(The linked web app doesn’t work on mobile in portrait mode, sorry!)
The biggest issue with this trick is that different engines calculate the filters differently, thus turning an okay-ish image into something that looks like a glitch.
On finnicky engines, I think if I were to seriously implement this for a project that needed to support arbitrary images, I'd do the dithering server side (assuming it's possible to develop some heuristics to select the correct transformation based on image type (text, low contrast, blurry, etc)), serve those to users, and allow them to customize the colouring filters. That way the dithering looks as good as it can per image, but it can then still be stylized to a user's preferences.
Hm, for this - yes I think it would make sense. That being said there are libraries that apply proper dithering without relying on SVG filters.
Also, I haven't tested this with canvas which I imagine would be much more consistent.
If I were to do this... more seriously, I'd keep some parameters exposed in the shortcode/component rendering the dithered images (or data-attrs) so that I could fine-tune them on a case-per-case basis. (I originally wanted to replace the bio photos on my site, like this one https://untested.sonnet.io/notes/wislawa-szymborska/)
There's a ton of articles about this but for something more fun and even better looking I recommend the Coding Train video on Weighted Voronoi Stippling: https://www.youtube.com/watch?v=Bxdt6T_1qgc
(The linked web app doesn’t work on mobile in portrait mode, sorry!)
The biggest issue with this trick is that different engines calculate the filters differently, thus turning an okay-ish image into something that looks like a glitch.