search

mandalinadagi / WGSR

56 stars 4 forks source link

Positive results #5

Open neosr-project opened 7 months ago

neosr-project commented 7 months ago

Hi, wanted to share some positive results using your research. I took the liberty to simplify your loss implementation into a single file, in neosr.

Left is without, Right is with Wavelet-Guided loss. All tests performed using manual seed and torch deterministic algorithms (making it bitwise reproducible). First few iters has instabilities, but it soon recovers and stabilizes gan. Convergence video: https://github.com/mandalinadagi/WGSR/assets/132400428/cf0da31e-70c0-4e76-8d76-3ad6d3707da7

Side-by-side: wg_comparison

mandalinadagi commented 7 months ago

Thank you so much for your work, are those results generated from bicubic downsampled images or from real-world samples?

neosr-project commented 7 months ago

Those tests are on bicubic degraded, yes. We have also tested it in real-world SR with positive results :+1:

@Phhofm finetuned a ATD model using it

mandalinadagi commented 7 months ago

Great, I will be adding those results and give reference to your repo for enabling the utilization of our wavelet domain loss with different methods. Thanks again, cheers!

neosr-project commented 7 months ago

Please feel free to use however you wish, thanks. Here are the original images: wg_test.zip. And models (SPAN network, inference can be done using test.py with this configuration): models.zip

Phhofm commented 7 months ago

I had also seen positive results of handling output artifacts using this research while training a dat2 model. The following are examples of where I first trained for 90k iters, then enabled the wavelet-guided loss in neosr (only change i made), then continued training until 190k iters.

Slowpics slider comparison, 3 examples

wavelet_example1 wavelet_example2 wavelet_example3

mandalinadagi commented 7 months ago

great news, thank you!

caijie0620 commented 7 months ago

Hi, wanted to share some positive results using your research. I took the liberty to simplify your loss implementation into a single file, in neosr.

Left is without, Right is with Wavelet-Guided loss. All tests performed using manual seed and torch deterministic algorithms (making it bitwise reproducible). First few iters has instabilities, but it soon recovers and stabilizes gan. Convergence video: https://github.com/mandalinadagi/WGSR/assets/132400428/cf0da31e-70c0-4e76-8d76-3ad6d3707da7

Side-by-side: wg_comparison

Why do I think the left looks more visually appealing? The right one has more artifacts.

neosr-project commented 7 months ago

Why do I think the left looks more visually appealing? The right one has more artifacts.

Baseline is sharper, doesn't mean it has more 'details'. Wavelet decreased the artifacts created by GAN training, which is the purpose of this research. It's visible on the t-shirt on the first sample.