Jake ArchibaldFolks who want to see JPEG-XL supported in more browsers, what is it about the format that attracts you to its use on the web compared to currently supported formats?
KKT@jaffathecake Unlike aspects of WebP & AVIF, JPEG XL is incredibly well thought out, and can be used in every stage of the image pipeline – RAW camera data through optimized web delivery.
@kkthompson which aspects of AVIF are you referring to?
@jaffathecake Mostly covered here already, but a few things AVIF doesn't have: best-of-breed lossless (including non-photographic images like logos when you don't have a vector), encoding speed, high max bit depths, very large max image size, progressive decoding, generational loss resilience, JPEG recompression… The JPEG XL spec was well conceived to grow into the future.
Kornel@kkthompson JXL is state-of-the-art for lossless, but I don't think actual true lossless is useful on the Web. It's for editing, not publishing.
People usually say they want "lossless" when they mean they don't want JPEG-like smudges, but that's a false dichotomy. It's possible to have blur-free perfectly sharp edge-preserving lossy compression, and AV1 supports it via palette-based blocks.
Henri Sivonen@kornel What tool configuration should I look for to do Rec 2100 HDR with AVIF at the quality level that’s below mathematically lossles, below flip-test visually lossless but above the level where you can see compression artifacts by zooming in? I.e. level that looks as good as lossless if you aren’t doing a flip test. (Something to compete with libjxl distance 1.4 or so.)
@hsivonen Unfortunately I can't help you with HDR, I haven't looked into HDR color spaces yet.
I suspect rav1e has a bug in allocation of luma vs chroma bitrate, and doesn't support palette blocks. If you're evaluating AVIF, use libaom or svt-av1.
@kornel Thanks. I’ve been told that SVT-AV1 starts by quartering chroma (which might arguably be a reasonable first step, but it’s also interesting that libjxl supposedly doesn’t do something as blunt). I have a hard time figuring out what the libaom quality parameter does conceptually (apart from the obvious lower compresses more), so I have a hard time figuring out what outcome I should expect with future input that I haven’t tested specifically.