the FSE layer isn't responsible for finding these sorts of patterns in an image codec. The domain modeling turns that sort of pattern into repeated data and then the FSE goes to town on the output.

zlib/deflate already has the same issue. It is mitigated somewhat by PNG row filters.

PNG's basic idea is to store the difference between the current pixel and the pixel above it, left of it or to the top-left (chosen once per row), then apply standard deflate compression to that. The first step basically turns gradients into repeating patterns of small numbers, which compress great. You can get decent improvements by just switching deflate for zstd

> You can get decent improvements by just switching deflate for zstd

Maybe I sounded too critical of zstd. To be clear: I use it for general-purpose compression where available, the only exception would be where eeking out the last % gain is important and slow decompression is acceptable and it has one of these patterns that Bzip2 does better in the first place

That it's better than deflate (afaik aka gzip and zlib, just with different header fields) is not surprising since that was iirc the defined goal of Zstandard project

JPEG XL is not “apply zstd to the pixel data”. What the parent comment meant is that JPEG XL, like zstd, happens to use ANS among its set of techniques. It handles gradients just fine.