re: parallelization, could a crude 3Dfft-based postprocessing achieve a slightly improved result relative to the current splat-ish approach while still being a fast-running approximation?

i.e. train a very small ML model on various camera parameters vs resulting reciprocal space transfer function.