lmvgjp commented 11 months ago

Hi Yuanzhi,

Thank you so much for taking your time to advise me! you are very kind greetings