Optical Generative Models (www.nature.com)

Researchers demonstrated “optical generative models” that synthesize images all‑optically by combining a shallow digital encoder with a reconfigurable free‑space diffractive decoder. Inspired by diffusion models, a small digital network (three fully connected layers) maps random 2D Gaussian noise into 2D phase patterns—optical generative seeds—which are precomputed and loaded onto a spatial light modulator (SLM). Under coherent illumination these phase seeds propagate through a jointly trained diffractive decoder (optimized with, e.g., 400×400 learnable phase features spanning 0–2π) to produce images on a sensor. The encoder and diffractive decoder are trained using pairs generated by a frozen teacher denoising diffusion probabilistic model (DDPM). The team experimentally demonstrated monochrome and multicolour generation (multiwavelength illumination) of MNIST digits, Fashion‑MNIST, Butterflies‑100, Celeb‑A faces and Van Gogh–style artworks with image quality metrics (IS, FID) and classifier-based tests comparable to digital generative models. Significance: because the optical decoder is static after training and synthesis is performed by free‑space light propagation, snapshot image generation requires no electronic compute during inference apart from illumination and SLM refresh (frame‑rate limited), offering a potentially orders‑of‑magnitude improvement in energy efficiency and scalability for generative AI tasks. The approach is hardware‑flexible (free‑space or integrated photonics), supports multicolour outputs, can optimize for diffraction efficiency during training, and could enable low‑power, high‑throughput content generation, image/video synthesis, or edge AI inference.