% jxl_from_tree /usr/share/libjxl-testdata/jxl/splines.tree splines.jxl
% jxlinfo splines.jxl
JPEG XL image, 320x320, (possibly) lossless, 8-bit RGB
Color space: RGB, D65, sRGB primaries, sRGB transfer function, rendering intent: Relative
% md5sum splines.jxl
222fcf2d528904bc796a7c3a3c64cd76 splines.jxl
However:
% djxl splines.jxl splines.ppm
% md5sum splines.ppm
may produce either:
77c4ccf6f23b320819610ebd5e1b2af0 splines.ppm
or
9ca111503859edaa6c3b1cb92ff657b7 splines.ppm
Splines are not currently used by the encoder at all, this is a test bitstream that was artificially
produced. Splines are additional image features that are represented internally in a vector form (control
points for catmull-rom splines) and painted by the decoder on top of (added numerically) the underlying
pixels. The jxlinfo tool only inspects the image header and sees that the pixels were encoded in RGB (not
XYB), so it offers the suggestion that the image is possibly lossless, but this is in any case just a
possibility (there is no way for it to know the history and workflow that was used to produce the image).
In this case the pixels themselves are encoded losslessly using Modular, but they are just all black. The
rest is painted with splines.
Spline rendering is defined mathematically in the spec and with infinite precision it is precisely
defined; however practical implementations do things with limited precision and there can be tiny
differences caused by rounding errors (which can be platform dependent and compiler dependent). The same
is true for other coding tools like the DCT. For that reason we have conformance tolerances.