Surprising that blurring AB of LAB works better than blurring H of HCL!

Yep I was surprised too. On top of that there's a lot of work to do to account for polar variables (blur sin and cos then take the atan).

👌

Weight a pixel's neighbors by similarity in lightness perhaps? And/or use a gaussian blur kernel?

the kernel in array-blur is already very similar to a gaussian, but I'm interested in your suggestion about weighting by lightness. Maybe this would be easier to do in WebGL though?

Not necessarily, I think? The steps would be: - extract lightness channel L - blur L to remove noise - specify max lightness difference d within which a pixel can contribute to the kernel - for each pixel: -- init hue sum `sh=0`, weight sum `sw=0`, lookup and store lightness at current pixel `pl` -- for each pixel in kernel: --- lookup pixel lightness `kl` --- calc weight: `kw=clamp(0, 1, abs(pl-kl))` --- if `kw==0` skip --- lookup kernel pixel hue `kh` --- `sh+=kw*kh` --- `sw+=kw`

Oops, small error for weight: `kw=1-clamp(0, d, abs(pl-kl))/d`

For later: https://observablehq.com/@stwind/glsl-gaussian-blur (→ link to library, → Github link to blog post)