How do you generate such an image？

[JimjimChen]

Can you tell me how you do the experiment? Is there any open source code?

[sefibk]

I don't understand.
This repo is exactly the open-source code

[sefibk]

AFAIK Michaeli & Irani didn't publish their code. I got Irani's lab to run the kernel estimation on the DIV2KRK dataset that you see in the figures

[JimjimChen]

How is the Ground-truth kernel obtained ，the second row in Figure 5 and the first row in Figure 6

[sefibk]

It is randomly generated, then we downscale a HR image to obtain the LR version.
The details can be found in chapter 5.2 of the paper

[JimjimChen]

and what does the "Convolving all filters of G results with the SR Kernel K" means? The "K" is the Ground-truth kernel? K is another input of Generator?

[sefibk]

The whole essence of the work is to estimate K. it is not given as input.
I recommend reading the paper

[JimjimChen]

It is randomly generated, then we downscale a HR image to obtain the LR version. The details can be found in chapter 5.2 of the paper

Could you tell me the specific operation? ”randomly generated“ means kernel estimation of GT image？if not， how to randomly generated and visualization？Is there a open code？

[sefibk]

I think you are missing the essence and you should re-read the paper.
in short:
We first generate the dataset with randomly generate kernels (AKA GT kernel).
Then apply one GT kernel for each HR image and obtain LR images.
We then apply kernel-GAN to estimate that GT kernel, but note that all the GAN sees in the LR image.

The generation of the dataset can be found here:
https://github.com/assafshocher/BlindSR_dataset_generator
The GT kernel's are generated randomly so each image is then downscaled with a different kernel.