It's probably doing some equivalent of solving a hard inverse problem approximately using numeric method, likely with as lest as many unknowns per pixel as the image, in a noisy domain, and with an expensive cost function for the optimization.
Not saying they are doing exactly that, but something in that realm/scale. 100kOps per pixel is really not that much in those kind of problems.
Not saying they are doing exactly that, but something in that realm/scale. 100kOps per pixel is really not that much in those kind of problems.