JPEG XL

this thread is for posting visualizations of prediction residuals and which predictors and contexts are used where in images. feel free to post images here to have them tested as well

JXL logo input/context/predictors/residuals

there's almost nothing left in the residuals, not even near the edges of the letters. this is effort 10 with all predictors (`-P 15`), any other predictor setting leaves significant residuals there

I think I need to change the way my imagemagick command to downscale images works. it appears to add a lot of contrast between edges to restore sharpness or something, but that just makes prediction (and possibly context modeling) worse

Box downscaling is the best if you want sharpness with no ringing

thanks, already found that out 🙂 i also compared how those sharpness effects effect context modeling and it results in a *lot* more contexts scattered all over the place, almost like noise. though of course noise in contexts doesn't necessarily mean worse file size since the MA tree can still be pretty small and just applied in a noisy way

hmm, should i prefer `-filter box -resize` or `-scale`?

Spider Mario says `-scale` works better

It wouldn't hurt to try both though. I don't use image magick

Keep in mind the box filter works best with 1/2, 1/3, 1/4, etc. downscales

they look nearly identical. i'll use scale to be safe

peaks original/context/predictors/residuals

I also like setting the predictor to `0` so that the context is just decided by the original values without being influenced by predictors