A research team from the University of California, Los Angeles has published a paper in the latest issue of Nature stating that they have developed a new type of image generator that uses light beams instead of traditional computing hardware to generate images. Compared to standard artificial intelligence (AI) tools, the energy consumption of this device to generate an image can be reduced to one in 100000, only a few millijoules. At present, most AI models that generate images based on text usually use the so-called "diffusion" method: AI first learns a large number of images, masters how to gradually destroy them with statistical noise, and then encodes these changing patterns into a set of rules. When receiving new noisy images, AI can perform reverse operations based on these rules, go through multiple steps of processing, and finally generate coherent images that match the text instructions. However, generating high-resolution realistic images requires a large number of diffusion steps, high computational power requirements, and huge energy consumption. Taking OpenAI as an example, its new image generator generated over 700 million images in its first week of operation. To meet this demand, a large amount of energy and water are required for powering and cooling the machines running the model. The new optical AI image generator may alleviate the problem of huge energy consumption. The research team stated that traditional digital diffusion models require hundreds to thousands of iterations to generate images, while the new system does not require any additional calculations except for initial encoding. The system first uses a digital encoder trained on a publicly available image dataset to generate static encoding that can be converted into images. Subsequently, with the help of a liquid crystal screen called "Spatial Light Modulator (SLM)", the code was physically imprinted onto a laser beam. When the laser passes through the second SLM used for decoding, the target image will be instantly presented on the screen recorded by the camera. In testing, the team used the new system to generate simple images such as black and white numbers (1 to 9), basic clothing, and Van Gogh style full-color paintings. The results show that the generated effect is roughly equivalent to that of traditional image generators. When drawing Van Gogh style images, the new system consumes only a few millijoules of energy per image, while traditional diffusion models require hundreds or even thousands of joules. The team stated that although this system still needs further optimization to replace widely used image generation tools, its low-power characteristics are particularly suitable for application scenarios such as wearable devices (such as AI glasses). (New Society)