Generative AI tools, such as creating art, writing code, and drafting emails, are eating up more and more energy. That’s a problem if you want to scale them smartly. So researchers led by Aydogan Ozcan at the University of California Los Angeles decided to bypass the computational heavy lifting entirely, and they found a clever way using light, reports Tech Xplore.
The usual AI image generation relies on diffusion: train a model by repeatedly applying noise until images vanish, then reverse that process for the prompt. It works, but it’s energy-hungry. Instead, this team builds noise using a digital encoder (trained on public datasets), then uses a spatial light modulator (SLM) to imprint that noise onto a laser beam. A second SLM decodes the beam into the final image. The result? The heavy work happens optically, not electronically. Most of the electricity-guzzling computation goes away.
Shiqi Chen, the lead author, points out that the optical generative model can churn out countless images with almost no computing power, making it both scalable and energy-efficient.
They tested it on a variety of visuals, e.g., celebrity portraits, butterflies, and even full-color Van Gogh-style artworks. The result is comparable in quality to traditional AI generators, but made with a fraction of the energy.
What this really means is that we’re looking at a breakthrough that cuts AI’s carbon footprint without sacrificing output quality. It could power VR and AR content, or even work on smartphones or smart eyewear. Fast, low-energy image generation is now within reach.