It is pretty much Helge's pixel particles but implemented through (slower) ICE implementation. Basically it uses barycentric coordinates to get the equivalent position in a 2d (uv) and 3d (xyz) triangle.
I remember some bottlenecks like: - Having to create a triangulated clone of the geometry, because it makes difference to how stuff works in 2d. I rember trying to avoid it with all my efforts, but it was useless... - Outputing locations instead of positions was a problem, but now thinking about it I cant remember why. It should be quite possible with reinterpret location on geo, right? I dont know... It should be fairly documented if you want to dive into it (at least that is how I recall it).
