dithering with a color palette gives unpleasant / wrong result

[olaf-rogalsky]

In the following example

import ColorSchemes, DitherPunk

# numeric matrix
mat = let n = 61
	[2 * x^2 + y^2 for x in -n:n, y in -n:n]
end;

# image of matrix with viridis color scheme
img = get.(Ref(ColorSchemes.viridis), mat, Ref(extrema(mat)))

# palette of 8 colors from viridis color scheme
c8 = get(ColorSchemes.viridis, 0.0:1//7:1.0)

# dither image
dither_img = DitherPunk.dither(img, c8)

the dithered image does not look right. For comparison I also show the result of https://ditherit.com with the same palette.

img	dither_img	ditherit

This behaviour somehow depends on the image size. For a small image size let n = 31 the results looks better

img	dither_img

Except for Atkinson all error diffusion algorithms have this problem

let mat, img, n = 91, dithers
	mat = [2 * x^2 + y^2 for x in -n:n, y in -n:n]
	img = get.(Ref(ColorSchemes.viridis), mat, Ref(extrema(mat)))
	dithers = [Atkinson(), SimpleErrorDiffusion(), FloydSteinberg(), Burkes(),
               Stucki(), JarvisJudice(), Sierra(), TwoRowSierra(),
               SierraLite(), Fan93(), ShiauFan(), ShiauFan2()]
	hvcat(4, reshape([dither(img, d, c8) for d in dithers], (4,3))...)
end

[adrhill]

Interesting. Could you try what happens when passing the additional kwarg dither(img, c8, clamp_error=false)?

[olaf-rogalsky]

Interesting. Could you try what happens when passing the additional kwarg dither(img, c8, clamp_error=false)?

the third image from left in the first row is FloydSteinberg() with clamp_error = false