The bottleneck might not be the generation itself but the digital preprocessing. Most of the power in current image generation goes to the transformer computations, not the final pixel output.
Still, optical computing for specific ML operations has been promising for years. Curious what the error rates look like compared to digital implementations.