A new protocol and tool for PNG file attachments

When my articles include diagrams to illustrate a concept, such as a state machine, I will check the Graphviz, gnuplot, or SVG source into source control alongside the image in case I need to make changes in the future. Sometimes I even make the image itself a link to its source file. I’ve thought it would be convenient if the raster image somehow contained its own source as metadata so that they don’t get separated. I looked around and wasn’t satisfied with the solutions I found, so I wrote one: pngattach.

My approach introduces a new private chunk type atCh (“attachment”) which contains a file name, a flag to indicate if the attachment is compressed, and an optionally DEFLATE-compressed blob of file contents. I tried to follow the spirit of PNG chunk formatting, but without the constraints I hoped to avoid. A single PNG can contain multiple attachments, e.g. source file, Makefile, README, license file, etc. The protocol places constraints on the file names to keep it simple and to avoid shenanigans: no control bytes (anything below ASCII space), no directories, and cannot start with a period (no special hidden files). If that’s too constraining, you could attach a ZIP or TAR.

PNG chunk format

PNG files begin with a fixed 8-byte header followed by of a series of chunks. Each chunk has an 8-byte header and 4-byte footer. The chunk header is a 32-bit big endian chunk length (not counting header or footer) and a 4-byte tag identifying its type. The length allows implementations to skip chunks it doesn’t recognize.

LLLL TTTT ...chunk... CCCC

The footer is a big endian CRC-32 checksum of the 4-byte type tag and the chunk body itself.

Chunk tags are interpreted as 4 ASCII characters, where the capitalization of each letter encodes 4 additional boolean flags. The flags in my tag, atCh, indicate it’s a non-critical private chunk which doesn’t depend on the image data.

PNG always ends with a zero-length IEND chunk, which works out to a kind of 12-byte constant footer.

Existing chunk types

The PNG standard currently defines three kinds of chunks for storing text metadata: tEXt, iTXt, zTXt. The first is limited to Latin-1 with LF newlines, and so cannot store UTF-8 source text. The latter two were introduced in the PNG 1.2 specification (November 1999), and allow (only) UTF-8 content with LF newlines. All three have a 1 to 79-byte Latin-1 “key” field, and the latter two some additional fields describing the language of the text.

The key field is null-terminated, making it 80 bytes maximum when treated as a null-terminated string. I believe this constraint exists to aid implementations, which can rely on this hard upper limit for the key lengths they’re expected to handle. Otherwise a key could have been up to 4GiB in length.

I had considered using part of the key as a file name, prefixed with a custom namespace (ex. attachment:FILENAME) to distinguish it from other text chunks. However, I didn’t like the constraints this placed on the file name, plus I wanted to support arbitrary file content, not limited to a particular subformat.

As prior art, there’s a draw.io/diagrams.net format which embeds a source string without file name. The source string is encoded in base64 (i.e. unconstrained by PNG), wrapped in XML, then incorrectly encoded as an iTXt chunk. The XML alone was enough to keep me away from using this format.

pngattach details

In my attachment protocol, the file name is an arbitrary length, null-terminated byte string (preferably UTF-8), much like a key field, with the previously-mentioned anti-shenanigans restrictions. The file name is followed by a byte, 0 or 1, indicating if the content is compressed using PNG’s officially-supported compression format. The rest is the arbitrary content bytes, which presumably the recipient will know how to use.

LLLL atCh example.txt 0 F ...contents... CCCC

I expect any experienced programmer could write a basic attachment extractor in their language of choice inside of 30 or so minutes. Hooking up a DEFLATE library for decompression would be the most difficult part.

Since it supports multiple attachments and behaves like an archive format, my tool supports flags much like tar: -c to create attachments (default and implicit), -t to list attachments, and -x to extract attachments. PNG data is always passed on standard input and standard output.

For example, to render a Graphviz diagram and attach the source all at once:

$ dot -Tpng graph.dot | pngattach graph.dot >graph.png

Later on someone might extract it and tweak it, like so (-v verbose, lists files as they’re extracted, like tar):

$ pngattach -xv <graph.png
$ vi graph.dot
$ dot -Tpng graph.dot >graph.png

Like tar, it can also write attachments to standard output with -O. For example, to re-render the image as an SVG:

$ pngattach -xO <graph.png | dot -Tsvg >graph.svg

Strictly processing standard input to output, rather than taking the input as an argument, is something I’ve been trying lately. I’m pretty happy with my command line design for pngattach. The real test will happen in the future, when I’ve forgotten the details and have to figure it out again from my own documentation.

Curiously, lots of common software refuses to handle PNGs containing large chunks, and so your PNG may not display if you attach a file even as small as a few MiB. A defense against denial of service?

Example PNG

I haven’t gone back and embedded attachments in any older articles, but I may do so in future articles. If you wanted to try it out for yourself, either with my tool or writing your own for fun, this PNG contains a compressed attachment:

I produced it like so (with the help of ImageMagick):

$ echo P3 1 1 1 0 1 0 |
      convert ppm:- resize 200 png:- |
      pngattach message.txt >atch-test.png

Error handling (addendum)

Another technique I’ve been trying is Go-style error value returns in C programs, where the errors-as-values are const char * pointers to static string buffers. The contents contain an error message to be displayed to the user, and errors may be wrapped in more context (what file, what operation, etc.) as the stack unwinds. A null pointer means no error, i.e. nil. I’ve used this extensively in pngattach. Examples of the style:

    int *p;

    if (nelem > (size_t)-1/sizeof(*p)) {
        return "out of memory";  // overflow

    p = malloc(nelem*sizeof(*p));
    if (!p) {
        return "out of memory";

    // ...

    if (!fwrite(buf, len, 1, stdout)) {
        return "write error";

An errwrap() function builds a new error string in a static buffer. This simple solution wouldn’t work in a multi-threaded program, but that’s not the case here. Mine toggles between two static buffers so that it can wrap recursively.

const char *
errwrap(const char *pre, const char *post)
    static char errtmp[2][256], i;
    int n = i = !i;  // toggle between two static buffers
    snprintf(errtmp[n], sizeof(errtmp[n]), "%s: %s", pre, post);
    return errtmp[n];

Then I can do stuff like:

    FILE *f = fopen(path, "wb");
    if (!f) {
        return errwrap("failed to open file", path);

And that can keep being wrapped on the way up:

    err = png_write(path);
    if (err) {
        return errwrap("writing PNG", err);

So that ultimately the user sees something like:

pngattach: writing PNG: failed to open file: example.png

That’s always printed by a single error printout block at the top level, where all errors are ultimately routed.

int main(int argc, char **argv)
    // ...

    err = run(options);
    if (err) {
        fprintf(stderr, "pngattach: %s\n", err);
        return 1;
    return 0;

Have a comment on this article? Start a discussion in my public inbox by sending an email to ~skeeto/public-inbox@lists.sr.ht [mailing list etiquette] , or see existing discussions.

null program

Chris Wellons

wellons@nullprogram.com (PGP)
~skeeto/public-inbox@lists.sr.ht (view)