Re: backwards compatible half-float PNG test app source+examples

Note the proposed half float encoding exists in *two colorspaces (or
encodings) simultaneously *because it's been tone mapped: 16-bits sRGB and
another space with a half float encoding, which may be undefined in the
file.

Because it's in 16-bit sRGB it's backwards compatible with virtually all
existing viewers/readers. That's a very powerful property.


On Mon, Nov 13, 2023 at 10:25 AM Richard Geldreich <rich@binomial.info>
wrote:

> It's interesting that the sBIT chunk describes a process quite similar to
> what we're doing to the integer bits of half-floats, to get the high bytes
> in a form that resembles sRGB pixels. The "sample values should be scaled
> to the full range of possible values" - which is exactly what the proposed
> half float encoding does to the high bytes after the sign bit is rotated
> left into the LSB.
>
> Half-floats are a IEEE spec, and very popular on GPU's, so I don't see why
> PNG cannot define a chunk and a spec for how the standardized half floats
> are mapped.
>
>
> On Mon, Nov 13, 2023 at 10:18 AM Richard Geldreich <rich@binomial.info>
> wrote:
>
>> >If you add the `sRGB` tag, so that it is clearly declared that such is
>> the case, then at least you have specified the color appropriately.
>>
>> Not everyone specifies the color appropriately. In practice, PNG is not
>> just a format used for *only* viewing files on a monitor. It's also an
>> interchange format used between programs. In common cases the colorspace is
>> implied or communicated external to .PNG.
>>
>> Yet users are still able to preview and view their files using common
>> OS's, browsers. It's not ideal but the software/OS/tool ecosystem works
>> this way in practice.
>>
>> On Mon, Nov 13, 2023 at 10:13 AM Leonard Rosenthol <lrosenth@adobe.com>
>> wrote:
>>
>>> If you add the `sRGB` tag, so that it is clearly declared that such is
>>> the case, then at least you have specified the color appropriately.
>>>
>>>
>>>
>>> And if someone wants to treat the values as something else – such as HDR
>>> data – they can do so, out of scope to the spec…which is also fine(-ish).
>>>
>>>
>>>
>>> Leonard
>>>
>>>
>>>
>>> *From: *Richard Geldreich <rich@binomial.info>
>>> *Date: *Monday, November 13, 2023 at 10:10 AM
>>> *To: *Leonard Rosenthol <lrosenth@adobe.com>
>>> *Cc: *public-png@w3.org <public-png@w3.org>
>>> *Subject: *Re: backwards compatible half-float PNG test app
>>> source+examples
>>>
>>> *EXTERNAL: Use caution when clicking on links or opening attachments.*
>>>
>>>
>>>
>>> > BUT if they are “backwards compatible” (as you discuss), then at least
>>> for that use case, they still need to be relevant to a color space.  And
>>> what is that?
>>>
>>>
>>>
>>> If we take an .EXR image downloaded from the web without
>>> "chromaticities", "whiteLuminance", or "adoptedNeutral" fields (which are
>>> defined as strictly optional in the spec itself, and aren't always used) -
>>> what colorspace is the output half-float PNG?
>>>
>>>
>>>
>>> The high bytes in the half float PNG (which is what Windows and other
>>> views will preview and view 16-bit PNG's as) in the lossless encoding I've
>>> proposed will be in sRGB.
>>>
>>>
>>>
>>> On Mon, Nov 13, 2023 at 10:00 AM Leonard Rosenthol <lrosenth@adobe.com>
>>> wrote:
>>>
>>> But any pixel value is only useful when considered in a given color
>>> space – unless, as Chris noted in his message, that you don’t consider
>>> these values for display and only for “data storage”.  And if that is the
>>> case, then PNG isn’t the right place for them.   BUT if they are “backwards
>>> compatible” (as you discuss), then at least for that use case, they still
>>> need to be relevant to a color space.  And what is that?
>>>
>>>
>>>
>>> Yes, I agree that EXR is a complex format – been working with that team
>>> on adding C2PA support, so I get it.
>>>
>>>
>>>
>>> But just because you have a hammer, doesn’t make everything a nail…
>>>
>>>
>>>
>>> Leonard
>>>
>>>
>>>
>>> *From: *Richard Geldreich <rich@binomial.info>
>>> *Date: *Monday, November 13, 2023 at 9:52 AM
>>> *To: *Leonard Rosenthol <lrosenth@adobe.com>
>>> *Cc: *public-png@w3.org <public-png@w3.org>
>>> *Subject: *Re: backwards compatible half-float PNG test app
>>> source+examples
>>>
>>> *EXTERNAL: Use caution when clicking on links or opening attachments.*
>>>
>>>
>>>
>>> >First off, HDR doesn’t mean anything without (at least) a CICP table.
>>> So are you adding a standard cICP tag as well to your images?
>>>
>>> The half float encoding != colorspace. Ward: "Because it [half floats]
>>> can represent negative primary values along with positive ones, the OpenEXR
>>> format covers the entire visible gamut and a range of about 10.7 orders of
>>> magnitude with a relative precision of 0.1%." It can handle any colorspace.
>>> Also according to Ward, many .EXR images don't actually encode the
>>> *optional* " chromaticities", "whiteLuminance", and "adoptedNeutral"
>>> .EXR fields in the file, yet somehow we've survived.
>>>
>>> >Second, what is the benefit to your published, but proprietary solution
>>> over an open specification such as OpenEXR?
>>>
>>> The Windows OS will not preview or view .EXR files, but it will preview
>>> and view half-float encoded PNG files if saved correctly in a lossless
>>> manner which we've already solved.
>>>
>>>
>>>
>>> "TinyEXR" is 10,000 lines of code, not counting the miniz library for
>>> Deflate support (which I wrote BTW). It can only load ~70% of the EXR's in
>>> the wild, so as a reliable interchange format it's terrible.OpenEXR itself
>>> is too much code and too bloated (dozens of source files).
>>>
>>>
>>>
>>> The half-float encoding is an industry standard (IEEE 754-2008), and is
>>> widely supported in GPU hardware and in GPU shader languages:
>>> https://en.wikipedia.org/wiki/Half-precision_floating-point_format
>>>
>>>
>>>
>>>
>>>
>>> On Mon, Nov 13, 2023 at 9:33 AM Leonard Rosenthol <lrosenth@adobe.com>
>>> wrote:
>>>
>>> Richard – I don’t understand.
>>>
>>>
>>>
>>> First off, HDR doesn’t mean anything without (at least) a CICP table.
>>> So are you adding a standard cICP tag as well to your images?  If you are
>>> doing that, then at least you’re creating something that can be used in
>>> other workflows.  This is especially important as you think about how these
>>> images will be composited with other HDR content, but in other formats…for
>>> example, consider a web page with both one of your PNGs and a JPEG (using
>>> the upcoming standard)…
>>>
>>>
>>>
>>> Second, what is the benefit to your published, but proprietary solution
>>> over an open specification such as OpenEXR?  Just that it happens to work
>>> in PNG instead of some other format?  From a wider industry perspective, I
>>> don’t see the benefit and instead lots of downsides to users worldwide who
>>> will now be confused why their images don’t work as expected.
>>>
>>>
>>>
>>> I’m certainly in favor of moving HDR forward – but IMO, we need to do it
>>> consistently as an industry and not a series of “one offs”….
>>>
>>>
>>>
>>> Leonard
>>>
>>>
>>>
>>> *From: *Richard Geldreich <rich@binomial.info>
>>> *Date: *Monday, November 13, 2023 at 9:10 AM
>>> *To: *Leonard Rosenthol <lrosenth@adobe.com>
>>> *Cc: *public-png@w3.org <public-png@w3.org>
>>> *Subject: *Re: backwards compatible half-float PNG test app
>>> source+examples
>>>
>>> *EXTERNAL: Use caution when clicking on links or opening attachments.*
>>>
>>>
>>>
>>> > Before introducing another PNG-specific approach, we should see where
>>> the industry is going and see if we can align with them first (which I
>>> believe we can/should).
>>>
>>>
>>>
>>> We need 16-bit/component half float encoding support (not "color" -
>>> that's a different problem) right now for our HDR codec work, so I'm going
>>> to open source this couple hundred line HDR PNG solution with a small easy
>>> to understand readme doc this week, using a ZLIB+Public Domain license.
>>> I'll then work with PNG library authors to get it supported into their libs
>>> and modify existing open source viewers to load HDR PNG's.
>>>
>>>
>>>
>>> We're already using half-float HDR PNG images for testing and
>>> development in our codec work here. It's backwards compatible with all
>>> existing readers and OS's, unlike .EXR, so it's a no brainer for us.
>>>
>>>
>>>
>>> Best regards,
>>>
>>> -Rich
>>>
>>>
>>>
>>> On Mon, Nov 13, 2023 at 8:35 AM Leonard Rosenthol <lrosenth@adobe.com>
>>> wrote:
>>>
>>> Richard – very interesting work, thanks for sharing!
>>>
>>>
>>>
>>> However, it isn’t in line with where the rest of the industry is going
>>> with respect to HDR imaging.  Currently ISO TC 42, TC 171, TC 130, JTC 1/SC
>>> 29 and the ICC are jointly working on a set of proposals for how to encode
>>> various aspects of HDR data – including gain maps – into common formats and
>>> technologies (including PNG).  The groups will be meeting in Tokyo in
>>> February for a full day working session to continue to move this work
>>> forward.  I believe that Chris Lilley will be there (perhaps remotely?) as
>>> well as myself.
>>>
>>>
>>>
>>> We already have some complex logic in PNG with respect to color and HDR
>>> (cICP vs. iCCP vs. gAMA vs sRGB vs ….).  Before introducing another
>>> PNG-specific approach, we should see where the industry is going and see if
>>> we can align with them first (which I believe we can/should).
>>>
>>>
>>>
>>> Leonard
>>>
>>>
>>>
>>> *From: *Richard Geldreich <rich@binomial.info>
>>> *Date: *Friday, November 10, 2023 at 11:05 PM
>>> *To: *public-png@w3.org <public-png@w3.org>
>>> *Subject: *Re: backwards compatible half-float PNG test app
>>> source+examples
>>>
>>> *EXTERNAL: Use caution when clicking on links or opening attachments.*
>>>
>>>
>>>
>>> Here's how the half float pixel values are losslessly recovered from
>>> 16-bit/component HDR PNG images in the "hdrpng" example code. This is the
>>> entire decoding procedure.
>>>
>>>
>>>
>>> This code extracts the low and high bytes from the stored PNG's 16-bpp
>>> pixels, remaps the high byte through the lookup table read from the "hdRa"
>>> ancillary chunk (below), shifts right the 16-bit value from [0,15] bits
>>> (this is typically 0, but for darker images it may be a small # of bits),
>>> and then rotates the sign bit back into the MSB from the LSB of the 16-bit
>>> value. This standard half float value can then be processed in an HDR
>>> pipeline.
>>>
>>>
>>>
>>> The lossless encoder that takes half-floats and outputs 16-bit PNG
>>> values is more complex, but not by much. It has to determine the shift
>>> amount by examining all the half floats in the image and a high byte
>>> histogram, and then compute a 256 entry lookup table using some sort of
>>> tone mapping algorithm. The table must be computed in a way that results in
>>> no loss.
>>>
>>>
>>>
>>>
>>>
>>> On Fri, Nov 10, 2023 at 4:48 AM Richard Geldreich <rich@binomial.info>
>>> wrote:
>>>
>>> If you want to see these example HDR .PNG's, unpacked to HDR .EXR files,
>>> using an in-browser HDR viewer, I've unpacked them (using the example
>>> hdrpng tool) to github here:
>>>
>>> https://github.com/richgel999/png16/tree/main/bin/unpacked
>>>
>>>
>>>
>>> The EXR viewer app (it's pretty good - I use it for testing on SDR
>>> monitors):
>>>
>>> https://viewer.openhdr.org/
>>>
>>>
>>>
>>>
>>>
>>> On Fri, Nov 10, 2023 at 4:34 AM Richard Geldreich <rich@binomial.info>
>>> wrote:
>>>
>>> I've found a lossless and trivally invertible transform that takes
>>> half-float HDR values (typically read from .HDR or .EXR images) and packs
>>> them to 16-bit unsigned pixels that are completely compatible with existing
>>> non-HDR aware PNG software. It uses a simple invertible and lossless global
>>> tone mapping operator that operates directly on the half-float values. Old
>>> readers view and see these PNG's as 16-bit PNG's (so usually 48bpp per
>>> pixel for RGB).
>>>
>>>
>>>
>>> The high bytes are tone mapped so the 16bpp images appear passable to
>>> existing readers. New readers can parse a small ~257 byte ancillary chunk
>>> which contains a byte remapping table used to remap the high bytes of the
>>> 16-bit components in the PNG image back to half-floats. (It's a little bit
>>> more complex than this to deal with signed floats and very low values, but
>>> that's the gist of it.) It's lossless for all valid half-float values
>>> (normals, denormals, signed). I filter out any NaN's/Inf"s in this test.
>>>
>>>
>>>
>>> You can see a bunch of example 48-bpp PNG's packed from .EXR images in
>>> this way here:
>>>
>>> https://github.com/richgel999/png16/tree/main/bin
>>>
>>>
>>>
>>> The C++ source to the example "hdrpng" tool is here. The example app
>>> excluding image reading/writing is only ~450 lines of code. (It currently
>>> compiles with VS 2022 under Windows - I'll add a Linux cmake file next.)
>>>
>>> https://github.com/richgel999/png16/
>>>
>>>
>>>
>>> It uses the popular open source lodepng library, unmodified, to write
>>> and read 16-bit PNG files and manipulate the ancillary "hdRa" chunk.
>>> hdrpng supports packing and tone mapping .EXR images to .PNG, unpacking HDR
>>> .PNG to .EXR, and an .EXR file comparison mode to verify that the half
>>> float values can be 100% recovered from the PNG file with no loss. The tone
>>> mapping in this example is automatic.
>>>
>>>
>>>
>>> I've tested the resulting 48bpp PNG files with pngcheck, Windows
>>> Explorer, Chrome, Paint Shop Pro, several tools, and the Windows Photo
>>> viewer app. So far, so good - they all look fine.
>>>
>>>
>>>
>>> Here are a couple example 48bpp PNG files. (Not sure what gmail will do
>>> to them, but this is what they look like.) These are how existing PNG
>>> readers view these HDR files. The half-float data is 100% preserved in
>>> these .PNG files, so HDR capable viewers are able to retrieve the original
>>> half-float pixels and do their own tone mapping or HDR processing.
>>>
>>>
>>>
>>>
>>>
>>> These PNG's validate successfully using pngcheck, because they are
>>> completely standard PNG files that any reader can load:
>>>
>>>
>>>
>>>