search

chaiNNer-org / spandrel

Spandrel gives your project support for various PyTorch architectures meant for AI Super-Resolution, restoration, and inpainting. Based on the model support implemented in chaiNNer.
MIT License
139 stars 12 forks source link

DRUNet support #14

Closed Wurlit closed 8 months ago

Wurlit commented 1 year ago

I'd like to ask for DRUNet model support, as an alternative denoising option. CUNet has troubles with very noisy images, something that DRUNet claims to overcome, as can be seen here: https://github.com/cszn/DPIR

Thanks.

adegerard commented 1 year ago

Made some tests. Used the image "(a) Noisy image with noise level 200". It does not give the same result... Maybe something wrong?!? The SCUNet he implemented later is better (same author)

Wurlit commented 1 year ago

There was an issue about real world images here: https://github.com/cszn/DPIR/issues/4 where the author suggested using https://github.com/cszn/DPIR/blob/master/main_dpir_sisr_real_applications.py for real applications. Maybe that's the issue?

Can you also test the b/w Pink Floyd image I tried with SCUNet a while back here: https://github.com/chaiNNer-org/chaiNNer/issues/2169 and see if it's any better?

Thank you for the interest.

Out of topic: Can I also post another feature request for his deblurring and super resolution models shown on the same page below the denoiser? The examples look also very neat.

adegerard commented 1 year ago

This was not an issue, it was just a guy who didn't know which script has to be used for rl. There are multiple parameters which have to be modified by hand. Here is your image once 'denoised' pr2169_x1_drunet_gray

Here is a comparison of the images I used some time ago: noisy input_noisy_dimensions DPIR (do not remember the parameters, but spent too long time) input_noisy_dimensions_x2_drunet_color SCUNet (GAN) input_noisy_dimensions_scunet_gan

Before asking to integrate other algorithms, would you please try them to check it's worth it? Thanks.

Wurlit commented 1 year ago

I understand your confusion, I didn't expect such deviating results. If you don't think it's worth the trouble, I guess you can close the issue, I don't know. Thank you anyway, just for putting the effort.

RunDevelopment commented 8 months ago

I'm gonna close this now, since SCUNet seems to be better in every way and (more importantly) actually used by the community.