ESRGAN dose not perform well with larger image with pretrained model

[sweihub]

Hi Mr. Xintao

I tested the pretrained model with the old baby pictures, this is the partial view of the result, you can see there're blurred small squares, the result is not as good as expected, maybe I need to train with high quality portrait dataset?

I checked the training code (BasicSR), so if I want the result to be 2K, that means the training ground truth size should be 2K/4 = 512, (train_MSRResNet_x4.yml --> gt_size: 512), the arbitrary image super resolution will not produce good result, am I right?

Thanks!

The original photo for your reference, which is 960x1280.