I don't have any knowledge of this specific implementation, but my guess would be that it is a question of optimization. If you were to implement something simi

I don't have any knowledge of this specific implementation, but my guess would be that it is a question of optimization. If you were to implement something similar to what's physically happening in a camera with a long exposure, you'd have to do a large number of oversamples (e.g. generate 10 sub-frames for every output frame) and merge them together. That's a ton of extra rendering. If, on the other hand, you can get the 3D renderer to generate "smeared geometry" (based on its knowledge of the motion speed, direction, and virtual shutter duration) and render each output frame once, that will get you faster render times.