[retroencabulato]

I don't think that article made strong arguments.

Using nearest-neighbour resampling on the low-res image is an absolute joke. He didn't even look at an objective quality measurement (PSNR). The human visual system is very sensitive to edges, and the high-res image has more pronounced blocking artefacts. Downsampling a high-res image is an unnecessary load on the end user, the 8x8 block transform was chosen for good reason.

I'm not necessarily saying the low-res is superior, but I disagree this ad hoc method is the 'best' way (compared to optimising the coding).

[jacobolus]

PSNR is not a very good test of human perception though, and optimizing for PSNR has had unfortunate consequences on image resizing algorithms IMO.

Whether downscaling a high res image is an excessive load on the end user depends a lot on the end user. From what I understand (but I’m not an expert, so please correct if this is wrong), bandwidth is the main bottleneck not only on I/O latency but also on CPU use, not image rescaling. I’m guessing even mobile phones of the last few generations don’t even break a sweat when downscaling images (can’t they use GPU for this?). As always, it would be a good idea to actually test CPU use / latency / battery drain from rendering images at different sizes/JPEG quality levels on the target client device.

[retroencabulato]

Other than pathological examples, PSNR is pretty useful for gauging the quality of photographs, and is very common in literature. Problems aside, it would have been nice to see some rate-distortion curves.

I agree eye balling the results is just as important, but I don't believe everyone should adopt this method because one dude thinks it looks better. Personally, I dislike the blocking artefacts around the neck and badge of the high-res image, even if some details are sharper.