Quality of RealSet65 results

[avgalichin]

Hello!

Thank you for the great work.

I've been running the version-2: Conference paper, 15 diffusion steps, trained with 500k iterations model in your written colaboratory script, as well as independent manual tests with different seeds on RealSet65 dataset.

But I haven't managed to achieve the same visual quality compared to the results presented in the paper. Although the details are still pretty good, it is still worser than the papers from my sight.

Specifically, I couldn't get the same quality of the right "eye" in the Lincoln.png example, and got some artifacts and slightly worser quality for the "0030.jpg" sample.

The generated results by my run:

Please, note the right eye quality in the first picture and some artifacts on the clothes near the neck in the second image, which are not presented in the paper results.

Maybe I miss something, or it is related to randomness? Can you please comment on this issue.

Thank you very much in advance!

[zsyOAOA]

Please set the chop_size as 256 by using the following commands:

python inference_resshift.py -i [image folder/image path] -o [result folder] --task realsr --scale 4 --version v2 --chop_size 256

This will produce the results consistent with our paper. @avgalichin

[avgalichin]

Hello!

I've tried, but still couldn't obtain the same result.

I run the following cell:

I've also tried to run the same cell with different seeds (--seed) and this can give some noticeable differents in the generated content. For instance, here:

The first one is generated with the default seed by running the command you've shown with --chop_size 256, while the second one was generated with some other random seed.

If there is no errors in the command I run, maybe the main difference is due to randomness?

[zsyOAOA]

I have tried directly using the python script (not colab), and obtained consistent results. Could you have a try with the python script? I haven't carefully checked the consistency of the colab demo.