To what extent does this reduce image file size? Could I use this as an alternative to dithering for lazily loading large images on a website, for example?
In the thumbnail demo, the LQIP-custom approach (simple resize to low-res jpg thumbail+optimize jpg) approach preserves the more salient features better and has compression on-par-or-better than SQIP with lower processing times. So in my opinion the simple extreme resize+jpgoptim is preferable for thumbnails.
Thumbnails are only small part of LQIP story though and I can imagine RH12503/Triangula having much nicer results for larger images than fogleman/primitive. OP should consider writing an axe312ger/sqip plugin.
I triangulated a 1988×1491 jpg using 10,000 points and managed to reduce the size to 20% of the original size, but the triangles could still obviously be seen.
It still would be cool to see this compared to those in Low Quality Image Placeholder implementations and find out if the extra work on nicer aethetics is preserved when the blur applied.