(I sent this earlier to Chris, without CC'ing the PNG Working Group.)
>Gain Maps are an (allegedly) efficient way to provide an SDR baseline
image plus an HDR alternate image. The claim of storage efficiency has
not been backed up by any measurements that I have seen.
FWIW: From a glTF, real time rendering, video game, and browser texturing
perspective, unless Gain Maps *are 100% lossless* with respect to the
interchange of IEEE 754-2008 half float pixels (FP16), this isn't a
compelling solution for our use cases. We need no loss whatsoever for
denormals, all valid normals, signed values, all valid 5-bit exponents,
NaN's, and Inf's.
PNG is a lossless format, and we need the ability to losslessly interchange
half float source textures and images while preserving compatibility with
existing 16-bit PNG software. We also need the ability to store half float
data in any colorspace, or even in no colorspace: such as normal maps,
which the use of SDR PNG's is pervasive, or height maps, etc. (Both normal
and height maps are commonly viewable SDR PNG content right now.)
I searched this doc for "half" and found nothing. The only mention of
quality was with respect to the SDR image. Gain Maps also look more complex
vs. just storing remapped half values in the PNG file, involving a resample
stage for example. The source data for this approach (the "Image File" in
this diagram from the doc below) would be an HDR image, *even a "haLf"
float PNG image*.
[image: image.png]
