Here’s a couple of things that would be really useful:
-
Faster data transfer between Blender image data and Numpy. Currently it takes seconds to transfer 2k textures. Ideal solution would be a Numpy proxy, that has next to zero overhead (Numpy using the same data pointer as Blender internally). There was already an attempt at solving this but I don’t know if it went anywhere: Bpy.data.images perf issues
-
Raycasting with packets (multiple, like 100k rays inside a vector which would take only 1 function call) instead of single function calls would make the BVH much, much more powerful from the Python side. Again, Numpy would be a good solution here. I’m assuming that the overhead here is function calls, as you’re making them millions per second.