Galaxy4594

[Oxipng](https://github.com/shssoichiro/oxipng) and [ECT](https://github.com/fhanau/Efficient-Compression-Tool) are PNG optimizers that losslessly recompress PNGs. In this case, the colorspace of the b/w image is RGB, so these programs will detect that and convert it...

> 409,961 gray_C0.jxl 120,626 gray.jxl How strange, what are your encoding settings? Operating system? Libjxl version? I tested with the same b/w image you posted. I'm on Windows libjxl 0.11:...

Yup, I can confirm. Effort 3 is the culprit, bump up modular mode to effort 4 and the file sizes are much smaller. `cjxl gray.png gray.jxl -m 1 -e 3`...

Bumping up the max property values didn't really decrease file size as much as I hoped (and sometimes even made files bigger!), so I'm not sure if that is the...

~~My histogram retune does come at a small cost of encoding performance (around 1-2%). Your PR is good middle ground, making denser files compared to main but not as dense...

Hmm, well re-testing with my PR synced to main, the performance gap is gone (in fact it might be slightly faster on my machine?). Maybe this idea still worth pursuing...

It doesn't happen for v0.10, maybe recent the changes in e11 are the culprit? It's funny how #3779 and #3782 increase density but in this edge case it's a regression....

> Does Butteraugli really affect lossless? No. > Theoretically the more context available, the better estimation possible. That's true, streaming speeds up encoding and reduces ram usage at a small...

The primary difference between efforts 9 and 10 is the choice of predictor. Predictor 6 produced the smallest output overall; however, it's interesting that effort 9 _still_ resulted in a...

There might be more relevant information here: #3841 You can probably just change a couple of flags in the build process and get smaller binaries.