Could I use the dsgan with my own image?

[zrongcheng]

Due to the sec 4.1 in paper, I had a confusion.
If could, I want train the pairs('generated DF2K lr' and 'DF2K hr'), does it could get on?
Thanks.

[manuelfritsche]

I am not sure if I understand your question correctly, but yes you can use your own corrupted images for z.

[zrongcheng]

My question is all in the picture above, notice the red word.
But i found the dsgan/data_loader.py, the meaning is below.

[manuelfritsche]

Yes, we used this configuration for training because the goal was to generate images with the same characteristics as the original input images. Instead of the "Cropped Noise image" you can also feed in any other dataset. However, this will require some small modifications with the data_loaders used. For example, you could use the DiscDataloader from dsgan/data_loader.py to serve the second dataset for the discriminator and set cropped=False in the TrainDataset. If you simply zip these two datasets together in the loop, then the rest should work fine. For evaluation puropses you could do the same thing for the valdiation dataset, but most of the measures there are not even using this second dataset, so it is not essential.

[zrongcheng]

Great, thank you very much, I understand what you mean.