> at the cost of placing a convnet in the middle (so much higher computational cost, which can be offloaded to a GPU/AI accelerator if available)
It also makes it harder to undo the effects of the demosaicking algorithm, which may be important if you're doing things like subpixel superresolution.