sefibk / KernelGAN

Other
337 stars 77 forks source link

For real images, the reconstruction results in a smooth transition #42

Closed zerodohero closed 4 years ago

zerodohero commented 4 years ago

As shown in the figure above, the reconstructed image on the left is overly smoothAre all the hyperparameters in the code tuned?Does the adjustment of hyperparameters have much effect on the results? 20200730173915

sefibk commented 4 years ago

The specified parameters fit the "average" image after reviewing several. Images with more or with less noise require adjusting the network. You can try to tune the --noise_scale added to ZSSR to get a better image. FYI - KernelGAN works best for natural images. Every so often, you might get images where performance will deteriorate...

zerodohero commented 4 years ago

The specified parameters fit the "average" image after reviewing several. Images with more or with less noise require adjusting the network. You can try to tune the --noise_scale added to ZSSR to get a better image. FYI - KernelGAN works best for natural images. Every so often, you might get images where performance will deteriorate...

Thank you very much

I will try your advice