solidpixel
commented
4 days ago Rejecting for now - needs more work.
Closed solidpixel closed 4 days ago
solidpixel
commented
4 days ago Rejecting for now - needs more work.
Status:
Suspect I have the wrong end of the stick here - MRSEE on log values is still going to be incorrect. For MRSEE to make sense we need to use linear colors, diving in/out of LNS encoding when we need to test quantization. This is going to end up a lot more expensive ...
Description
HDR texture values are stored logarithmically.
Using absolute sum of squares on the logarithmic values (the existing behaviour on main) causes the compressor to spend too much effort preserving imperceptible shifts in dark channel values at the expense of bright values in the same block. This performs poorly in blocks with sharp luminance changes (dark texels in a bright block) and in blocks with saturated color values (dark channels in bright pixels).
Using absolute sum of squares on linearized HDR values (attempted by the first patch in this PR) avoids the compressor fixating on dark values, but instead causes the compressor to spend too much effort preserving bright values. This is because the errors in the bright channels can be orders of magnitude bigger than the errors in the dark channels, and dark values can end up quantizing close to black.
Using relative sum of square on the logarithmic values, proposed by Ryg in the blog below, encourages the compressor to find a balance of relative error across the whole block, favoring neither light nor dark channels. Blog: https://fgiesen.wordpress.com/2024/11/14/mrsse/ .
This is a perceptual error metric for HDR textures, so we expect that this will make the mPSNR score worse because we are no longer optimizing for absolute error. Test failures in HDR textures are therefore expected, and LDR textures are not affected.