Generative artificial intelligence (AI) tools use a lot of energy and water. This is because they need powerful computers that produce a large carbon footprint and require water to cool the equipment. This raises concerns about their sustainability, as water and energy are limited resources needed by humans, animals, and plants.
Researchers at UCLA have developed a new technology that could make generative AI more sustainable. They created models that generate images using photonics, a method that uses light for computing instead of electricity, which is used in traditional electronics. This approach, detailed in a study published in Nature, reduces energy use significantly. Unlike current AI models that need many steps to create images, the new system uses light to produce images in one step, making it faster and more efficient.
Optical generative models for efficiency and versatility
The UCLA system combines a digital encoder (a program that converts data into a format for processing) and an optical decoder (a device that uses light to interpret data). This setup generates images quickly without the repeated calculations required by other AI models. The system is also flexible, allowing the same hardware to handle different tasks with simple updates. Tests showed it can create images of digits, fashion items, butterflies, and human faces, matching the quality of advanced digital models. It even produced artwork similar to Vincent van Gogh’s style in one step, compared to 1,000 steps for traditional models, using much less energy.
This technology also offers privacy benefits. Images can be encoded with specific light wavelengths, only viewable with matching decoders, like a lock and key. It could be used in devices like smart glasses or mobile phones, enabling fast, low-power AI. By reducing energy and water use, this approach supports sustainable AI for applications like medical imaging and edge computing (processing data on local devices).
