results are super blurry

[ihavetoomanyquestions]

I am trying to use SRGAN for increasing resolution of some artwork. However, the results I got are super blurry, basically it seems to be doing opposite of what I was expecting: instead of a higher resolution image, I got an image where all nearby colors are mixed together (see this: https://i.imgur.com/1AZpoiZ.png)

My config is as follows:

CUDA_VISIBLE_DEVICES=0,1 python main.py
--output_dir ./experiment_SRResnet/
--summary_dir ./experiment_SRResnet/log/
--mode train
--is_training True
--task SRResnet
--batch_size 2
--flip True
--random_crop True
--crop_size 24
--input_dir_LR ./data/img_lowres/
--input_dir_HR ./data/img_highres/
--num_resblock 16
--name_queue_capacity 4096
--image_queue_capacity 4096
--perceptual_mode MSE
--queue_thread 4
--ratio 0.001
--learning_rate 0.0001
--decay_step 400000
--decay_rate 0.1
--stair False
--beta 0.9
--max_iter 1000000
--save_freq 20000

I am so upset after training the model for so long (1M iterations) and getting this :( What might go wrong?

PS: my images are png files, and I created the low-res version by resizing the high-res ones to 128 x height (the high-res ones are around 800 x height), and I have 3000 images in total.

[harsh306]

Yes,

1. You may try to change the crop size
2. Test intermediate results like after 2k iterations.
3. Analyze your PSNR plot and content loss plots

[webdevserv]

I have the same problem. The results are blurry.

[webdevserv]

PSNR Achieved: 43.998844