Something I've wondered is whether technology like this could eventually be self-defeating for hardware manufacturers. Rather than the evolution of graphics deriving from improved accuracy of the optical simulation fuelled by advances in computational power, it may instead derive from optimising subjective video quality, similarly to video codecs.
While accurately simulating optics is needfully computationally expensive and gives special-purpose graphics hardware an advantage, it's not clear that psychologically subjective high quality graphics (i.e. generating visuals which are inaccurate but convincing to humans) has such a need.
>While accurately simulating optics is needfully computationally expensive and gives special-purpose graphics hardware an advantage, it's not clear that psychologically subjective high quality graphics (i.e. generating visuals which are inaccurate but convincing to humans) has such a need.
What you're describing is rasterization, which is what the industry standard is (at least for games) for decades.
Techniques used to create realism with rasterization (e.g. normal mapping; shadow mapping; screen-space anti-aliasing) are still simulations of optics, just not entirely faithful ones.
Generating visuals with an autoencoder, albeit hinted by noisy physically-based raytracing, is not an optical technique; detail is generated from a visual statistical model, not an optical simulation.
The raytracing is, but you don't see the result of the raytracing, you see the output of a neural network inventing detail based on higher definition training data. It's like seeing some blurry dots through a microscope, then drawing a sketch of detailed cells, based on your memory of pictures you've seen. The microscope is an optical system, but the sketch is the result of memory and style transfer, not simulation of optics. Hypothetically, you could have no understanding of the behaviour of light in producing the detailed sketch.
I think the success of deep learning is quite unfortunate. There are a lot of areas where "throw an ANN at it" has become a go-to even though they're basically inscrutable blackboxes with minimal theoretical guarantees.
While accurately simulating optics is needfully computationally expensive and gives special-purpose graphics hardware an advantage, it's not clear that psychologically subjective high quality graphics (i.e. generating visuals which are inaccurate but convincing to humans) has such a need.