Poor image quality

[Zhentao-Liu]

Thanks for you code, it's a great job. I use python sr.py -p val -c config/sr_sr3_64_512.json and your pretrained models to perform super-resolution. But I get poor results which are noisy and blury. step=2000. Could you give some advice?

[Lxy98Code]

Thanks for you code, it's a great job. I use python sr.py -p val -c config/sr_sr3_64_512.json and your pretrained models to perform super-resolution. But I get poor results which are noisy and blury. step=2000. Could you give some advice?

Hi, Zhentao. I think your results are reasonable since the pretrained model author shared is limited to Nvidia 1080Ti. Enlarge the iterations when you train the model, and the results should be better. I did the same thing as you. However, I found even worse results with same command python sr.py -p val -c config/sr_sr3_64_512.json and pretrained model 'I830000_E32', step=2000. Except for the 'resume state', I didn't change any other config. Do you have any experience about these worse results? Could you share any suggestion? Thanks in advance!

[Zhentao-Liu]

Thanks for your reply. You mean I should increase the train step to get better results? Maybe I should try it later. As for your results, it's really wired. I use the same command and the same pretrained model, and I get my results. Isn't there anything wrong? Maybe you should check it more carefully.

[Lxy98Code]

Yes, I think so. Take a try to increase the training step for your own model. You can find more details from this issue https://github.com/Janspiry/Image-Super-Resolution-via-Iterative-Refinement/issues/78.

[Zhentao-Liu]

Yes, I think so. Take a try to increase the training step for your own model. You can find more details from this issue #78.

Thanks very much.

[huchi00057]

Thanks for you code, it's a great job. I use python sr.py -p val -c config/sr_sr3_64_512.json and your pretrained models to perform super-resolution. But I get poor results which are noisy and blury. step=2000. Could you give some advice?

Hi, Zhentao. I think your results are reasonable since the pretrained model author shared is limited to Nvidia 1080Ti. Enlarge the iterations when you train the model, and the results should be better. I did the same thing as you. However, I found even worse results with same command python sr.py -p val -c config/sr_sr3_64_512.json and pretrained model 'I830000_E32', step=2000. Except for the 'resume state', I didn't change any other config. Do you have any experience about these worse results? Could you share any suggestion? Thanks in advance!

Actually, i had the same trouble like you in my custom weight. But if i executed it on Colab, everything would be perfect.

[BlouseDong]

Hello, I have a question for you。 “python infer.py -c [config file]”
What is the config file here? How should I define input low pixel images? In other words, I don't know how to conduct testing.

[927514606]

Thanks for you code, it's a great job. I use python sr.py -p val -c config/sr_sr3_64_512.json and your pretrained models to perform super-resolution. But I get poor results which are noisy and blury. step=2000. Could you give some advice?

Hi, Zhentao. I think your results are reasonable since the pretrained model author shared is limited to Nvidia 1080Ti. Enlarge the iterations when you train the model, and the results should be better. I did the same thing as you. However, I found even worse results with same command python sr.py -p val -c config/sr_sr3_64_512.json and pretrained model 'I830000_E32', step=2000. Except for the 'resume state', I didn't change any other config. Do you have any experience about these worse results? Could you share any suggestion? Thanks in advance!

Actually, i had the same trouble like you in my custom weight. But if i executed it on Colab, everything would be perfect. Hello, I just started to use colab, may I ask how you train in colab, can you share? Thank you very much