JPEG XL

Not sure (hate Slack), but maybe thread better…?

Lossy jxl always has at least two passes: one with a 1:8 resolution image (enough for a blurry preview that is good enough to understand what's in the image, but bad enough to see it's not the final image), and one with the final image. The passes themselves can be ordered in other orders than top to bottom too, e.g. you can do middle-first or salient regions first, and get interesting loading behavior even with just two passes.

More passes can be added, it should not make a big difference in filesize but it will generally come at a small cost compared to doing just the basic progressive.

The basic progressiveness that you always get is also a good default, which is why we made it the default 🙂

cjxl currently doesn't offer a lot of control over progressive passes, it only has a few options for non-default things to do. The main thing is to do more "AC" passes, which is similar to what you can do in progressive jpeg, and gives you intermediate steps between 1:8 and 1:1. JXL is more flexible than JPEG here though: in JPEG the quality/resolution improvement of each pass is always affecting the whole image, while in JXL you can e.g. make a pass that adds a lot of detail in, say, the faces or the text, while other regions are left blurry in that pass.

Another thing JXL can do is "progressive DC", i.e. passes that are even lower resolution than 1:8.

jxlinfo alas cannot give much info about progressive passes since libjxl itself does not expose much information about it, besides implicitly by generating more events during actual decoding. But jxlinfo doesn't do actual decoding, it only does the header parsing bits of the process without asking libjxl to do the actual decode.

Thank you! So could maybe djxl give info about what progressive-related options were used during encoding? I guess I will have to experiment with ``` --progressive (sounds magic, is it default?) --group_order (are its --center_x&y basically the saliency demo knobs?) --progressive_ac (what’s its input? number of extra steps?) --qprogressive_ac (I really need to read some book or somepin) --progressive_dc (so 0 removes even the 1:8? 1 is default 1:8 and 2 introduces 1:4?) ``` ? Yeah, can’t read code, but maybe will find some discussion of them on github. Thanks you for pointers!

You cannot remove the 1:8. --progressive_dc=1 will add 1:16, 1:32 etc. Setting it to higher values than 1 is only useful for super huge images where you can do recursively progressive DC (only useful if even the 1:64 image is still a useful preview).

Group order / center x/y is indeed like in the saliency demo

Progressive ac and qprogressive ac are two different "scan scripts" that are available in libjxl, for the 1:4 / 1:2 passes. I don't remember exactly what the difference is, iirc one of them does more (in progressive jpeg terms) spectral selection, while the other does more successive approximation.