As MIT Technology Review put it, “generating an image using a powerful AI model takes as much energy as fully charging your smartphone.” That figure came from a single study and describes one particularly heavy model, not image generation in general but a striking comparison is it.

The study behind the number

The figure traces to a paper called “Power Hungry Processing: Watts Driving the Cost of AI Deployment?”, posted online in November 2023 by Sasha Luccioni and Yacine Jernite of Hugging Face and Emma Strubell of Carnegie Mellon University. Many energy studies measure the one-off cost of training a model. This team measured something different: the cost of using a model afterwards, the part that repeats every time someone types a prompt.

The scope was broad. The researchers tested 88 models across 30 datasets and 10 tasks. Image generation topped the table by a wide margin. In the authors’ own summary, they wrote: “We found that Stable Diffusion XL uses nearly 1 phone charge worth of energy per generation.” That one model, the least efficient they tested, is where the smartphone headline comes from.

When MIT Technology Review reported the finding, the paper had not yet been peer reviewed, and the smartphone figure describes Stable Diffusion XL. Read it as one careful measurement of a heavy model, not a universal rule about every image tool.

Why a single image costs so much

The reason a picture costs more than a paragraph comes down to how each task works. Text generation builds an answer one piece at a time. Image generation starts from random visual noise and cleans it up over many passes until a clear picture appears. Every pass runs the full model, and that repeated clean-up is where the energy goes.

The contrast in the study is stark. Making a thousand text answers used only about 16 per cent of a smartphone charge, against nearly a full charge for a single image from the heaviest model. The study also found that a large all-purpose model can use about 30 times more energy than a smaller, single-purpose model doing the same simple sorting task.

The smartphone figure is not fixed. It depends on the model, the resolution, and the hardware.

Scale changes the arithmetic

Any single image uses very little energy. The concern the study raises is what happens when you multiply it. Making a thousand images with Stable Diffusion XL produced about the same carbon as driving 4.1 miles in a petrol car. A single image barely registers; the arithmetic only bites at scale.

The paper also found that use at scale can rival the cost of building the model in the first place. For one family of models, the authors estimated it would take between 200 and 500 million uses to match the energy spent on training.

Images may only be today’s ceiling. A later analysis noted that a high-quality five-second AI video can use roughly 700 times the energy of a high-quality image, which makes the per-image figure look more like a floor than a ceiling.

Jesse Dodge, a research scientist at the Allen Institute for AI who was not part of the study, told MIT Technology Review that the results show “this idea that the new wave of AI systems are much more carbon intensive” than what existed even a few years ago.

What the researchers say should change

The study’s practical point is about disclosure. Someone choosing one image tool over another has no way to compare energy costs, because most companies do not publish them. As Luccioni put it, “Every time you use an AI to generate an image, you don’t know how much energy that’s using.” Energy can be estimated for open models but remains largely opaque for closed commercial ones.

Hugging Face has since built an AI Energy Score that rates models on efficiency, meant to sit alongside the accuracy scores the field already watches. Her argument is for pressure and rules rather than guesswork. As one example, she suggests that “if your model is being used by, say, 10 million users a day or more, it has to have an energy score of B+ or higher.”

Whether efficiency reporting becomes as routine as the accuracy scores every model already advertises, or stays optional while daily generation counts keep climbing, remains unsettled.