Fun fact: d3.blurImage was extracted and refined from D3-contour's code. For faster and more accurate values, you might want to use d3.blur2 on a raster of numbers, rather than blurImage on a raster of pixels.

Haha, good to know! The literal blur approach is not efficient anyways. I featured it mainly for sensemaking. For an efficient implementation, I suggest recoding the fast KDE approach on a byte array representing a grid of heights. It boils down to a fast implementation of a high precision Gaussian blur, and it dodges the aliasing problem we see below, which is due to the height being coded on 8 bits (i.e., from 0 to 255 as a RGB channel).

blur2 is not a “literal blur”—this would indeed be very slow. It uses the very fast stackblur algorithm, and its performance is comparable to fast-kde. See https://observablehq.com/@fil/kde-estimation for a side-by-side of the two codes and results. The algo is described in https://observablehq.com/@d3/d3-blur