Upscaling a LR 512px image to 1k or 2k resolution (Factor of 2x or 4x)

[sanchit-ahuja]

This is with respect to Issue #19 . In there you have answered that to evaluate an image other than one from Div2K validation set, we need to downscale. So is it not at all possible to work with varied Low-res images like upscaling a 1080p to 2k or 512p to 1k etc?

[krasserm]

In there you have answered that to evaluate an image other than one from Div2K validation set, we need to downscale.

In #19 the discussion was mainly about how to reproduce downgrading as done in the DIV2K dataset. Technically, you can feed the pre-trained models with whatever LR image you like but the quality of the SR image generated with the pre-trained models will depend on how the LR image was obtained.

So is it not at all possible to work with varied Low-res images like upscaling a 1080p to 2k or 512p to 1k etc?

If your LR images have been obtained with a bicubic downgrading function you'll likely get useful results with the pre-trained models. If your LR images have been obtained with a significantly different downgrading function (for example, by taking images with a low-quality mobile phone plus applying strong JPEG compression) you probably want to re-train the models with a different downgrading function. The DIV2K data provider supports several pre-defined downgrading functions (see downgrade parameter in the constructor). If you want to specify your own downgrading function you need to write your own LR/HR data provider.

[sanchit-ahuja]

Thanks a lot for a quick reply.

[sanchit-ahuja]

So I tried upscaling a 512X512 to 2048X2048 image. There were couple of problems.

1. The inference time is around 10 seconds to get the output image.
2. The PSNR for upscaling these images is quite low.
   Is it possible to increase the PSNR? Will retraining the model on 512*512 images be helpful?