It'd be awesome if we could give the spandrel a facelift with some new functionality. :)

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

It'd be awesome if we could give the spandrel a facelift with some new functionality. :) #152

Closed zelenooki87 closed 6 months ago

zelenooki87 commented 7 months ago

following architectures:

FFTFormer: https://github.com/kkkls/fftformer M3SNet: https://github.com/Tombs98/M3SNet MHNet: https://github.com/Tombs98/MHNet Maxim: https://github.com/vztu/maxim-pytorch/tree/main And: Restormer: https://github.com/swz30/restormer (because, KBNet and Nafnet do not work properly for some images - I opened a bug issue about this)

RunDevelopment commented 7 months ago

I added support for Restormer (#170), FFTFormer (#171), and M3SNet (#172). The other 2 are weird:

The official Maxim implementation uses a different framework that isn't compatible with PyTorch. You linked a PyTorch re-implementation of Maxim.
MHNet is a mess. The author deleted the arch code (it's still accessible via git history) and there are no pretrained models. The paper makes it out to be an amazing arch, so I wonder why the repo has these problems. Anyway, I already put together the arch code, but it's useless without pretrained models to test it. Without pretrained models, I cannot add MHNet.

zelenooki87 commented 7 months ago

Thank you very much, friend! I saw them in benchmark tests on the paperswithcode website, so I thought they would be okay. I haven't tested them all. Thank you very much again!

zelenooki87 commented 7 months ago

Is it possible to add support for RESTORMER's gaussian_color_denoising_blind.pth model? ( https://drive.google.com/drive/folders/1Qwsjyny54RZWa7zC4Apg7exixLBo4uF0 ) Chaninner not supports it. Other restormer models that I have tested working okay.

Thank you again.

RunDevelopment commented 7 months ago

Chaninner not supports it.

Because we haven't updated spandrel yet. Chainner currently doesn't support any of the above architectures, but it will in a few days :)

Other restormer models that I have tested working okay.

How...

zelenooki87 commented 7 months ago

I apologize, my mistake! Up until recently, it was working with custom builds. I would periodically update it by manually copying files from the source folder. Unfortunately, that's practically impossible now. Thanks for the info, I'll wait for it to be fixed.

By the way, about the white balance we discussed in another thread (if you haven't seen it yet), is there a chance we can integrate that as well?

RunDevelopment commented 7 months ago

By the way, about the white balance we discussed in another thread (if you haven't seen it yet), is there a chance we can integrate that as well?

It's not rejected, but nobody is currently working on it. It's a relatively low-prio item, so it might take some time until we implement it.

RunDevelopment commented 6 months ago

I'm going to close this issue now.

I'm unsure about the PyTorch reimplementation of Maxim. It might be useful, but it feels a little weird to add it. Please make a separate issue if the architecture is dearly wanted.

MHNet is rejected due to problems with the repo (author deleted arch code, no pretrained models). If those problems are somehow resolved in the future, we can add it then.