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

LR image generation #10

Closed JustinAsdz closed 4 years ago

JustinAsdz commented 4 years ago

Hi Fritsche: I've got a question about training the DSGAN from SCRATCH, whether the LR images are directly generated by bicubic downsampling , and if crop is used.

Thanks

JustinAsdz commented 4 years ago

Is it just bicubic from fig.3 in the paper? or 512x512 patches for training DSGAN and 128x128 for discriminator as described in part 4.1 of the paper

manuelfritsche commented 4 years ago

We randomly crop each training image to get images of the size 512x512. Using these cropped images, we generate the LR image by downsampling them, which results in 128x128 images. The discriminator image is created by simply cropping a 128x128 image.