Anatomy of a UDeep file

A UDeep figure is an ordinary image file with one thing added: a declaration of how its data became the picture, stored where the image format keeps metadata (profile 8.8). Below, the two valid conformance files are taken apart as they are. Both show the same synthetic data, a ramp with one bright spot, 40 by 60 values; the page is made from the files, so it shows what they hold.

A gain-map JPEG

valid.jpg, 5,009 bytes: two JPEG images in one file. The first is the ordinary figure, which any program shows. The second, after it, is the gain map, which a screen that can show HDR uses to make the bright parts brighter than white. The declaration is in the first image's XMP segment, beside the gain map's own description.

the declaration's segmentother metadataimage data

In a figure this small (60 by 40 pixels) the metadata is a large share of the file. In the 18 HDR figures of this site's gallery, the declaration is 1.5 to 1.7 KiB of files from 184 KiB to 1 MiB, under 1 % of each.

The primary image: the ordinary figure

Part	At byte	Bytes	What it holds
`SOI`	0	2	start of the primary image
`APP1`	2	1,790	XMP: the gain map's container description (`hdrgm`, `Container`) and the UDeep declaration (profile 8.8)
`APP2`	1,792	36	ISO 21496-1 gain-map metadata, binary: its version
`APP2`	1,828	90	Multi-Picture Format index: where the gain-map image starts, and its size
`APP0`	1,918	18	JFIF header
`SOF0`	1,936	19	frame: 60 × 40 pixels, 3 components, 8-bit
`DQT`	1,955	69	quantisation table
`DQT`	2,024	69	quantisation table
`DHT`	2,093	33	Huffman table
`DHT`	2,126	183	Huffman table
`DHT`	2,309	33	Huffman table
`DHT`	2,342	183	Huffman table
`SOS`	2,525	760	start of scan, then the coded image data: the picture itself: the fair SDR figure, the one every screen and print shows
`EOI`	3,285	2	end of the primary image

The gain-map image

Part	At byte	Bytes	What it holds
`SOI`	3,287	2	start of the gain-map image
`APP1`	3,289	478	XMP: the gain map's own parameters (`hdrgm`)
`APP2`	3,767	93	ISO 21496-1 gain-map metadata, binary: the gain map's parameters
`APP0`	3,860	18	JFIF header
`SOF0`	3,878	19	frame: 60 × 40 pixels, 3 components, 8-bit
`DQT`	3,897	69	quantisation table
`DQT`	3,966	69	quantisation table
`DHT`	4,035	33	Huffman table
`DHT`	4,068	183	Huffman table
`DHT`	4,251	33	Huffman table
`DHT`	4,284	183	Huffman table
`SOS`	4,467	540	start of scan, then the coded image data: the gain map: how much brighter each pixel may be on an HDR screen
`EOI`	5,007	2	end of the gain-map image

A 16-bit PQ PNG

record.png, 4,603 bytes: the same figure as a data record, fine enough that it reads back as the data. It has no ordinary SDR picture of its own, so its declaration names the one that goes with it, by digest.

Part	At byte	Bytes	What it holds
`signature`	0	8	the eight bytes every PNG starts with
`IHDR`	8	25	60 × 40 pixels, 16-bit, RGB
`cICP`	33	16	colour signalling 9, 16, 0, 1: BT.2020 primaries, the PQ transfer function, RGB, full range (profile 9.1)
`iTXt`	49	1,261	`XML:com.adobe.xmp`, not compressed: the UDeep declaration (profile 8.8), naming its SDR figure (9.1.2)
`IDAT × 1`	1,310	3,281	the image: PQ codes, 16 bits a channel, deflated
`IEND`	4,591	12	the end

The declaration

The properties the JPEG declares, in the order the reference writes them, each with its meaning from the namespace page. The PNG's declaration is the same but for what depends on the encoding: the widest step, the rounding against the noise, and that it is a data record.

Property	Value in `valid.jpg`	Meaning
`udeep:DataUnit`	`counts`	unit of the data values
`udeep:Stretch`	`linear`	`linear`, `log`, `asinh` or `proportional`
`udeep:RangeLow`	`0`	the declared range, in data units after `StretchOffset` has been subtracted (8.3)
`udeep:RangeHigh`	`100`	the declared range, in data units after `StretchOffset` has been subtracted (8.3)
`udeep:Knee`	`50`	data value mapped to diffuse white: the data value itself, before `StretchOffset` is subtracted (8.3)
`udeep:KneeRule`	`preset; half of full scale`	`physical`, `statistical`, `preset` or `percentile`, then a description
`udeep:ReferenceHeadroom`	`2`	headroom the figure was authored for (linear ratio), greater than 1
`udeep:MaxDataStep`	`0.34332799267566827`	largest Δx_q over the declared range, in data units
`udeep:NoiseModel`	`constant`	`constant`, `poisson`, `relative`, `tolerance` or `unknown` (§5.2)
`udeep:NoiseParameter`	`1`	the model's parameter (§5.2), greater than zero; required unless the model is `unknown`, and absent then
`udeep:StretchWidth`	`1`	the asinh width
`udeep:StretchOffset`	`0`	the value subtracted from the data before the stretch
`udeep:AboveKnee`	`hermite; exit=1`	form of the segment above the knee, with its parameters: `hermite; exit=`s (§6.7)
`udeep:Masked`	`0`	number of pixels in the validity mask
`udeep:DerivedMasked`	`0`	the number of pixels that rule holds at white, beyond the validity mask
`udeep:Record`	`rendering`	`data-record` or `rendering` (§7.3)
`udeep:Preset`	`extend`	optional: `highlight`, `extend` or another named preset (informative)
`udeep:ColourMap`	`fire/3`	the map from light to colour (§6.10): `grey`, or a registered map as `name/version` (for example `fire/2`)
`udeep:RoundingToNoise`	`0.09911025449581548`	optional: the worst value over the declared range of the rounding error divided by the stated uncertainty (§7.2.1); omitted when the noise model is `unknown`
`udeep:DataDigest`	`sha256:c9f0ab72273ef44158a82022efea27d2c6124fd266da4210d7107fcd07c7e0c3`	optional: the digest of the array that was mapped, as defined in 8.6
`udeep:DataSource`	`doi:10.0000/udeep-conformance-example`	optional: where the data is kept: `doi:…` where the data has one, otherwise a URL, otherwise a description; several are separated by `;`
`udeep:SdrDigest`	`sha256:4242032dcdc7fe5a08c277a0807c36a9774468b1e9d428752427eafbbc994de8`	optional, except in a file with no SDR base of its own, which MUST declare it (9.1.2): the digest of the fair SDR rendering's 8-bit RGB pixels, before any lossy coding, as defined in 8.6
`udeep:DataDigestSha3`	`sha3-256:7c72d7eb969b2ea7f559bf2ef3d48639378d03090d73cbc3f2c6025f4cb95d93`	optional, and only with `DataDigest`: the same bytes digested with SHA3-256 (8.6), a hash function of another design, so that the link to the data holds if either function is ever broken
`udeep:SdrDigestSha3`	`sha3-256:8871bba8d6001e8f0525634bd1a8c8c8bc1b2b011fe3ca097799394f33099dce`	optional, and only with `SdrDigest`: the SDR figure's bytes digested with SHA3-256 (8.6)
`udeep:Version`	`0.3`	profile version, `"0.3"`
`udeep:ImageState`	`data-referred`	`data-referred`

The whole XMP packet, one attribute to a line

<x:xmpmeta
    xmlns:x="adobe:ns:meta/"
    x:xmptk="udeep">
<rdf:RDF
    xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#">
<rdf:Description
    rdf:about=""
    xmlns:Container="http://ns.google.com/photos/1.0/container/"
    xmlns:Item="http://ns.google.com/photos/1.0/container/item/"
    xmlns:hdrgm="http://ns.adobe.com/hdr-gain-map/1.0/"
    hdrgm:Version="1.0"><Container:Directory><rdf:Seq><rdf:li
    rdf:parseType="Resource"><Container:Item
    Item:Semantic="Primary"
    Item:Mime="image/jpeg"/></rdf:li><rdf:li
    rdf:parseType="Resource"><Container:Item
    Item:Semantic="GainMap"
    Item:Mime="image/jpeg"
    Item:Length="1722"/></rdf:li></rdf:Seq></Container:Directory></rdf:Description>
<rdf:Description
    rdf:about=""
    xmlns:udeep="https://ultradeeprange.github.io/udeep/ns/0.3/"
    udeep:DataUnit="counts"
    udeep:Stretch="linear"
    udeep:RangeLow="0"
    udeep:RangeHigh="100"
    udeep:Knee="50"
    udeep:KneeRule="preset; half of full scale"
    udeep:ReferenceHeadroom="2"
    udeep:MaxDataStep="0.34332799267566827"
    udeep:NoiseModel="constant"
    udeep:NoiseParameter="1"
    udeep:StretchWidth="1"
    udeep:StretchOffset="0"
    udeep:AboveKnee="hermite; exit=1"
    udeep:Masked="0"
    udeep:DerivedMasked="0"
    udeep:Record="rendering"
    udeep:Preset="extend"
    udeep:ColourMap="fire/3"
    udeep:RoundingToNoise="0.09911025449581548"
    udeep:DataDigest="sha256:c9f0ab72273ef44158a82022efea27d2c6124fd266da4210d7107fcd07c7e0c3"
    udeep:DataSource="doi:10.0000/udeep-conformance-example"
    udeep:SdrDigest="sha256:4242032dcdc7fe5a08c277a0807c36a9774468b1e9d428752427eafbbc994de8"
    udeep:DataDigestSha3="sha3-256:7c72d7eb969b2ea7f559bf2ef3d48639378d03090d73cbc3f2c6025f4cb95d93"
    udeep:SdrDigestSha3="sha3-256:8871bba8d6001e8f0525634bd1a8c8c8bc1b2b011fe3ca097799394f33099dce"
    udeep:Version="0.3"
    udeep:ImageState="data-referred"/>
</rdf:RDF>
</x:xmpmeta>

Where it must be, and what a reader does

One declaration a file, in one XMP packet. A reader takes it only from the place below, and refuses a file that holds two.
In a gain-map JPEG, the primary image's XMP segment (APP1), never the gain-map image's. One segment holds at most 65,533 bytes; this declaration is 1,757 bytes with its packet.
In a PNG, an iTXt chunk named XML:com.adobe.xmp, not compressed, before the image data. In an SVG, the <metadata> element.
Properties are written as attributes; a reader also takes them written as elements, as other XMP tools may rewrite them.
A signature, when there is one, is added after the declaration, so that it covers it.

The conformance files break these rules one at a time, with the verdict a validator must give.