If the unit specifier byte is 1, the units are meters; if it is 0, the units are unspecified, and only the relative dimensions are known. Currently, no other values are valid. Note that the format of the chunk precludes pixel sizes greater than one meter, which should not be a significant hardship for most applications, but it allows pixels as small as 4.7 Ångstroms, which is roughly the size of a single atom.

For the previous scanning example, 600 dpi is equal to 23,622.05 pixels per meter, so both the x and y values would be 23,622, and the unit specifier would be 1. The second example is slightly trickier. First, it is necessary to know that practically all current computer displays have a physical aspect ratio of 4:3,[89] which means the viewable portion of the display (the glass) is three-quarters as high as it is wide. Thus, the horizontal pixels-per-unit in the case of a 1280 × 1024 display is proportional to (1280/4) or 320, while the vertical pixels-per-unit is proportional to (1024/3) or 341.333333. Because we don't have an absolute scale, we are free to multiply these values by a common factor; doing so will preserve some of the precision that would otherwise be lost due to truncation of the decimal part of the second value (the .3333 part). One choice would be a power of 10, such as 1,000; then the stored values would be 320,000 and 341,333, respectively. But in this case, we can do better: we know that the fractional part is simply one-third, so multiplying both values by 3 will preserve the aspect ratio exactly. Thus the chunk would contain the values (3 × 1280/4) or 960, (3 × 1024/3) or 1,024, and 0 for the unit specifier. Values of 15, 16, and 0 would work equally well.

[89] This will change with the convergence of computers and high-definition TV. Displays for the latter have a 16:9 aspect ratio, which apparently is the geometric mean of standard television and computer displays (4:3) and of modern, panoramic films (typically 2.35:1, but it varies).

A decoder that encounters a pHYs chunk with different values for the x and y axes has several options. The simplest and least correct approach is to ignore the chunk; most current viewers do this. A better approach is to interpolate the pixels in one of the dimensions; this gives the correct overall appearance but introduces noticeable artifacts--for the preceding example, it involves either duplicating every 15th column stretching the image horizontally, or deleting every 16th row shrinking the image vertically. The best approach is to resample the image, a procedure that amounts to converting the image to a continuous (or analog) representation and then overlaying the desired pixel grid on that. This is, by far, the most expensive approach in terms of CPU usage, but the results are excellent.