The paper describes the development of a fully all-optical chip called LightGen designed to generate semantic visual content using light[1]. The authors integrated more than 2 million photonic "neurons" on the chip using diffractive metasystems and 3D packaging, enabling high-resolution image processing directly on the chip[2][3]. They introduced the concept of optical latent space, which allows the network dimension to be changed by optical means without electronic conversions[1][4]. For training, they used Bayesian (probabilistic) algorithms tailored to the physical properties of the optical system[1][4]. Experimentally, LightGen has demonstrated capabilities including high-resolution semantic image generation, de-noising, style transfer, and three-dimensional image generation and manipulation[1][3]. The measured end-to-end computation speed and energy efficiency were each more than two orders of magnitude higher compared to state-of-the-art electronic chips according to the authors[1][4]. The results are presented in the journal Science and provide evidence that all-optical solutions can fundamentally speed up and streamline large-scale visual generative models[1].