How does this impact some objective quality metrics like SSIM, PSNR-HSV, VMAF?

[jrmuizel]

It would be interesting to see the impact deblocking has especially at lower bitrates.

[jyrkialakuijala]

We only tested with butteraugli which is a maximum error based metric. We saw reduction of error values in smooth areas, but most often the maximum error is in a non-smooth area and is not affected by better performance in slow gradients.

A L2-norm based quality metric should be able to see the improvement, but we didn't test that.

Overall we consider the approach compatible with a browser visualizing jpegs, but we didn't have time to try out integration or performance optimization yet.

(This algorithm is embarrassingly simple to implement.)