nagadomi/waifu2x

Image error using CUnet and denoise only

Closed this issue · 2 comments

VireV commented

I was looking to denoise a very noisy image using waifu2x. Obviously, the image is way too noisy to get any good output from but I did it anyway. The original image:

ESzKB3RUwAEHZoV

The results from other models seemed to give poor results so I tried the CUnet model. However, when I did, the output produced very distorted:

This is with denoise only and set to level 1:
ESzKB3RUwAEHZoV(CUnet)(noise)(Level1)

This is with denoise only and set to level 2:
ESzKB3RUwAEHZoV(CUnet)(noise)(Level2)

Is this a problem with the model itself or a bug elsewhere in the program?

Thank you.
(I am using waifu2x-caffe)

nagadomi commented

This is not the waifu2x-caffe's repository. It is https://github.com/lltcggie/waifu2x-caffe

Is this a problem with the model itself or a bug elsewhere in the program?

There is no distortion problem with my implementation http://waifu2x.udp.jp/ .
When using CUNet model, Split size must be a multiple of 4. Try 128 or 64.

  • 👍1
VireV commented

Sorry about that, I'll close it now.