Feature request: Quality-optimized subsampling

[tdaede]

There are methods for producing 4:2:0 subsampled video that aim to provide more optimal results once displayed (both subjectively, and via cross-coefficient PSNR and CIEDE2000 metrics):

https://norkin.org/pdf/SPIE_2016_HDR_conversion_metrics.pdf

Despite the HDR title, the methods are also useful for SDR content. The paper provides 3 methods, of which the "Closed Form 1" method is probably most suitable for a SIMD implementation.

Also, these algorithms currently only perturb the luma coefficient. This helps make the problem tractable but there are likely better algorithms that could potentially generate better chroma values as well.

HDRTools also has an open source implementation that can be checked against (HDRTools is more a reference implementation rather than a practical library like zimg).

[tdaede]

Yes, this is specifically for encoding only (or slightly more generally, when downsampling chroma).

The method in the paper converts RGB->YUV and does the downsample in one step. This doesn't map well to any of the existing zimg filter types. A YUV->YUV filter type can also be derived fairly efficiently as long as the colorspace is known, and the input is non constant luminance. I don't know if that would be worth implementing over simply forcing a conversion to RGB with an existing colorspace filter, though.

[Artoria2e5]

The webp people has a different, iterative thing that touches not just luma. It's available as a a libyuvsharp your program can link to, but it's only intended for static images given the speed.

mozilla/mozjpeg#193 is related