Understanding the Energy Cost of AI-Generated Images
As MIT Technology Review put it, “generating an image using a powerful AI model takes as much energy as fully charging your smartphone.” This striking comparison comes from a single study focused on a particular heavy AI model, rather than a general rule applicable to all image generation tools. Yet, it sheds important light on the environmental impact of today’s AI technologies.
The study behind the number
The energy figure originates from a November 2023 paper titled “Power Hungry Processing: Watts Driving the Cost of AI Deployment?” authored by Sasha Luccioni and Yacine Jernite of Hugging Face along with Emma Strubell from Carnegie Mellon University. Unlike many prior analyses that focus primarily on the one-time cost of training AI models, this research uniquely measures the ongoing energy consumption incurred every time a model is used after deployment – for example, when a user inputs a prompt to generate an image or text.
The scope of the study was extensive: the researchers tested 88 models across 30 datasets and 10 different tasks. Image generation clearly stood out as the most energy-intensive activity. Specifically, the authors highlight that “Stable Diffusion XL uses nearly 1 phone charge worth of energy per generation,” which is the basis for the now-famous smartphone energy comparison. This particular model was identified as the least energy efficient among those tested.
At the time MIT Technology Review reported these findings, the paper was still awaiting peer review. It’s important to interpret this smartphone charge estimate as a precise measurement for one demanding model, rather than a universal benchmark for all AI image generation tools.
Why a single image costs so much
The key reason that creating a single AI-generated image consumes more energy than generating a text response lies in the fundamental mechanics of these tasks. Text generation models produce output sequentially, composing answers one token at a time. In contrast, image generation begins with random visual noise and then progressively refines this noise through many iterative passes until a clear image forms. Each pass requires running the entire model, amplifying the cumulative energy cost.
The study’s findings illustrate this contrast vividly: generating a thousand text responses consumed only about 16 per cent of a smartphone charge, compared to nearly a full charge for just one image created by the heaviest model. Additionally, the research revealed that a large, all-purpose AI model can consume nearly 30 times more energy than a smaller, specialized model performing the same simple task.
It’s essential to understand that the smartphone charge figure varies depending on factors such as the specific model used, image resolution, and the hardware powering the process.
Scale changes the arithmetic
While the energy consumption for a single image might seem minimal, the environmental implications become significant when scaled. For instance, generating a thousand images with Stable Diffusion XL produces carbon emissions comparable to driving 4.1 miles in a petrol car. This highlights how cumulative usage dramatically amplifies the energy footprint.
Interestingly, the paper also suggests that the energy consumed through widespread model usage can rival, or even surpass, the energy originally invested in training the model. For one model family, the authors estimate that between 200 and 500 million uses would be required to equal the training energy cost.
Moreover, image generation may represent just the beginning of this energy challenge. Follow-up analyses indicate that creating a high-quality five-second AI video might require roughly 700 times the energy of a single high-quality image, suggesting that video generation could significantly escalate energy demands.
Jesse Dodge, a research scientist at the Allen Institute for AI who was not involved in the study, told MIT Technology Review that these results emphasize “this idea that the new wave of AI systems are much more carbon intensive” than those just a few years ago.
What the researchers say should change
The study’s authors emphasize the need for transparency in energy consumption. Currently, users choosing between AI image generation tools have no clear way to compare their environmental impacts, as most companies do not disclose energy usage metrics. As Luccioni explained, “Every time you use an AI to generate an image, you don’t know how much energy that’s using.” While energy consumption can be roughly estimated for open-source models, it remains largely opaque for proprietary commercial offerings.
In response, Hugging Face has developed an AI Energy Score that evaluates models based on their efficiency. This rating aims to complement accuracy scores, encouraging industry-wide accountability. Luccioni has proposed concrete guidelines, suggesting that if a model is used by, for example, 10 million users daily, it should maintain an energy efficiency rating of B+ or higher. Such measures would incentivize developers to optimize their models’ environmental footprint.
Whether energy efficiency reporting becomes as standardized and routine as accuracy reporting remains an open question. Meanwhile, as daily use of AI-generated content continues to surge, the environmental costs warrant ongoing scrutiny and proactive management.
For further reading, visit Here.
