search

manuelfritsche / real-world-sr

[ICCVW 2019] PyTorch implementation of DSGAN and ESRGAN-FS from the paper "Frequency Separation for Real-World Super-Resolution". This code was the winning solution of the AIM challenge on Real-World Super-Resolution at ICCV 2019
MIT License
162 stars 37 forks source link

Trouble training ESRGAN with jpeg artifact images. #6

Closed jfun9494 closed 4 years ago

jfun9494 commented 4 years ago

Hello! Thank you for posting your code on GitHub. I successfully recreated the first part of your method (training DSGAN). Specifically I trained for jpeg artifacts and generated a paired dataset for the SR network (ESRGAN-FS). The generated dataset looks convincing, DSGAN was able to generate jpeg artifacts that look real. Then I trained ESRGAN-FS with your given codes/options/jpeg/train_TDSR.yml config file (all settings unchanged and placed the RRDB_ESRGAN_x4.pth pretrained model in the appropriate folder).

However, I am getting very weird results on the validation set. The images look cartoonish and the details are missing. This is for the TDSR case (50000 iteration), where I am trying to get clean HR images from generated LR images from DSGAN with jpeg artifacts. ESRGAN output

Now, the weird thing is that, I even checked the supervised training where LR images are real images with jpeg artifacts (not generated) and still get the same cartoonish ouput. So the problem is with ESRGAN-FS, not DSGAN generated dataset. In addition, I also train the original ESRGAN (from BasicSR repo), and the got the same problem.

Can you help me with this? Maybe the issue is with the training configs (learning rate, loss weights) in the options files?

manuelfritsche commented 4 years ago

Hi jfun9494. The results look as expected. The TDSR task is quite difficult and the results are not perfect yet.

jfun9494 commented 4 years ago

Thank you for the response! Yes I also found TDSR to be quite a challenging task. Also thanks for sharing the pretrained models.