Portable Network Graphics (PNG) Specification (Third Edition)

All PNG images contain a single static image.

Some PNG images — called Animated PNG (APNG) — also contain a frame-based animation sequence, the animated image. The first frame of this may be — but need not be — the static image. Non-animation-capable displays (such as printers) will display the static image rather than the animation sequence.

The static image, and each individual frame of an animated image, corresponds to a reference image and is stored as a PNG image.

This specification specifies the PNG datastream, and places some requirements on PNG encoders, which generate PNG datastreams, PNG decoders, which interpret PNG datastreams, and PNG editors, which transform one PNG datastream into another. It does not specify the interface between an application and either a PNG encoder, decoder, or editor. The precise form in which an image is presented to an encoder or delivered by a decoder is not specified. Four kinds of image are distinguished.

Source image

The source image is the image presented to a PNG encoder.

Reference image

The reference image, which only exists conceptually, is a rectangular array of rectangular pixels, all having the same width and height, and all containing the same number of unsigned integer samples, either three (red, green, blue) or four (red, green, blue, alpha). The array of all samples of a particular kind (red, green, blue, or alpha) is called a channel. Each channel has a sample depth in the range 1 to 16, which is the number of bits used by every sample in the channel. Different channels may have different sample depths. The red, green, and blue samples determine the intensities of the red, green, and blue components of the pixel's colour; if they are all zero, the pixel is black, and if they all have their maximum values (2^sampledepth-1), the pixel is white. The alpha sample determines a pixel's degree of opacity, where zero means fully transparent and the maximum value means fully opaque. In a three-channel reference image all pixels are fully opaque. (It is also possible for a four-channel reference image to have all pixels fully opaque; the difference is that the latter has a specific alpha sample depth, whereas the former does not.) Each horizontal row of pixels is called a scanline. Pixels are ordered from left to right within each scanline, and scanlines are ordered from top to bottom. Every reference image can be represented exactly by a PNG datastream and every PNG datastream can be converted into a reference image. A PNG encoder may transform the source image directly into a PNG image, but conceptually it first transforms the source image into a reference image, then transforms the reference image into a PNG image. Depending on the type of source image, the transformation from the source image to a reference image may require the loss of information. That transformation is beyond the scope of this International Standard. The reference image, however, can always be recovered exactly from a PNG datastream.

PNG image

The PNG image is obtained from the reference image by a series of transformations: alpha separation, indexing, RGB merging, alpha compaction, and sample depth scaling. Five types of PNG image are defined (see 6.1 Colour types and values). (If the PNG encoder actually transforms the source image directly into the PNG image, and the source image format is already similar to the PNG image format, the encoder may be able to avoid doing some of these transformations.) Although not all sample depths in the range 1 to 16 bits are explicitly supported in the PNG image, the number of significant bits in each channel of the reference image may be recorded. All channels in the PNG image have the same sample depth. A PNG encoder generates a PNG datastream from the PNG image. A PNG decoder takes the PNG datastream and recreates the PNG image.

Delivered image

The delivered image is constructed from the PNG image obtained by decoding a PNG datastream. No specific format is specified for the delivered image. A viewer presents an image to the user as close to the appearance of the original source image as it can achieve.

The relationships between the four kinds of image are illustrated in Figure 1.

The relationships between samples, channels, pixels, and sample depth are illustrated in Figure 2.

The RGB colour space in which colour samples are situated may be specified in one of four ways:

1. by CICP image format signaling metadata;
2. by an ICC profile;
3. by specifying explicitly that the colour space is sRGB when the samples conform to this colour space;
4. by specifying a gamma value and the 1931 CIE x,y chromaticities of the red, green, and blue primaries used in the image and the reference white point.

For high-end applications the first two methods provides the most flexibility and control. The third method enables one particular, but extremely common, colour space to be indicated. The fourth method, which was standardized before ICC profiles were widely adopted, enables the exact chromaticities of the RGB data to be specified, along with the gamma correction to be applied (see C. Gamma and chromaticity). However, colour-aware applications will prefer one of the first three methods, while colour-unaware applications will typically ignore all four methods.

Gamma correction is not applied to the alpha channel, if present. Alpha samples always represent a linear fraction of full opacity.

Mastering metadata may also be provided

The transformation of the reference image results in one of five types of PNG image (see Figure 6) :

Truecolour with alpha

Each pixel consists of four samples: red, green, blue and alpha.

Greyscale with alpha

Each pixel consists of two samples: a grey sample and an alpha sample.

Truecolour

Each pixel consists of a triplet of samples: red, green, blue. An optional alpha channel can be specified as a single triplet of red, green, blue samples: pixels of the image whose red, green, blue samples are identical to the red, green, blue samples of the alpha channel are fully transparent; others are fully opaque. If the alpha channel is not present, all pixels are fully opaque.

Greyscale

Each pixel consists of a single grey sample, which represents overall luminance (on a scale from black to white). An optional alpha channel can be specified as a single grey sample: pixels of the image whose grey sample is identical to the grey sample of the alpha channel are fully transparent; others are fully opaque. If the alpha channel is not present, all pixels are fully opaque.

Indexed-colour

Each pixel consists of an index into a palette (and into an associated table of alpha values, if present).

The format of each pixel depends on the PNG image type and the bit depth. For PNG image types other than indexed-colour, the bit depth specifies the number of bits per sample, not the total number of bits per pixel. For indexed-colour images, the bit depth specifies the number of bits in each palette index, not the sample depth of the colours in the palette or alpha table. Within the pixel the samples appear in the following order, depending on the PNG image type.

1. Truecolour with alpha: red, green, blue, alpha.
2. Greyscale with alpha: grey, alpha.
3. Truecolour: red, green, blue.
4. Greyscale: grey.
5. Indexed-colour: palette index.

Ancillary information may be associated with an image. Decoders may ignore all or some of the ancillary information. The types of ancillary information provided are described in Table 1.

Errors in a PNG datastream fall into two general classes:

1. transmission errors or damage to a computer file system, which tend to corrupt much or all of the datastream;
2. syntax errors, which appear as invalid values in chunks, or as missing or misplaced chunks. Syntax errors can be caused not only by encoding mistakes, but also by the use of registered or private values, if those values are unknown to the decoder.

PNG decoders should detect errors as early as possible, recover from errors whenever possible, and fail gracefully otherwise. The error handling philosophy is described in detail in 13.1 Error handling.

The PNG datastream consists of a PNG signature followed by a sequence of chunks. It is the result of encoding a PNG image.

The first eight bytes of a PNG datastream always contain the following hexadecimal values:

89 50 4E 47 0D 0A 1A 0A

This signature indicates that the remainder of the datastream contains a single PNG image, consisting of a series of chunks beginning with an IHDR chunk and ending with an IEND chunk.

This signature differentiates a PNG datastream from other types of datastream and allows early detection of some transmission errors.

Each chunk consists of three or four fields (see Figure 10). The meaning of the fields is described in Table 4. The chunk data field may be empty.

The chunk data length may be any number of bytes up to the maximum; therefore, implementors cannot assume that chunks are aligned on any boundaries larger than bytes.

Chunk types are chosen to be meaningful names when the bytes of the chunk type are interpreted as ISO 646 letters [ISO646]. Chunk types are assigned so that a decoder can determine some properties of a chunk even when the type is not recognized. These rules allow safe, flexible extension of the PNG format, by allowing a PNG decoder to decide what to do when it encounters an unknown chunk.

The naming rules are normally of interest only when the decoder does not recognize the chunk's type, as specified at 13. PNG decoders and viewers.

Four bits of the chunk type, the property bits, namely bit 5 (value 32) of each byte, are used to convey chunk properties. This choice means that a human can read off the assigned properties according to whether the letter corresponding to each byte of the chunk type is uppercase (bit 5 is 0) or lowercase (bit 5 is 1).

The property bits are an inherent part of the chunk type, and hence are fixed for any chunk type. Thus, CHNK and cHNk would be unrelated chunk types, not the same chunk with different properties.

The semantics of the property bits are defined in Table 5.

The hypothetical chunk type "cHNk" has the property bits:

cHNk  <-- 32 bit chunk type represented in text form
||||
|||+- Safe-to-copy bit is 1 (lower case letter; bit 5 is 1)
||+-- Reserved bit is 0     (upper case letter; bit 5 is 0)
|+--- Private bit is 0      (upper case letter; bit 5 is 0)
+---- Ancillary bit is 1    (lower case letter; bit 5 is 1)

Therefore, this name represents an ancillary, public, safe-to-copy chunk.

CRC fields are calculated using standardized CRC methods with pre and post conditioning, as defined by [ISO-3309] and [ITU-T-V.42]. The CRC polynomial employed— which is identical to that used in the GZIP file format specification [RFC1952]— is

x³² + x²⁶ + x²³ + x²² + x¹⁶ + x¹² + x¹¹ + x¹⁰ + x⁸ + x⁷ + x⁵ + x⁴ + x² + x + 1

In PNG, the 32-bit CRC is initialized to all 1's, and then the data from each byte is processed from the least significant bit (1) to the most significant bit (128). After all the data bytes are processed, the CRC is inverted (its ones complement is taken). This value is transmitted (stored in the datastream) MSB first. For the purpose of separating into bytes and ordering, the least significant bit of the 32-bit CRC is defined to be the coefficient of the x31 term.

Practical calculation of the CRC often employs a precalculated table to accelerate the computation. See D. Sample CRC implementation.

The constraints on the positioning of the individual chunks are listed in Table 6 and illustrated diagrammatically for static images in Figure 11 and Figure 12, for animated images where the static image forms the first frame in Figure 13 and Figure 14, and for animated images where the static image is not part of the animation in Figure 15 and Figure 16. These lattice diagrams represent the constraints on positioning imposed by this specification. The lines in the diagrams define partial ordering relationships. Chunks higher up shall appear before chunks lower down. Chunks which are horizontally aligned and appear between two other chunk types (higher and lower than the horizontally aligned chunks) may appear in any order between the two higher and lower chunk types to which they are connected. The superscript associated with the chunk type is defined in Table 7. It indicates whether the chunk is mandatory, optional, or may appear more than once. A vertical bar between two chunk types indicates alternatives.

Values greater than or equal to 128 in the following fields are private field values:

• compression method
• interlace method
• filter method

These private field values are neither defined nor reserved by this specification.

Private field values MAY be used for experimental or private semantics.

Private field values SHOULD NOT appear in publicly available software or datastreams since they can result in datastreams that are unreadable by PNG decoders as detailed at 13. PNG decoders and viewers.

As explained in 4.5 PNG image there are five types of PNG image. Corresponding to each type is a colour type, which is the sum of the following values: 1 (palette used), 2 (truecolour used) and 4 (alpha used). greyscale and truecolour images may have an explicit alpha channel. The PNG image types and corresponding colour types are listed in Table 8.

The allowed bit depths and sample depths for each PNG image type are listed in Image header.

Greyscale samples represent luminance if the transfer curve is indicated (by gAMA, sRGB, or iCCP) or device-dependent greyscale if not. RGB samples represent calibrated colour information if the colour space is indicated (by gAMA and cHRM, or sRGB, or iCCP, or uncalibrated device-dependent colour if not.

Sample values are not necessarily proportional to light intensity; the gAMA chunk specifies the relationship between sample values and display output intensity. Viewers are strongly encouraged to compensate properly. See 4.3 Colour spaces, 13.13 Decoder gamma handling and C. Gamma and chromaticity.

In a PNG datastream transparency may be represented in one of four ways, depending on the PNG image type (see alpha separation and alpha compaction).

1. Truecolour with alpha, greyscale with alpha: an alpha channel is part of the image array.
2. truecolour, greyscale: A tRNS chunk contains a single pixel value distinguishing the fully transparent pixels from the fully opaque pixels.
3. Indexed-colour: A tRNS chunk contains the alpha table that associates an alpha sample with each palette entry.
4. truecolour, greyscale, indexed-colour: there is no tRNS chunk present and all pixels are fully opaque.

An alpha channel included in the image array has 8-bit or 16-bit samples, the same size as the other samples. The alpha sample for each pixel is stored immediately following the greyscale or RGB samples of the pixel. An alpha value of zero represents full transparency, and a value of 2^sampledepth - 1 represents full opacity. Intermediate values indicate partially transparent pixels that can be composited against a background image to yield the delivered image.

The colour values in a pixel are not premultiplied by the alpha value assigned to the pixel. This rule is sometimes called "unassociated" or "non-premultiplied" alpha. (Another common technique is to store sample values premultiplied by the alpha value; in effect, such an image is already composited against a black background. PNG does not use premultiplied alpha. In consequence an image editor can take a PNG image and easily change its transparency.) See 12.3 Alpha channel creation and 13.16 Alpha channel processing.

All integers that require more than one byte shall be in network byte order (as illustrated in Figure 17 ): the most significant byte comes first, then the less significant bytes in descending order of significance (MSB LSB for two-byte integers, MSB B2 B1 LSB for four-byte integers). The highest bit (value 128) of a byte is numbered bit 7; the lowest bit (value 1) is numbered bit 0. Values are unsigned unless otherwise noted. Values explicitly noted as signed are represented in two's complement notation.

PNG four-byte unsigned integers are limited to the range 0 to 2³¹-1 to accommodate languages that have difficulty with unsigned four-byte values.

A PNG image (or pass, see 8. Interlacing and pass extraction) is a rectangular pixel array, with pixels appearing left-to-right within each scanline, and scanlines appearing top-to-bottom. The size of each pixel is determined by the number of bits per pixel.

Pixels within a scanline are always packed into a sequence of bytes with no wasted bits between pixels. Scanlines always begin on byte boundaries. Permitted bit depths and colour types are restricted so that in all cases the packing is simple and efficient.

In PNG images of colour type 0 (greyscale) each pixel is a single sample, which may have precision less than a byte (1, 2, or 4 bits). These samples are packed into bytes with the leftmost sample in the high-order bits of a byte followed by the other samples for the scanline.

In PNG images of colour type 3 (indexed-colour) each pixel is a single palette index. These indices are packed into bytes in the same way as the samples for colour type 0.

When there are multiple pixels per byte, some low-order bits of the last byte of a scanline may go unused. The contents of these unused bits are not specified.

PNG images that are not indexed-colour images may have sample values with a bit depth of 16. Such sample values are in network byte order (MSB first, LSB second). PNG permits multi-sample pixels only with 8 and 16-bit samples, so multiple samples of a single pixel are never packed into one byte.

A filter method is a transformation applied to an array of scanlines with the aim of improving their compressibility.

PNG standardizes one filter method and several filter types that may be used to prepare image data for compression. It transforms the byte sequence into an equal length sequence of bytes preceded by a filter type byte (see Figure 18 for an example).

The encoder shall use only a single filter method for an interlaced PNG image, but may use different filter types for each scanline in a reduced image. An intelligent encoder can switch filters from one scanline to the next. The method for choosing which filter to employ is left to the encoder.

The filter type byte is not considered part of the image data, but it is included in the datastream sent to the compression step. See 9. Filtering.

Two interlace methods are defined in this International Standard, methods 0 and 1. Other values of interlace method are reserved for future standardization.

With interlace method 0, the null method, pixels are extracted sequentially from left to right, and scanlines sequentially from top to bottom. The interlaced PNG image is a single reduced image.

Interlace method 1, known as Adam7, defines seven distinct passes over the image. Each pass transmits a subset of the pixels in the reference image. The pass in which each pixel is transmitted (numbered from 1 to 7) is defined by replicating the following 8-by-8 pattern over the entire image, starting at the upper left corner:

1 6 4 6 2 6 4 6
7 7 7 7 7 7 7 7
5 6 5 6 5 6 5 6
7 7 7 7 7 7 7 7
3 6 4 6 3 6 4 6
7 7 7 7 7 7 7 7
5 6 5 6 5 6 5 6
7 7 7 7 7 7 7 7

Figure 4.8 shows the seven passes of interlace method 1. Within each pass, the selected pixels are transmitted left to right within a scanline, and selected scanlines sequentially from top to bottom. For example, pass 2 contains pixels 4, 12, 20, etc. of scanlines 0, 8, 16, etc. (where scanline 0, pixel 0 is the upper left corner). The last pass contains all of scanlines 1, 3, 5, etc. The transmission order is defined so that all the scanlines transmitted in a pass will have the same number of pixels; this is necessary for proper application of some of the filters. The interlaced PNG image consists of a sequence of seven reduced images. For example, if the PNG image is 16 by 16 pixels, then the third pass will be a reduced image of two scanlines, each containing four pixels (see figure 4.8).

Scanlines that do not completely fill an integral number of bytes are padded as defined in 7.2 Scanlines.

Filtering transforms the PNG image with the goal of improving compression. The overall process is depicted in Figure 7 while the specifics of serializing and filtering a scanline are shown in Figure 18.

PNG allows for a number of filter methods. All the reduced images in an interlaced image shall use a single filter method. Only filter method 0 is defined by this specification. Other filter methods are reserved for future standardization. Filter method 0 provides a set of five filter types, and individual scanlines in each reduced image may use different filter types.

PNG imposes no additional restriction on which filter types can be applied to an interlaced PNG image. However, the filter types are not equally effective on all types of data. See 12.7 Filter selection.

Filtering transforms the byte sequence in a scanline to an equal length sequence of bytes preceded by the filter type. Filter type bytes are associated only with non-empty scanlines. No filter type bytes are present in an empty pass. See 13.10 Interlacing and progressive display.

Filters are applied to bytes, not to pixels, regardless of the bit depth or colour type of the image. The filters operate on the byte sequence formed by a scanline that has been represented as described in 7.2 Scanlines. If the image includes an alpha channel, the alpha data is filtered in the same way as the image data.

Filters may use the original values of the following bytes to generate the new byte value:

Figure 19 shows the relative positions of the bytes x, a, b, and c.

Filter method 0 defines five basic filter types as listed in Table 10. Orig(y) denotes the original (unfiltered) value of byte y. Filt(y) denotes the value after a filter type has been applied. Recon(y) denotes the value after the corresponding reconstruction function has been applied. The Paeth filter type PaethPredictor [Paeth] is defined below.

Filter method 0 specifies exactly this set of five filter types and this shall not be extended. This ensures that decoders need not decompress the data to determine whether it contains unsupported filter types: it is sufficient to check the filter method in 11.2.1 IHDR Image header.

For all filters, the bytes "to the left of" the first pixel in a scanline shall be treated as being zero. For filters that refer to the prior scanline, the entire prior scanline and bytes "to the left of" the first pixel in the prior scanline shall be treated as being zeroes for the first scanline of a reduced image.

To reverse the effect of a filter requires the decoded values of the prior pixel on the same scanline, the pixel immediately above the current pixel on the prior scanline, and the pixel just to the left of the pixel above.

Unsigned arithmetic modulo 256 is used, so that both the inputs and outputs fit into bytes. Filters are applied to each byte regardless of bit depth. The sequence of Filt values is transmitted as the filtered scanline.

The sum Orig(a) + Orig(b) shall be performed without overflow (using at least nine-bit arithmetic). floor() indicates that the result of the division is rounded to the next lower integer if fractional; in other words, it is an integer division or right shift operation.

The Paeth filter type computes a simple linear function of the three neighbouring pixels (left, above, upper left), then chooses as predictor the neighbouring pixel closest to the computed value. The algorithm used in this specification is an adaptation of the technique due to Alan W. Paeth [Paeth].

The PaethPredictor function is defined in the code below. The logic of the function and the locations of the bytes a, b, c, and x are shown in Figure 20. Pr is the predictor for byte x.

p = a + b - c
pa = abs(p - a)
pb = abs(p - b)
pc = abs(p - c)
if pa <= pb and pa <= pc then Pr = a
else if pb <= pc then Pr = b
else Pr = c
return Pr

The calculations within the PaethPredictor function shall be performed exactly, without overflow.

The order in which the comparisons are performed is critical and shall not be altered. The function tries to establish in which of the three directions (vertical, horizontal, or diagonal) the gradient of the image is smallest.

Exactly the same PaethPredictor function is used by both encoder and decoder.

Only PNG compression method 0 is defined by this International Standard. Other values of compression method are reserved for future standardization. PNG compression method 0 is deflate compression with a sliding window (which is an upper bound on the distances appearing in the deflate stream) of at most 32768 bytes. Deflate compression is derived from LZ77.

Deflate-compressed datastreams within PNG are stored in the zlib format, which has the structure:

zlib is specified at [rfc1950].

For PNG compression method 0, the zlib compression method/flags code shall specify method code 8 (deflate compression) and an LZ77 window size of not more than 32768 bytes. The zlib compression method number is not the same as the PNG compression method number in the IHDR chunk. The additional flags shall not specify a preset dictionary.

If the data to be compressed contain 16384 bytes or fewer, the PNG encoder may set the window size by rounding up to a power of 2 (256 minimum). This decreases the memory required for both encoding and decoding, without adversely affecting the compression ratio.

The compressed data within the zlib datastream are stored as a series of blocks, each of which can represent raw (uncompressed) data, LZ77-compressed data encoded with fixed Huffman codes, or LZ77-compressed data encoded with custom Huffman codes. A marker bit in the final block identifies it as the last block, allowing the decoder to recognize the end of the compressed datastream. Further details on the compression algorithm and the encoding are given in the deflate specification [rfc1951].

The check value stored at the end of the zlib datastream is calculated on the uncompressed data represented by the datastream. The algorithm used to calculate this is not the same as the CRC calculation used for PNG chunk CRC field values. The zlib check value is useful mainly as a cross-check that the deflate algorithms are implemented correctly. Verifying the individual PNG chunk CRCs provides confidence that the PNG datastream has been transmitted undamaged.

The sequence of filtered scanlines is compressed and the resulting data stream is split into IDAT chunks. The concatenation of the contents of all the IDAT chunks makes up a zlib datastream. This datastream decompresses to filtered image data.

It is important to emphasize that the boundaries between IDAT chunks are arbitrary and can fall anywhere in the zlib datastream. There is not necessarily any correlation between IDAT chunk boundaries and deflate block boundaries or any other feature of the zlib data. For example, it is entirely possible for the terminating zlib check value to be split across IDAT chunks.

Similarly, there is no required correlation between the structure of the image data (i.e., scanline boundaries) and deflate block boundaries or IDAT chunk boundaries. The complete filtered PNG image is represented by a single zlib datastream that is stored in a number of IDAT chunks.

PNG also uses compression method 0 in iTXt, iCCP, and zTXt chunks. Unlike the image data, such datastreams are not split across chunks; each such chunk contains an independent zlib datastream (see 10.1 Compression method 0).

This clause defines chunk used in this specification.

See C. Gamma and chromaticity for a brief introduction to gamma issues.

PNG encoders capable of full colour management will perform more sophisticated calculations than those described here and may choose to use the iCCP chunk. If it is known that the image samples conform to the sRGB specification [SRGB], encoders are strongly encouraged to write the sRGB chunk without performing additional gamma handling. In both cases it is recommended that an appropriate gAMA chunk be generated for use by PNG decoders that do not recognize the iCCP or sRGB chunks.

A PNG encoder has to determine:

1. what value to write in the gAMA chunk;
2. how to transform the provided image samples into the values to be written in the PNG datastream.

The value to write in the gAMA chunk is that value which causes a PNG decoder to behave in the desired way. See 13.13 Decoder gamma handling.

The transform to be applied depends on the nature of the image samples and their precision. If the samples represent light intensity in floating-point or high precision integer form (perhaps from a computer graphics renderer), the encoder may perform gamma encoding (applying a power function with exponent less than 1) before quantizing the data to integer values for inclusion in the PNG datastream. This results in fewer banding artifacts at a given sample depth, or allows smaller samples while retaining the same visual quality. An intensity level expressed as a floating-point value in the range 0 to 1 can be converted to a datastream image sample by:

integer_sample = floor((2sampledepth-1) * intensityencoding_exponent + 0.5)

If the intensity in the equation is the desired output intensity, the encoding exponent is the gamma value to be used in the gAMA chunk.

If the intensity available to the PNG encoder is the original scene intensity, another transformation may be needed. There is sometimes a requirement for the displayed image to have higher contrast than the original source image. This corresponds to an end-to-end transfer function from original scene to display output with an exponent greater than 1. In this case:

gamma = encoding_exponent/end_to_end_exponent

If it is not known whether the conditions under which the original image was captured or calculated warrant such a contrast change, it may be assumed that the display intensities are proportional to original scene intensities, i.e. the end-to-end exponent is 1 and hence:

gamma = encoding_exponent

If the image is being written to a datastream only, the encoder is free to choose the encoding exponent. Choosing a value that causes the gamma value in the gAMA chunk to be 1/2.2 is often a reasonable choice because it minimizes the work for a PNG decoder displaying on a typical video monitor.

Some image renderers may simultaneously write the image to a PNG datastream and display it on-screen. The displayed pixels should be gamma corrected for the display system and viewing conditions in use, so that the user sees a proper representation of the intended scene.

If the renderer wants to write the displayed sample values to the PNG datastream, avoiding a separate gamma encoding step for the datastream, the renderer should approximate the transfer function of the display system by a power function, and write the reciprocal of the exponent into the gAMA chunk. This will allow a PNG decoder to reproduce what was displayed on screen for the originator during rendering.

However, it is equally reasonable for a renderer to compute displayed pixels appropriate for the display device, and to perform separate gamma encoding for data storage and transmission, arranging to have a value in the gAMA chunk more appropriate to the future use of the image.

Computer graphics renderers often do not perform gamma encoding, instead making sample values directly proportional to scene light intensity. If the PNG encoder receives sample values that have already been quantized into integer values, there is no point in doing gamma encoding on them; that would just result in further loss of information. The encoder should just write the sample values to the PNG datastream. This does not imply that the gAMA chunk should contain a gamma value of 1.0 because the desired end-to-end transfer function from scene intensity to display output intensity is not necessarily linear. However, the desired gamma value is probably not far from 1.0. It may depend on whether the scene being rendered is a daylight scene or an indoor scene, etc.

When the sample values come directly from a piece of hardware, the correct gAMA value can, in principle, be inferred from the transfer function of the hardware and lighting conditions of the scene. In the case of video digitizers ("frame grabbers"), the samples are probably in the sRGB colour space, because the sRGB specification was designed to be compatible with modern video standards. Image scanners are less predictable. Their output samples may be proportional to the input light intensity since CCD sensors themselves are linear, or the scanner hardware may have already applied a power function designed to compensate for dot gain in subsequent printing (an exponent of about 0.57), or the scanner may have corrected the samples for display on a monitor. It may be necessary to refer to the scanner's manual or to scan a calibrated target in order to determine the characteristics of a particular scanner. It should be remembered that gamma relates samples to desired display output, not to scanner input.

Datastream format converters generally should not attempt to convert supplied images to a different gamma. The data should be stored in the PNG datastream without conversion, and the gamma value should be deduced from information in the source datastream if possible. Gamma alteration at datastream conversion time causes re-quantization of the set of intensity levels that are represented, introducing further roundoff error with little benefit. It is almost always better to just copy the sample values intact from the input to the output file.

If the source datastream describes the gamma characteristics of the image, a datastream converter is strongly encouraged to write a gAMA chunk. Some datastream formats specify the display exponent (the exponent of the function which maps image samples to display output rather than the other direction). If the source file's gamma value is greater than 1.0, it is probably a display exponent, and the reciprocal of this value should be used for the PNG gamma value. If the source file format records the relationship between image samples and a quantity other than display output, it will be more complex than this to deduce the PNG gamma value.

If a PNG encoder or datastream converter knows that the image has been displayed satisfactorily using a display system whose transfer function can be approximated by a power function with exponent display_exponent, the image can be marked as having the gamma value:

gamma = 1/display_exponent

It is better to write a gAMA chunk with a value that is approximately correct than to omit the chunk and force PNG decoders to guess an approximate gamma value. If a PNG encoder is unable to infer the gamma value, it is preferable to omit the gAMA chunk. If a guess has to be made this should be left to the PNG decoder.

gamma does not apply to alpha samples; alpha is always represented linearly.

See also 13.13 Decoder gamma handling.

See C. Gamma and chromaticity for references to colour issues.

PNG encoders capable of full colour management will perform more sophisticated calculations than those described here and may choose to use the iCCP chunk. If it is known that the image samples conform to the sRGB specification [SRGB], PNG encoders are strongly encouraged to use the sRGB chunk.

If it is possible for the encoder to determine the chromaticities of the source display primaries, or to make a strong guess based on the origin of the image, or the hardware running it, the encoder is strongly encouraged to output the cHRM chunk. If this is done, the gAMA chunk should also be written; decoders can do little with a cHRM chunk if the gAMA chunk is missing.

There are a number of recommendations and standards for primaries and white points, some of which are linked to particular technologies, for example the CCIR 709 standard [ITU-R-BT.709] and the SMPTE-C standard [SMPTE-170M].

There are three cases that need to be considered:

1. the encoder is part of the generation system;
2. the source image is captured by a camera or scanner;
3. the PNG datastream was generated by translation from some other format.

In the case of hand-drawn or digitally edited images, it is necessary to determine what monitor they were viewed on when being produced. Many image editing programs allow the type of monitor being used to be specified. This is often because they are working in some device-independent space internally. Such programs have enough information to write valid cHRM and gAMA chunks, and are strongly encouraged to do so automatically.

If the encoder is compiled as a portion of a computer image renderer that performs full-spectral rendering, the monitor values that were used to convert from the internal device-independent colour space to RGB should be written into the cHRM chunk. Any colours that are outside the gamut of the chosen RGB device should be mapped to be within the gamut; PNG does not store out-of-gamut colours.

If the computer image renderer performs calculations directly in device-dependent RGB space, a cHRM chunk should not be written unless the scene description and rendering parameters have been adjusted for a particular monitor. In that case, the data for that monitor should be used to construct a cHRM chunk.

A few image formats store calibration information, which can be used to fill in the cHRM chunk. For example, TIFF 6.0 files [TIFF-6.0] can optionally store calibration information, which if present should be used to construct the cHRM chunk.

Video created with recent video equipment probably uses the CCIR 709 primaries and D65 white point [ITU-R-BT.709], which are given in Table 28.

An older but still very popular video standard is SMPTE-C [SMPTE-170M] given in Table 29.

It is not recommended that datastream format converters attempt to convert supplied images to a different RGB colour space. The data should be stored in the PNG datastream without conversion, and the source primary chromaticities should be recorded if they are known. Colour space transformation at datastream conversion time is a bad idea because of gamut mismatches and rounding errors. As with gamma conversions, it is better to store the data losslessly and incur at most one conversion when the image is finally displayed.

See 13.14 Decoder colour handling.

The alpha channel can be regarded either as a mask that temporarily hides transparent parts of the image, or as a means for constructing a non-rectangular image. In the first case, the colour values of fully transparent pixels should be preserved for future use. In the second case, the transparent pixels carry no useful data and are simply there to fill out the rectangular image area required by PNG. In this case, fully transparent pixels should all be assigned the same colour value for best compression.

Image authors should keep in mind the possibility that a decoder will not support transparency control in full (see 13.16 Alpha channel processing). Hence, the colours assigned to transparent pixels should be reasonable background colours whenever feasible.

For applications that do not require a full alpha channel, or cannot afford the price in compression efficiency, the tRNS transparency chunk is also available.

If the image has a known background colour, this colour should be written in the bKGD chunk. Even decoders that ignore transparency may use the bKGD colour to fill unused screen area.

If the original image has premultiplied (also called "associated") alpha data, it can be converted to PNG's non-premultiplied format by dividing each sample value by the corresponding alpha value, then multiplying by the maximum value for the image bit depth, and rounding to the nearest integer. In valid premultiplied data, the sample values never exceed their corresponding alpha values, so the result of the division should always be in the range 0 to 1. If the alpha value is zero, output black (zeroes).

When encoding input samples that have a sample depth that cannot be directly represented in PNG, the encoder shall scale the samples up to a sample depth that is allowed by PNG. The most accurate scaling method is the linear equation:

output = floor((input * MAXOUTSAMPLE / MAXINSAMPLE) + 0.5)

where the input samples range from 0 to MAXINSAMPLE and the outputs range from 0 to MAXOUTSAMPLE (which is 2^sampledepth-1).

A close approximation to the linear scaling method is achieved by "left bit replication", which is shifting the valid bits to begin in the most significant bit and repeating the most significant bits into the open bits. This method is often faster to compute than linear scaling.

Assume that 5-bit samples are being scaled up to 8 bits. If the source sample value is 27 (in the range from 0-31), then the original bits are:

4 3 2 1 0
---------
1 1 0 1 1

Left bit replication gives a value of 222:

7 6 5 4 3  2 1 0
----------------
1 1 0 1 1  1 1 0
|=======|  |===|
    |      Leftmost Bits Repeated to Fill Open Bits
    |
Original Bits

which matches the value computed by the linear equation. Left bit replication usually gives the same value as linear scaling, and is never off by more than one.

A distinctly less accurate approximation is obtained by simply left-shifting the input value and filling the low order bits with zeroes. This scheme cannot reproduce white exactly, since it does not generate an all-ones maximum value; the net effect is to darken the image slightly. This method is not recommended in general, but it does have the effect of improving compression, particularly when dealing with greater-than-8-bit sample depths. Since the relative error introduced by zero-fill scaling is small at high sample depths, some encoders may choose to use it. Zero-fill shall not be used for alpha channel data, however, since many decoders will treat alpha values of all zeroes and all ones as special cases. It is important to represent both those values exactly in the scaled data.

When the encoder writes an sBIT chunk, it is required to do the scaling in such a way that the high-order bits of the stored samples match the original data. That is, if the sBIT chunk specifies a sample depth of S, the high-order S bits of the stored data shall agree with the original S-bit data values. This allows decoders to recover the original data by shifting right. The added low-order bits are not constrained. All the above scaling methods meet this restriction.

When scaling up source image data, it is recommended that the low-order bits be filled consistently for all samples; that is, the same source value should generate the same sample value at any pixel position. This improves compression by reducing the number of distinct sample values. This is not a mandatory requirement, and some encoders may choose not to follow it. For example, an encoder might instead dither the low-order bits, improving displayed image quality at the price of increasing file size.

In some applications the original source data may have a range that is not a power of 2. The linear scaling equation still works for this case, although the shifting methods do not. It is recommended that an sBIT chunk not be written for such images, since sBIT suggests that the original data range was exactly 0..2^S-1.

Suggested palettes may appear as sPLT chunks in any PNG datastream, or as a PLTE chunk in truecolour PNG datastreams. In either case, the suggested palette is not an essential part of the image data, but it may be used to present the image on indexed-colour display hardware. Suggested palettes are of no interest to viewers running on truecolour hardware.

When an sPLT chunk is used to provide a suggested palette, it is recommended that the encoder use the frequency fields to indicate the relative importance of the palette entries, rather than leave them all zero (meaning undefined). The frequency values are most easily computed as "nearest neighbour" counts, that is, the approximate usage of each RGBA palette entry if no dithering is applied. (These counts will often be available "for free" as a consequence of developing the suggested palette.) Because the suggested palette includes transparency information, it should be computed for the un-composited image.

Even for indexed-colour images, sPLT can be used to define alternative reduced palettes for viewers that are unable to display all the colours present in the PLTE chunk. If the PLTE chunk appears without the bKGD chunk in an image of colour type 6, the circumstances under which the palette was computed are unspecified.

An older method for including a suggested palette in a truecolour PNG datastream uses the PLTE chunk. If this method is used, the histogram (frequencies) should appear in a separate hIST chunk. The PLTE chunk does not include transparency information. Hence for images of colour type 6 (truecolour with alpha), it is recommended that a bKGD chunk appear and that the palette and histogram be computed with reference to the image as it would appear after compositing against the specified background colour. This definition is necessary to ensure that useful palette entries are generated for pixels having fractional alpha values. The resulting palette will probably be useful only to viewers that present the image against the same background colour. It is recommended that PNG editors delete or recompute the palette if they alter or remove the bKGD chunk in an image of colour type 6.

For images of colour type 2 (truecolour), it is recommended that the PLTE and hIST chunks be computed with reference to the RGB data only, ignoring any transparent-colour specification. If the datastream uses transparency (has a tRNS chunk), viewers can easily adapt the resulting palette for use with their intended background colour (see 13.17 Histogram and suggested palette usage).

For providing suggested palettes, the sPLT chunk is more flexible than the PLTE chunk in the following ways:

1. With sPLT multiple suggested palettes may be provided. A PNG decoder may choose an appropriate palette based on name or number of entries.
2. In a PNG datastream of colour type 6 (truecolour with alpha channel), the PLTE chunk represents a palette already composited against the bKGD colour, so it is useful only for display against that background colour. The sPLT chunk provides an un-composited palette, which is useful for display against backgrounds chosen by the PNG decoder.
3. Since the sPLT chunk is an ancillary chunk, a PNG editor may add or modify suggested palettes without being forced to discard unknown unsafe-to-copy chunks.
4. Whereas the sPLT chunk is allowed in PNG datastreams for colour types 0, 3, and 4 (greyscale and indexed-colour), the PLTE chunk cannot be used to provide reduced palettes in these cases.
5. More than 256 entries may appear in the sPLT chunk.

A PNG encoder that uses the sPLT chunk may choose to write a suggested palette represented by PLTE and hIST chunks as well, for compatibility with decoders that do not recognize the sPLT chunk.

This specification defines two interlace methods, one of which is no interlacing. Interlacing provides a convenient basis from which decoders can progressively display an image, as described in 13.10 Interlacing and progressive display.

For images of colour type 3 (indexed-colour), filter type 0 (None) is usually the most effective. Colour images with 256 or fewer colours should almost always be stored in indexed-colour format; truecolour format is likely to be much larger.

Filter type 0 is also recommended for images of bit depths less than 8. For low-bit-depth greyscale images, in rare cases, better compression may be obtained by first expanding the image to 8-bit representation and then applying filtering.

For truecolour and greyscale images, any of the five filters may prove the most effective. If an encoder uses a fixed filter, the Paeth filter type is most likely to be the best.

For best compression of truecolour and greyscale images, and if compression efficiency is valued over speed of compression, the recommended approach is adaptive filtering in which a filter type is chosen for each scanline. Each unique image will have a different set of filters which perform best for it. An encoder could try every combination of filters to find what compresses best for a given image. However, when an exhaustive search is unacceptable, here are some general heuristics which may perform well enough: compute the output scanline using all five filters, and select the filter that gives the smallest sum of absolute values of outputs. (Consider the output bytes as signed differences for this test.) This method usually outperforms any single fixed filter type choice.

Filtering according to these recommendations is effective in conjunction with either of the two interlace methods defined in this specification.

The encoder may divide the compressed datastream into IDAT chunks however it wishes. (Multiple IDAT chunks are allowed so that encoders may work in a fixed amount of memory; typically the chunk size will correspond to the encoder's buffer size.) A PNG datastream in which each IDAT chunk contains only one data byte is valid, though remarkably wasteful of space. (Zero-length IDAT chunks are also valid, though even more wasteful.)

A nonempty keyword shall be provided for each text chunk. The generic keyword "Comment" can be used if no better description of the text is available. If a user-supplied keyword is used, encoders should check that it meets the restrictions on keywords.

The iTXt chunk uses the UTF-8 encoding of Unicode and thus can store text in any language. The tEXt and zTXt chunks use the Latin-1 (ISO 8859-1) character encoding, which limits the range of characters that can be used in these chunks. Encoders should prefer iTXt to tEXt and zTXt chunks, in order to allow a wide range of characters without data loss. Encoders must convert characters that use local legacy character encodings to the appropriate encoding when storing text.

When creating iTXt chunks, encoders should follow UTF-8 encode in Encoding Standard.

Encoders should discourage the creation of single lines of text longer than 79 Unicode code points, in order to facilitate easy reading. It is recommended that text items less than 1024 bytes in size should be output using uncompressed text chunks. It is recommended that the basic title and author keywords be output using uncompressed text chunks. Placing large text chunks after the image data (after the IDAT chunks) can speed up image display in some situations, as the decoder will decode the image data first. It is recommended that small text chunks, such as the image title, appear before the IDAT chunks.

Errors in a PNG datastream will fall into two general classes, transmission errors and syntax errors (see 4.10 Error handling).

Examples of transmission errors are transmission in "text" or "ascii" mode, in which byte codes 13 and/or 10 may be added, removed, or converted throughout the datastream; unexpected termination, in which the datastream is truncated; or a physical error on a storage device, in which one or more blocks (typically 512 bytes each) will have garbled or random values. Some examples of syntax errors are an invalid value for a row filter, an invalid compression method, an invalid chunk length, the absence of a PLTE chunk before the first IDAT chunk in an indexed image, or the presence of multiple gAMA chunks. A PNG decoder should handle errors as follows:

1. Detect errors as early as possible using the PNG signature bytes and CRCs on each chunk. Decoders should verify that all eight bytes of the PNG signature are correct. A decoder can have additional confidence in the datastream's integrity if the next eight bytes begin an IHDR chunk with the correct chunk length. A CRC should be checked before processing the chunk data. Sometimes this is impractical, for example when a streaming PNG decoder is processing a large IDAT chunk. In this case the CRC should be checked when the end of the chunk is reached.
2. Recover from an error, if possible; otherwise fail gracefully. Errors that have little or no effect on the processing of the image may be ignored, while those that affect critical data shall be dealt with in a manner appropriate to the application.
3. Provide helpful messages describing errors, including recoverable errors.

Three classes of PNG chunks are relevant to this philosophy. For the purposes of this classification, an "unknown chunk" is either one whose type was genuinely unknown to the decoder's author, or one that the author chose to treat as unknown, because default handling of that chunk type would be sufficient for the program's purposes. Other chunks are called "known chunks". Given this definition, the three classes are as follows:

1. known chunks, which necessarily includes all of the critical chunks defined in this specification (IHDR, PLTE, IDAT, IEND)
2. unknown critical chunks (bit 5 of the first byte of the chunk type is 0)
3. unknown ancillary chunks (bit 5 of the first byte of the chunk type is 1)

See 5.4 Chunk naming conventions for a description of chunk naming conventions.

PNG chunk types are marked "critical" or "ancillary" according to whether the chunks are critical for the purpose of extracting a viewable image (as with IHDR, PLTE, and IDAT) or critical to understanding the datastream structure (as with IEND). This is a specific kind of criticality and one that is not necessarily relevant to every conceivable decoder. For example, a program whose sole purpose is to extract text annotations (for example, copyright information) does not require a viewable image but should decode UTF-8 correctly. Another decoder might consider the tRNS and gAMA chunks essential to its proper execution.

Syntax errors always involve known chunks because syntax errors in unknown chunks cannot be detected. The PNG decoder has to determine whether a syntax error is fatal (unrecoverable) or not, depending on its requirements and the situation. For example, most decoders can ignore an invalid IEND chunk; a text-extraction program can ignore the absence of IDAT; an image viewer cannot recover from an empty PLTE chunk in an indexed image but it can ignore an invalid PLTE chunk in a truecolour image; and a program that extracts the alpha channel can ignore an invalid gAMA chunk, but may consider the presence of two tRNS chunks to be a fatal error. Anomalous situations other than syntax errors shall be treated as follows:

1. Encountering an unknown ancillary chunk is never an error. The chunk can simply be ignored.
2. Encountering an unknown critical chunk is a fatal condition for any decoder trying to extract the image from the datastream. A decoder that ignored a critical chunk could not know whether the image it extracted was the one intended by the encoder.
3. A PNG signature mismatch, a CRC mismatch, or an unexpected end-of-stream indicates a corrupted datastream, and may be regarded as a fatal error. A decoder could try to salvage something from the datastream, but the extent of the damage will not be known.

When a fatal condition occurs, the decoder should fail immediately, signal an error to the user if appropriate, and optionally continue displaying any image data already visible to the user (i.e. "fail gracefully"). The application as a whole need not terminate.

When a non-fatal error occurs, the decoder should signal a warning to the user if appropriate, recover from the error, and continue processing normally.

When decoding an indexed-color PNG, if out-of-range indexes are encountered, decoders have historically varied in their handling of this error. Displaying the pixel as opaque black is one common error recovery tactic, and is now required by this specification. Older implementations will vary, and so the behavior must not be relied on by encoders.

Decoders that do not compute CRCs should interpret apparent syntax errors as indications of corruption (see also 13.2 Error checking).

Errors in compressed chunks ( IDAT, zTXt, iTXt, iCCP) could lead to buffer overruns. Implementors of deflate decompressors should guard against this possibility.

APNG is designed to allow incremental display of frames before the entire datastream has been read. This implies that some errors may not be detected until partway through the animation. It is strongly recommended that when any error is encountered decoders should discard all subsequent frames, stop the animation, and revert to displaying the static image. A decoder which detects an error before the animation has started should display the static image. An error message may be displayed to the user if appropriate.

Decoders shall treat out-of-order APNG chunks as an error. APNG-aware PNG editors should restore them to correct order, using the sequence numbers.

The PNG error handling philosophy is described in 13.1 Error handling.

An unknown chunk type is not to be treated as an error unless it is a critical chunk.

The chunk type can be checked for plausibility by seeing whether all four bytes are in the range codes 41-5A and 61-7A (hexadecimal); note that this need be done only for unrecognized chunk types. If the total datastream size is known (from file system information, HTTP protocol, etc), the chunk length can be checked for plausibility as well. If CRCs are not checked, dropped/added data bytes or an erroneous chunk length can cause the decoder to get out of step and misinterpret subsequent data as a chunk header.

For known-length chunks, such as IHDR, decoders should treat an unexpected chunk length as an error. Future extensions to this specification will not add new fields to existing chunks; instead, new chunk types will be added to carry new information.

Unexpected values in fields of known chunks (for example, an unexpected compression method in the IHDR chunk) shall be checked for and treated as errors. However, it is recommended that unexpected field values be treated as fatal errors only in critical chunks. An unexpected value in an ancillary chunk can be handled by ignoring the whole chunk as though it were an unknown chunk type. (This recommendation assumes that the chunk's CRC has been verified. In decoders that do not check CRCs, it is safer to treat any unexpected value as indicating a corrupted datastream.)

Standard PNG images shall be compressed with compression method 0. The compression method field of the IHDR chunk is provided for possible future standardization or proprietary variants. Decoders shall check this byte and report an error if it holds an unrecognized code. See 10. Compression for details.

A PNG datastream is composed of a collection of explicitly typed chunks. Chunks whose contents are defined by the specification could actually contain anything, including malicious code. Similarly there could be data after the IEND chunk which could contain anything, including malicious code. There is no known risk that such malicious code could be executed on the recipient's computer as a result of decoding the PNG image. However, a malicious application might hide such code inside an innocent-looking image file and then execute it.

The possible security risks associated with future chunk types cannot be specified at this time. Security issues will be considered when defining future public chunks. There is no additional security risk associated with unknown or unimplemented chunk types, because such chunks will be ignored, or at most be copied into another PNG datastream.

The iTXt, tEXt, and zTXt chunks contain keywords and data that are meant to be displayed as plain text. The iCCP and sPLT chunks contain keywords that are meant to be displayed as plain text. It is possible that if the decoder displays such text without filtering out control characters, especially the ESC (escape) character, certain systems or terminals could behave in undesirable and insecure ways. It is recommended that decoders filter out control characters to avoid this risk; see 13.7 Text chunk processing.

Every chunk begins with a length field, which makes it easier to write decoders that are invulnerable to fraudulent chunks that attempt to overflow buffers. The CRC at the end of every chunk provides a robust defence against accidentally corrupted data. The PNG signature bytes provide early detection of common file transmission errors.

A decoder that fails to check CRCs could be subject to data corruption. The only likely consequence of such corruption is incorrectly displayed pixels within the image. Worse things might happen if the CRC of the IHDR chunk is not checked and the width or height fields are corrupted. See 13.2 Error checking.

A poorly written decoder might be subject to buffer overflow, because chunks can be extremely large, up to 2³¹-1 bytes long. But properly written decoders will handle large chunks without difficulty.

Some image editing tools have historically performed redaction by merely setting the alpha channel of the redacted area to zero, without also removing the actual image data. Users who rely solely on the visual appearance of such images run a privacy risk because the actual image data can be easily recovered.

Similarly, some image editing tools have historically performed clipping by rewriting the width and height in IHDR without re-encoding the image data, which thus extends beyond the new width and height and may be recovered.

Images with eXIf chunks may contain automatically-included data, such as photographic GPS coordinates, which could be a privacy risk if the user is unaware that the PNG image contains this data. (Other image formats that contain EXIF, such as JPEG/JFIF, have the same privacy risk).

Decoders shall recognize chunk types by a simple four-byte literal comparison; it is incorrect to perform case conversion on chunk types. A decoder encountering an unknown chunk in which the ancillary bit is 1 may safely ignore the chunk and proceed to display the image. A decoder trying to extract the image, upon encountering an unknown chunk in which the ancillary bit is 0, indicating a critical chunk, shall indicate to the user that the image contains information it cannot safely interpret.

Decoders should test the properties of an unknown chunk type by numerically testing the specified bits. Testing whether a character is uppercase or lowercase is inefficient, and even incorrect if a locale-specific case definition is used.

Decoders should not flag an error if the reserved bit is set to 1, however, as some future version of the PNG specification could define a meaning for this bit. It is sufficient to treat a chunk with this bit set in the same way as any other unknown chunk type.

Decoders do not need to test the chunk type private bit, since it has no functional significance and is used to avoid conflicts between chunks defined by W3C and those defined privately.

All ancillary chunks are optional; decoders may ignore them. However, decoders are encouraged to interpret these chunks when appropriate and feasible.

Non-square pixels can be represented (see 11.3.4.3 pHYs Physical pixel dimensions), but viewers are not required to account for them; a viewer can present any PNG datastream as though its pixels are square.

Where the pixel aspect ratio of the display differs from the aspect ratio of the physical pixel dimensions defined in the PNG datastream, viewers are strongly encouraged to rescale images for proper display.

When the pHYs chunk has a unit specifier of 0 (unit is unknown), the behaviour of a decoder may depend on the ratio of the two pixels-per-unit values, but should not depend on their magnitudes. For example, a pHYs chunk containing (ppuX, ppuY, unit) = (2, 1, 0) is equivalent to one containing (1000, 500, 0); both are equally valid indications that the image pixels are twice as tall as they are wide.

One reasonable way for viewers to handle a difference between the pixel aspect ratios of the image and the display is to expand the image either horizontally or vertically, but not both. The scale factors could be obtained using the following floating-point calculations:

image_ratio = pHYs_ppuY / pHYs_ppuX
display_ratio = display_ppuY / display_ppuX
scale_factor_X = max(1.0, image_ratio/display_ratio)
scale_factor_Y = max(1.0, display_ratio/image_ratio)

Because other methods such as maintaining the image area are also reasonable, and because ignoring the pHYs chunk is permissible, authors should not assume that all viewing applications will use this scaling method.

As well as making corrections for pixel aspect ratio, a viewer may have reasons to perform additional scaling both horizontally and vertically. For example, a viewer might want to shrink an image that is too large to fit on the display, or to expand images sent to a high-resolution printer so that they appear the same size as they did on the display.

If practical, PNG decoders should have a way to display to the user all the iTXt, tEXt, and zTXt chunks found in the datastream. Even if the decoder does not recognize a particular text keyword, the user might be able to understand it.

When processing tEXt and zTXt chunks, decoders could encounter characters other than those permitted. Some can be safely displayed (e.g., TAB, FF, and CR, hexadecimal 09, 0C, and 0D, respectively), but others, especially the ESC character (hexadecimal 1B), could pose a security hazard (because unexpected actions may be taken by display hardware or software). Decoders should not attempt to directly display any non-Latin-1 characters (except for newline and perhaps TAB, FF, CR) encountered in a tEXt or zTXt chunk. Instead, they should be ignored or displayed in a visible notation such as "\nnn". See 13.3 Security considerations.

When processing iTXt chunks, decoders should follow UTF-8 decode in Encoding Standard.

Even though encoders are recommended to represent newlines as linefeed (hexadecimal 0A), it is recommended that decoders not rely on this; it is best to recognize all the common newline combinations (CR, LF, and CR-LF) and display each as a single newline. TAB can be expanded to the proper number of spaces needed to arrive at a column multiple of 8.

Decoders running on systems with a non-Latin-1 legacy character encoding should remap character codes so that Latin-1 characters are displayed correctly. Unsupported characters should be replaced with a system-appropriate replacement character (such as U+FFFD REPLACEMENT CHARACTER, U+003F QUESTION MARK, or U+001A SUB) or mapped to a visible notation such as "\nnn". Characters should be only displayed if they are printable characters on the decoding system. Some byte values may be interpreted by the decoding system as control characters; for security, decoders running on such systems should not display these control characters.

Decoders should be prepared to display text chunks that contain any number of printing characters between newline characters, even though it is recommended that encoders avoid creating lines in excess of 79 characters.

The compression technique used in this specification does not require the entire compressed datastream to be available before decompression can start. Display can therefore commence before the entire decompressed datastream is available. It is extremely unlikely that any general purpose compression methods in future versions of this specification will not have this property.

It is important to emphasize that IDAT chunk boundaries have no semantic significance and can occur at any point in the compressed datastream. There is no required correlation between the structure of the image data (for example, scanline boundaries) and deflate block boundaries or IDAT chunk boundaries. The complete image data is represented by a single zlib datastream that is stored in some number of IDAT chunks; a decoder that assumes any more than this is incorrect. Some encoder implementations may emit datastreams in which some of these structures are indeed related, but decoders cannot rely on this.

To reverse the effect of a filter, the decoder may need to use the decoded values of the prior pixel on the same line, the pixel immediately above the current pixel on the prior line, and the pixel just to the left of the pixel above. This implies that at least one scanline's worth of image data needs to be stored by the decoder at all times. Even though some filter types do not refer to the prior scanline, the decoder will always need to store each scanline as it is decoded, since the next scanline might use a filter type that refers to it. See 7.3 Filtering.

Decoders are required to be able to read interlaced images. If the reference image contains fewer than five columns or fewer than five rows, some passes will be empty. Encoders and decoders shall handle this case correctly. In particular, filter type bytes are associated only with nonempty scanlines; no filter type bytes are present in an empty reduced image.

When receiving images over slow transmission links, viewers can improve perceived performance by displaying interlaced images progressively. This means that as each reduced image is received, an approximation to the complete image is displayed based on the data received so far. One simple yet pleasing effect can be obtained by expanding each received pixel to fill a rectangle covering the yet-to-be-transmitted pixel positions below and to the right of the received pixel. This process can be described by the following ISO C code [ISO_9899]:

/*
    variables declared and initialized elsewhere in the code:
        height, width
    functions or macros defined elsewhere in the code:
        visit(), min()
 */

int starting_row[7]  = { 0, 0, 4, 0, 2, 0, 1 };
int starting_col[7]  = { 0, 4, 0, 2, 0, 1, 0 };
int row_increment[7] = { 8, 8, 8, 4, 4, 2, 2 };
int col_increment[7] = { 8, 8, 4, 4, 2, 2, 1 };
int block_height[7]  = { 8, 8, 4, 4, 2, 2, 1 };
int block_width[7]   = { 8, 4, 4, 2, 2, 1, 1 };

int pass;
long row, col;

pass = 0;
while (pass < 7)
{
    row = starting_row[pass];
    while (row < height)
    {
        col = starting_col[pass];
        while (col < width)
        {
            visit(row, col,
                  min(block_height[pass], height - row),
                  min(block_width[pass], width - col));
            col = col + col_increment[pass];
        }
        row = row + row_increment[pass];
    }
    pass = pass + 1;
}

The function visit(row,column,height,width) obtains the next transmitted pixel and paints a rectangle of the specified height and width, whose upper-left corner is at the specified row and column, using the colour indicated by the pixel. Note that row and column are measured from 0,0 at the upper left corner.

If the viewer is merging the received image with a background image, it may be more convenient just to paint the received pixel positions (the visit() function sets only the pixel at the specified row and column, not the whole rectangle). This produces a "fade-in" effect as the new image gradually replaces the old. An advantage of this approach is that proper alpha or transparency processing can be done as each pixel is replaced. Painting a rectangle as described above will overwrite background-image pixels that may be needed later, if the pixels eventually received for those positions turn out to be wholly or partially transparent. This is a problem only if the background image is not stored anywhere offscreen.

To achieve PNG's goal of universal interchangeability, decoders shall accept all types of PNG image: indexed-colour, truecolour, and greyscale. Viewers running on indexed-colour display hardware need to be able to reduce truecolour images to indexed-colour for viewing. This process is called "colour quantization".

A simple, fast method for colour quantization is to reduce the image to a fixed palette. Palettes with uniform colour spacing ("colour cubes") are usually used to minimize the per-pixel computation. For photograph-like images, dithering is recommended to avoid ugly contours in what should be smooth gradients; however, dithering introduces graininess that can be objectionable.

The quality of rendering can be improved substantially by using a palette chosen specifically for the image, since a colour cube usually has numerous entries that are unused in any particular image. This approach requires more work, first in choosing the palette, and second in mapping individual pixels to the closest available colour. PNG allows the encoder to supply suggested palettes, but not all encoders will do so, and the suggested palettes may be unsuitable in any case (they may have too many or too few colours). Therefore, high-quality viewers will need to have a palette selection routine at hand. A large lookup table is usually the most feasible way of mapping individual pixels to palette entries with adequate speed.

Numerous implementations of colour quantization are available. The PNG sample implementation, libpng (http://www.libpng.org/pub/png/libpng.html), includes code for the purpose.

Decoders may wish to scale PNG data to a lesser sample depth (data precision) for display. For example, 16-bit data will need to be reduced to 8-bit depth for use on most present-day display hardware. Reduction of 8-bit data to 5-bit depth is also common.

The most accurate scaling is achieved by the linear equation

output = floor((input * MAXOUTSAMPLE / MAXINSAMPLE) + 0.5)

where

MAXINSAMPLE = (2sampledepth)-1
MAXOUTSAMPLE = (2desired_sampledepth)-1

A slightly less accurate conversion is achieved by simply shifting right by (sampledepth - desired_sampledepth) places. For example, to reduce 16-bit samples to 8-bit, the low-order byte can be discarded. In many situations the shift method is sufficiently accurate for display purposes, and it is certainly much faster. (But if gamma correction is being done, sample rescaling can be merged into the gamma correction lookup table, as is illustrated in 13.13 Decoder gamma handling.)

If the decoder needs to scale samples up (for example, if the frame buffer has a greater sample depth than the PNG image), it should use linear scaling or left-bit-replication as described in 12.4 Sample depth scaling.

When an sBIT chunk is present, the reference image data can be recovered by shifting right to the sample depth specified by sBIT. Note that linear scaling will not necessarily reproduce the original data, because the encoder is not required to have used linear scaling to scale the data up. However, the encoder is required to have used a method that preserves the high-order bits, so shifting always works. This is the only case in which shifting might be said to be more accurate than linear scaling. A decoder need not pay attention to the sBIT chunk; the stored image is a valid PNG datastream of the sample depth indicated by the IHDR chunk; however, using sBIT to recover the original samples before scaling them to suit the display often yields a more accurate display than ignoring sBIT.

When comparing pixel values to tRNS chunk values to detect transparent pixels, the comparison shall be done exactly. Therefore, transparent pixel detection shall be done before reducing sample precision.

See C. Gamma and chromaticity for a brief introduction to gamma issues.

Viewers capable of full colour management will perform more sophisticated calculations than those described here.

For an image display program to produce correct tone reproduction, it is necessary to take into account the relationship between samples and display output, and the transfer function of the display system. This can be done by calculating:

sample = integer_sample / (2sampledepth - 1.0)
display_output = sample1.0/gamma
display_input = inverse_display_transfer(display_output)
framebuf_sample = floor((display_input * MAX_FRAMEBUF_SAMPLE)+0.5)

where integer_sample is the sample value from the datastream, framebuf_sample is the value to write into the frame buffer, and MAX_FRAMEBUF_SAMPLE is the maximum value of a frame buffer sample (255 for 8-bit, 31 for 5-bit, etc). The first line converts an integer sample into a normalized floating point value (in the range 0.0 to 1.0), the second converts to a value proportional to the desired display output intensity, the third accounts for the display system's transfer function, and the fourth converts to an integer frame buffer sample. Zero raised to any positive power is zero.

A step could be inserted between the second and third to adjust display_output to account for the difference between the actual viewing conditions and the reference viewing conditions. However, this adjustment requires accounting for veiling glare, black mapping, and colour appearance models, none of which can be well approximated by power functions. Such calculations are not described here. If viewing conditions are ignored, the error will usually be small.

The display transfer function can typically be approximated by a power function with exponent display_exponent, in which case the second and third lines can be merged into:

display_input = sample1.0/(gamma * display_exponent) = sampledecoding_exponent

so as to perform only one power calculation. For colour images, the entire calculation is performed separately for R, G, and B values.

The gamma value can be taken directly from the gAMA chunk. Alternatively, an application may wish to allow the user to adjust the appearance of the displayed image by influencing the gamma value. For example, the user could manually set a parameter user_exponent which defaults to 1.0, and the application could set:

gamma = gamma_from_file / user_exponent
decoding_exponent = 1.0 / (gamma * display_exponent)
   = user_exponent / (gamma_from_file * display_exponent)

The user would set user_exponent greater than 1 to darken the mid-level tones, or less than 1 to lighten them.

A gAMA chunk containing zero is meaningless but could appear by mistake. Decoders should ignore it, and editors may discard it and issue a warning to the user.

It is not necessary to perform a transcendental mathematical computation for every pixel. Instead, a lookup table can be computed that gives the correct output value for every possible sample value. This requires only 256 calculations per image (for 8-bit accuracy), not one or three calculations per pixel. For an indexed-colour image, a one-time correction of the palette is sufficient, unless the image uses transparency and is being displayed against a nonuniform background.

If floating-point calculations are not possible, gamma correction tables can be computed using integer arithmetic and a precomputed table of logarithms. Example code appears in [PNG-EXTENSIONS].

When the incoming image has unknown gamma value (gAMA, sRGB, and iCCP all absent), standalone image viewers should choose a likely default gamma value, but allow the user to select a new one if the result proves too dark or too light. The default gamma value may depend on other knowledge about the image, for example whether it came from the Internet or from the local system. For consistency, viewers for document formats such as HTML, or vector graphics such as SVG, should treat embedded or linked PNG images with unknown gamma value in the same way that they treat other untagged images.

In practice, it is often difficult to determine what value of display exponent should be used. In systems with no built-in gamma correction, the display exponent is determined entirely by the CRT. A display exponent of 2.2 should be used unless detailed calibration measurements are available for the particular CRT used.

Many modern frame buffers have lookup tables that are used to perform gamma correction, and on these systems the display exponent value should be the exponent of the lookup table and CRT combined. It may not be possible to find out what the lookup table contains from within the viewer application, in which case it may be necessary to ask the user to supply the display system's exponent value. Unfortunately, different manufacturers use different ways of specifying what should go into the lookup table, so interpretation of the system gamma value is system-dependent.

The response of real displays is actually more complex than can be described by a single number (the display exponent). If actual measurements of the monitor's light output as a function of voltage input are available, the third and fourth lines of the computation above can be replaced by a lookup in these measurements, to find the actual frame buffer value that most nearly gives the desired brightness.

See C. Gamma and chromaticity for references to colour issues.

In many cases, the image data in PNG datastreams will be treated as device-dependent RGB values and displayed without modification (except for appropriate gamma correction). This provides the fastest display of PNG images. But unless the viewer uses exactly the same display hardware as that used by the author of the original image, the colours will not be exactly the same as those seen by the original author, particularly for darker or near-neutral colours. The cHRM chunk provides information that allows closer colour matching than that provided by gamma correction alone.

The cHRM data can be used to transform the image data from RGB to XYZ and thence into a perceptually linear colour space such as CIE LAB. The colours can be partitioned to generate an optimal palette, because the geometric distance between two colours in CIE LAB is strongly related to how different those colours appear (unlike, for example, RGB or XYZ spaces). The resulting palette of colours, once transformed back into RGB colour space, could be used for display or written into a PLTE chunk.

Decoders that are part of image processing applications might also transform image data into CIE LAB space for analysis.

In applications where colour fidelity is critical, such as product design, scientific visualization, medicine, architecture, or advertising, PNG decoders can transform the image data from source RGB to the display RGB space of the monitor used to view the image. This involves calculating the matrix to go from source RGB to XYZ and the matrix to go from XYZ to display RGB, then combining them to produce the overall transformation. The PNG decoder is responsible for implementing gamut mapping.

Decoders running on platforms that have a Colour Management System (CMS) can pass the image data, gAMA, and cHRM values to the CMS for display or further processing.

PNG decoders that provide colour printing facilities can use the facilities in Level 2 PostScript to specify image data in calibrated RGB space or in a device-independent colour space such as XYZ. This will provide better colour fidelity than a simple RGB to CMYK conversion. The PostScript Language Reference manual [PostScript] gives examples. Such decoders are responsible for implementing gamut mapping between source RGB (specified in the cHRM chunk) and the target printer. The PostScript interpreter is then responsible for producing the required colours.

PNG decoders can use the cHRM data to calculate an accurate greyscale representation of a colour image. Conversion from RGB to grey is simply a case of calculating the Y (luminance) component of XYZ, which is a weighted sum of R, G, and B values. The weights depend upon the monitor type, i.e. the values in the cHRM chunk. PNG decoders may wish to do this for PNG datastreams with no cHRM chunk. In this case, a reasonable default would be the CCIR 709 primaries [ITU-R-BT.709]. The original NTSC primaries should not be used unless the PNG image really was colour-balanced for such a monitor.

The background colour given by the bKGD chunk will typically be used to fill unused screen space around the image, as well as any transparent pixels within the image. (Thus, bKGD is valid and useful even when the image does not use transparency.) If no bKGD chunk is present, the viewer will need to decide upon a suitable background colour. When no other information is available, a medium grey such as 153 in the 8-bit sRGB colour space would be a reasonable choice. Transparent black or white text and dark drop shadows, which are common, would all be legible against this background.

Viewers that have a specific background against which to present the image (such as web browsers) should ignore the bKGD chunk, in effect overriding bKGD with their preferred background colour or background image.

The background colour given by the bKGD chunk is not to be considered transparent, even if it happens to match the colour given by the tRNS chunk (or, in the case of an indexed-colour image, refers to a palette index that is marked as transparent by the tRNS chunk). Otherwise one would have to imagine something "behind the background" to composite against. The background colour is either used as background or ignored; it is not an intermediate layer between the PNG image and some other background.

Indeed, it will be common that the bKGD and tRNS chunks specify the same colour, since then a decoder that does not implement transparency processing will give the intended display, at least when no partially-transparent pixels are present.

The alpha channel can be used to composite a foreground image against a background image. The PNG datastream defines the foreground image and the transparency mask, but not the background image. PNG decoders are not required to support this most general case. It is expected that most will be able to support compositing against a single background colour.

The equation for computing a composited sample value is:

output = alpha * foreground + (1-alpha) * background

where alpha and the input and output sample values are expressed as fractions in the range 0 to 1. This computation should be performed with intensity samples (not gamma-encoded samples). For colour images, the computation is done separately for R, G, and B samples.

The following code illustrates the general case of compositing a foreground image against a background image. It assumes that the original pixel data are available for the background image, and that output is to a frame buffer for display. Other variants are possible; see the comments below the code. The code allows the sample depths and gamma values of foreground image and background image all to be different and not necessarily suited to the display system. In practice no assumptions about equality should be made without first checking.

This code is ISO C [ISO_9899], with line numbers added for reference in the comments below.

01  int foreground[4];  /* image pixel: R, G, B, A */
02  int background[3];  /* background pixel: R, G, B */
03  int fbpix[3];       /* frame buffer pixel */
04  int fg_maxsample;   /* foreground max sample */
05  int bg_maxsample;   /* background max sample */
06  int fb_maxsample;   /* frame buffer max sample */
07  int ialpha;
08  float alpha, compalpha;
09  float gamfg, linfg, gambg, linbg, comppix, gcvideo;

    /* Get max sample values in data and frame buffer */
10  fg_maxsample = (1 << fg_sample_depth) - 1;
11  bg_maxsample = (1 << bg_sample_depth) - 1;
12  fb_maxsample = (1 << frame_buffer_sample_depth) - 1;
    /*
     * Get integer version of alpha.
     * Check for opaque and transparent special cases;
     * no compositing needed if so.
     *
     * We show the whole gamma decode/correct process in
     * floating point, but it would more likely be done
     * with lookup tables.
     */
13  ialpha = foreground[3];

14  if (ialpha == 0) {
        /*
         * Foreground image is transparent here.
         * If the background image is already in the frame
         * buffer, there is nothing to do.
         */
15      ;
16  } else if (ialpha == fg_maxsample) {
        /*
         * Copy foreground pixel to frame buffer.
         */
17      for (i = 0; i < 3; i++) {
18          gamfg = (float) foreground[i] / fg_maxsample;
19          linfg = pow(gamfg, 1.0 / fg_gamma);
20          comppix = linfg;
21          gcvideo = pow(comppix, 1.0 / display_exponent);
22          fbpix[i] = (int) (gcvideo * fb_maxsample + 0.5);
23      }
24  } else {
        /*
         * Compositing is necessary.
         * Get floating-point alpha and its complement.
         * Note: alpha is always linear; gamma does not
         * affect it.
         */
25      alpha = (float) ialpha / fg_maxsample;
26      compalpha = 1.0 - alpha;

27      for (i = 0; i < 3; i++) {
            /*
             * Convert foreground and background to floating
             * point, then undo gamma encoding.
             */
28          gamfg = (float) foreground[i] / fg_maxsample;
29          linfg = pow(gamfg, 1.0 / fg_gamma);
30          gambg = (float) background[i] / bg_maxsample;

31          linbg = pow(gambg, 1.0 / bg_gamma);
            /*
             * Composite.
             */
32          comppix = linfg * alpha + linbg * compalpha;
            /*
             * Gamma correct for display.
             * Convert to integer frame buffer pixel.
             */
33          gcvideo = pow(comppix, 1.0 / display_exponent);
34          fbpix[i] = (int) (gcvideo * fb_maxsample + 0.5);
35      }
36  }

Variations:

1. If output is to another PNG datastream instead of a frame buffer, lines 21, 22, 33, and 34 should be changed along the following lines

   /*
    * Gamma encode for storage in output datastream.
    * Convert to integer sample value.
    */
   gamout = pow(comppix, outfile_gamma);
   outpix[i] = (int) (gamout * out_maxsample + 0.5);

   Also, it becomes necessary to process background pixels when alpha is zero, rather than just skipping pixels. Thus, line 15 will need to be replaced by copies of lines 17-23, but processing background instead of foreground pixel values.
2. If the sample depths of the output file, foreground file, and background file are all the same, and the three gamma values also match, then the no-compositing code in lines 14-23 reduces to copying pixel values from the input file to the output file if alpha is one, or copying pixel values from background to output file if alpha is zero. Since alpha is typically either zero or one for the vast majority of pixels in an image, this is a significant saving. No gamma computations are needed for most pixels.
3. When the sample depths and gamma values all match, it may appear attractive to skip the gamma decoding and encoding (lines 28-31, 33-34) and just perform line 32 using gamma-encoded sample values. Although this does not have too bad an effect on image quality, the time savings are small if alpha values of zero and one are treated as special cases as recommended here.
4. If the original pixel values of the background image are no longer available, only processed frame buffer pixels left by display of the background image, then lines 30 and 31 need to extract intensity from the frame buffer pixel values using code such as

   /*
    * Convert frame buffer value into intensity sample.
    */
   gcvideo = (float) fbpix[i] / fb_maxsample;
   linbg = pow(gcvideo, display_exponent);

   However, some roundoff error can result, so it is better to have the original background pixels available if at all possible.
5. Note that lines 18-22 are performing exactly the same gamma computation that is done when no alpha channel is present. If the no-alpha case is handled with a lookup table, the same lookup table can be used here. Lines 28-31 and 33-34 can also be done with (different) lookup tables.
6. Integer arithmetic can be used instead of floating point, providing care is taken to maintain sufficient precision throughout.

When displaying a PNG image with full alpha channel, it is important to be able to composite the image against some background, even if it is only black. Ignoring the alpha channel will cause PNG images that have been converted from an associated-alpha representation to look wrong. (Of course, if the alpha channel is a separate transparency mask, then ignoring alpha is a useful option: it allows the hidden parts of the image to be recovered.)

Even if the decoder does not implement true compositing logic, it is simple to deal with images that contain only zero and one alpha values. (This is implicitly true for greyscale and truecolour PNG datastreams that use a tRNS chunk; for indexed-colour PNG datastreams it is easy to check whether the tRNS chunk contains any values other than 0 and 255.) In this simple case, transparent pixels are replaced by the background colour, while others are unchanged.

If a decoder contains only this much transparency capability, it should deal with a full alpha channel by treating all nonzero alpha values as fully opaque or by dithering. Neither approach will yield very good results for images converted from associated-alpha formats, but this is preferable to doing nothing. Dithering full alpha to binary alpha is very much like dithering greyscale to black-and-white, except that all fully transparent and fully opaque pixels should be left unchanged by the dither.

For viewers running on indexed-colour hardware attempting to display a truecolour image, or an indexed-colour image whose palette is too large for the frame buffer, the encoder may have provided one or more suggested palettes in sPLT chunks. If one of these is found to be suitable, based on size and perhaps name, the PNG decoder can use that palette. Suggested palettes with a sample depth different from what the decoder needs can be converted using sample depth rescaling (see 13.12 Sample depth rescaling).

When the background is a solid colour, the viewer should composite the image and the suggested palette against that colour, then quantize the resulting image to the resulting RGB palette. When the image uses transparency and the background is not a solid colour, no suggested palette is likely to be useful.

For truecolour images, a suggested palette might also be provided in a PLTE chunk. If the image has a tRNS chunk and the background is a solid colour, the viewer will need to adapt the suggested palette for use with its desired background colour. To do this, the palette entry closest to the tRNS colour should be replaced with the desired background colour; or alternatively a palette entry for the background colour can be added, if the viewer can handle more colours than there are PLTE entries.

For images of colour type 6 (truecolour with alpha), any PLTE chunk should have been designed for display of the image against a uniform background of the colour specified by the bKGD chunk. Viewers should probably ignore the palette if they intend to use a different background, or if the bKGD chunk is missing. Viewers can use a suggested palette for display against a different background than it was intended for, but the results may not be very good.

If the viewer presents a transparent truecolour image against a background that is more complex than a uniform colour, it is unlikely that the suggested palette will be optimal for the composite image. In this case it is best to perform a truecolour compositing step on the truecolour PNG image and background image, then colour-quantize the resulting image.

In truecolour PNG datastreams, if both PLTE and sPLT chunks appear, the PNG decoder may choose from among the palettes suggested by both, bearing in mind the different transparency semantics described above.

The frequencies in the sPLT and hIST chunks are useful when the viewer cannot provide as many colours as are used in the palette in the PNG datastream. If the viewer has a shortfall of only a few colours, it is usually adequate to drop the least-used colours from the palette. To reduce the number of colours substantially, it is best to choose entirely new representative colours, rather than trying to use a subset of the existing palette. This amounts to performing a new colour quantization step; however, the existing palette and histogram can be used as the input data, thus avoiding a scan of the image data in the IDAT chunks.

If no suggested palette is provided, a decoder can develop its own, at the cost of an extra pass over the image data in the IDAT chunks. Alternatively, a default palette (probably a colour cube) can be used.

See also 12.5 Suggested palettes.

Authors are encouraged to look existing chunk types in both this specification and [PNG-EXTENSIONS] before considering introducing a new chunk types. The chunk types at [PNG-EXTENSIONS] are expected to be less widely supported than those defined in this specification.

Two examples of PNG editors are a program that adds or modifies text chunks, and a program that adds a suggested palette to a truecolour PNG datastream. Ordinary image editors are not PNG editors because they usually discard all unrecognized information while reading in an image.

To allow new chunk types to be added to PNG, it is necessary to establish rules about the ordering requirements for all chunk types. Otherwise a PNG editor does not know what to do when it encounters an unknown chunk.

EXAMPLE Consider a hypothetical new ancillary chunk type that is safe-to-copy and is required to appear after PLTE if PLTE is present. If a program attempts to add a PLTE chunk and does not recognize the new chunk, it may insert the PLTE chunk in the wrong place, namely after the new chunk. Such problems could be prevented by requiring PNG editors to discard all unknown chunks, but that is a very unattractive solution. Instead, PNG requires ancillary chunks not to have ordering restrictions like this.

To prevent this type of problem while allowing for future extension, constraints are placed on both the behaviour of PNG editors and the allowed ordering requirements for chunks. The safe-to-copy bit defines the proper handling of unrecognized chunks in a datastream that is being modified.

1. If a chunk's safe-to-copy bit is 1, the chunk may be copied to a modified PNG datastream whether or not the PNG editor recognizes the chunk type, and regardless of the extent of the datastream modifications.
2. If a chunk's safe-to-copy bit is 0, it indicates that the chunk depends on the image data. If the program has made any changes to critical chunks, including addition, modification, deletion, or reordering of critical chunks, then unrecognized unsafe chunks shall not be copied to the output PNG datastream. (Of course, if the program does recognize the chunk, it can choose to output an appropriately modified version.)
3. A PNG editor is always allowed to copy all unrecognized ancillary chunks if it has only added, deleted, modified, or reordered ancillary chunks. This implies that it is not permissible for ancillary chunks to depend on other ancillary chunks.
4. PNG editors shall terminate on encountering an unrecognized critical chunk type, because there is no way to be certain that a valid datastream will result from modifying a datastream containing such a chunk. (Simply discarding the chunk is not good enough, because it might have unknown implications for the interpretation of other chunks.) The safe/unsafe mechanism is intended for use with ancillary chunks. The safe-to-copy bit will always be 0 for critical chunks.

The rules governing ordering of chunks are as follows.

1. When copying an unknown unsafe-to-copy ancillary chunk, a PNG editor shall not move the chunk relative to any critical chunk. It may relocate the chunk freely relative to other ancillary chunks that occur between the same pair of critical chunks. (This is well defined since the editor shall not add, delete, modify, or reorder critical chunks if it is preserving unknown unsafe-to-copy chunks.)
2. When copying an unknown safe-to-copy ancillary chunk, a PNG editor shall not move the chunk from before IDAT to after IDAT or vice versa. (This is well defined because IDAT is always present.) Any other reordering is permitted.
3. When copying a known ancillary chunk type, an editor need only honour the specific chunk ordering rules that exist for that chunk type. However, it may always choose to apply the above general rules instead.

These rules are expressed in terms of copying chunks from an input datastream to an output datastream, but they apply in the obvious way if a PNG datastream is modified in place.

See also 5.4 Chunk naming conventions.

PNG editors that do not change the image data should not change the tIME chunk. The Creation Time keyword in the tEXt, zTXt, and iTXt chunks may be used for a user-supplied time.