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