I think accurate film emulation would require a fair amount of training material pairs (digital/film) to learn the transformation between colours, colour/scene dependent dynamic range compression, and other artefacts like local contrast. The paper mentions using 4000 training pairs for their HDR+ example