Behind each image generated by artificial intelligence is hiding a huge energy cost. Models like those of Openai, for example, can consume hundreds or thousands of joules for each individual image. Only in a week, the Openii image generator produced 700 million imageswith a significant environmental imprint in terms of energy consumption and water.
But now, a team of researchers fromUniversity of California, Los Angeles (UCLA) has developed a radically different system. Posted on Naturethis new approach generates images Using laser bundles and optical modulators, breaking down the energy consumption to a few Millijoule for image – millions of times less than traditional digital systems.
How artificial intelligence based on light works
The conventional method of the generation of images AI is based on a process called spreadwhich starts from a random image and, through hundreds or thousands of stepsgradually reduces the noise to obtain the final result. This technique, albeit effective, is extremely energetic.
On the contrary, the UCLA system generates an image in one optical stepwithout requiring any additional digital computation after the initial coding phase. In practice, a digital encodertrained on a standard dataset, creates one phase scheme – a sort of mathematical blueprint – which is projected on a Space light modulatora liquid crystal device.
When a bundle of laser light crosses this modulator, the pattern is transmitted to a second component, called diffractive decoder. At this stage, the image takes shape directly on a sensor, without the need for GPUs or complex digital calculations.
Test on portraits, butterflies and Van Gogh
The team test this model on subjects such as Human faces, handwritten figures, butterflies and works in Van Gogh’s style. While not reaching absolute perfection, the resulting images were statistically comparable to those generated by digital models based on diffusion.
Among the tested methods, the phase coding He got the highest score in terms of quality and variety of images. In particular, he produced Visually superior results for complex scenesas paintings with an elaborate artistic style.
Second Shiqi Chenthe first author of the study, “our optical generative models can summarize an unlimited number of images with a minimum energy consumption, offering a scalable and sustainable alternative to digital artificial intelligence”.
In addition to the energy aspect, technology opens new perspectives also in terms of Security and privacy. Each generated image has a unique optical pattern that can only be decoded through one compatible surfacecreating a sort of physical key mechanism and lock. This approach could prove useful for applications such as the safe communication or thecontrast.
The system exists in two variants: the Snapshot versionwhich generates the image in one optical step, and the iterative versionwhich replicates more closely the diffusion models, refining the images through subsequent impulses of light.
In perspective, this technology could be miniaturized in Integrated photo chipsreplacing bulky laser with nanophabic surfaces. This would allow integration into devices such as AR/VR glasses or tools of medical diagnostics.
The interest in solutions to the most sustainable grows in parallel with the spread of generative models. A 2023 study has estimated that the training of large models can issue as much as there are thousands of transcontinental flights. In this context, the Optical generation of images is proposed as a concrete solution for contain the environmental impact artificial intelligence.
As he said Aydogan Ozcansenior author of the study:
Our work shows that the perspective can be used for large -scale generative generative tasks.
Although technology is still young and not without limits – such as the alignment of the components or the resolution of the modulators – the potential is enormous.
In short, artificial intelligence could soon stop consuming energy at unsustainable rhythms And start “painting” using only the light.
