Deep thoughts on deep image compression

Deep image compositing has, in the last decade, become an industry standard approach to combining multiple computer-generated elements into a final frame. With rich support for multiple depth-specified samples per-pixel, deep images overcome many of the challenges previously faced when trying to combine multiple images using simple alpha channels and/or depth values. A practical challenge when using deep images, however, is managing the data footprint. The visual fidelity of the computer generated environments, characters and effects is continually growing, typically resulting in both a higher number of elements and greater complexity within each element. It is not uncommon to be using "gigabytes" to describe the size of deep image collections which, as more and more visual effects facilities establish a global presence, introduces a significant concern about timely overseas data transfer. Further, as deep images flow through compositing networks, the high sample count contributes to longer processing times. Our observation is that, with a richer contextual understanding of the target composite, systems - both automatic and artist-controlled - can be built to significantly compress deep images such that there is no perceptual difference in the final result.