Yeah that would be cool! I guess that's actually not that hard to implement since I just need to do some sort of 2D convolution with the aperture shape, instead of just cv2.blur.

The tricky thing is that in order to get different amount of blur at different depths, I interpolate between precomputed 3 levels of blur, since that vectorizes well at -O3, I'm not sure how to compute a per-pixel varying blur efficiently without going to the GPU.

That's what I thought this was doing, based on the description. The reason that Meet backgrounds look fake is because they do a Gaussian Blur instead of simulating a lens. A physical bokeh will blow out highlights, whereas a Gaussian Blur will smooth out bright spots.