Allow custom LRs to train on JPEG artifacts, dithering, etc

[n00mkrad]

Currently, ESRGAN outperforms SRFlow when it comes to uncompressed upscaling.

And SRFlow can't be trained on JPEGs or other artifacts.

Is it possible to use custom LR images that have been augmented?

[ALLinLLM]

And SRFlow can't be trained on JPEGs or other artifacts.

Why？

[n00mkrad]

Why？

Because the dataset preparation script seems to directly use downscaled versions of the HRs, not allowing any manual editing.

[martin-danelljan]

You can train SRFlow on any kind of LR images, eg with JPEG, noise, etc. Please just modify the data processing pipeline to include the degradations you want.

Regards, Martin