This is a neat article with some good experimentation, but... This is essentially someone seeing the precursor to the bilateral filter for the first time without ever having heard of a bilateral filter before.
A bilateral filter weights surrounding pixels by how close in value they are. This ends up creating a smoothing effect from soft clustering. Take a look at the 'rolling guidance filter' for an effect when used iteratively.
Really simple filters can yield very impressive results on images.
This makes me think of deep learning "style transfer" effects. They're nice but so are artistic filters that long predate these things and are tremendously simpler.
It seems to create aliasing artefacts at edges. I wonder if this can be suppressed by scaling up the image and the filter radius and then scaling back down?
Indeed, a bilateral filter is this approach of selection but segmenting on signal statistics of the surrounding image rather than just value.
(Usually variance for denoising, contrast for non-blocking sharpening, L2 metric like PSNR for deblocking, luminance for chroma artifact reduction after upscaling.)
Fascinating. Couldn't help but think that the result looks a lot like some recent game shaders, for grass and landscape, like seen in the witness. Maybe Jonathan can chime in.
https://pro.arcgis.com/en/pro-app/help/analysis/geostatistic...
https://en.wikipedia.org/wiki/Inverse_distance_weighting