Optical generative models for novel image synthesis from noise

We present an optical generative model inspired by diffusion models, where a shallow digital encoder rapidly converts random noise into phase patterns that serve as optical seeds for a desired data distribution. A jointly-trained, reconfigurable free-space optical decoder processes these seeds through optical wave propagation and diffraction, generating novel images consistent with the desired data distribution. We demonstrate the versatility of this method across diverse datasets, including handwritten digits, fashion items, butterflies, human faces, and artworks, where the optical generative models achieved image quality and diversity comparable to conventional diffusion model-based models. Experimentally, we realize optical generation of previously unseen images under visible light and extend this capability to produce Van Gogh-style artworks using both monochromatic and multi-wavelength illumination. These findings highlight optical generative models as a promising pathway towards scalable and energy-efficient inference, harnessing the speed, inherent parallelism, and low power consumption of optics/photonics for AI-driven novel content creation.