How much energy does an AI-generated image use?
Generating an image with AI costs roughly 9 Wh. Loading a photo that already exists costs roughly 0.1 Wh. That is a factor of 90, and the first number is the shakier of the two. The only published per-model measurements span 1.4 to 11.5 Wh for the same task, driven by model size alone. The second number is shakier still, for a completely different reason. Here are both calculations, term by term.
The 90:1 ratio between the two bars is an order of magnitude, not a measurement. On the AI side everything hangs on the model: two versions of Stable Diffusion, measured on the same rig, sit a factor of 8 apart. On the photo side the cost does not really divide by image at all. What holds either way: generating costs ten to a hundred times more than loading. It is also the only common AI action whose compute outweighs the screen that displays it.
Figures updated on June 30, 2026
Where does the 9 Wh figure for a generated image come from?
From a bracket, not a measurement. This site uses 9 Wh, which is 30 text prompts at 0.3 Wh each. Measurements do exist, but they cover open models. Luccioni, Jernite and Strubell ran a thousand generations per model on a single GPU, then published the bill. Stable Diffusion v1.5: 1.38 kWh per thousand images. Stable Diffusion XL, the least efficient model in their panel: 11.49 kWh. That works out at 1.4 Wh and 11.5 Wh per image. Same task, same rig, a factor of 8 for model size alone. No commercial service publishes its own number, not Midjourney, not DALL·E, not Imagen. Google does it for text; nobody does it for images. An open model on one GPU is also not a production service: no idle servers, no data centre overhead, no queue. That gap always runs the same way, which is why this site sits its estimate near the top of the bracket rather than in the middle.
What does loading an existing photo actually cost?
The calculation takes one line, and it does not mean much. A web photo weighs 1 to 2 MB. The reference intensity for fixed-line networks is 0.06 kWh per gigabyte, measured by Aslan and co-authors for the year 2015. Two megabytes therefore give 0.12 Wh, which is where the 0.1 Wh on the left bar comes from. Two things undercut it. The same study found that intensity halving every two years. Carried forward to today, the same rule would give a far smaller figure. More fundamentally, Mytton showed in 2024 that network energy is not proportional to the data carried. A router draws nearly as much idle as loaded. The marginal cost of your photo is close to zero. This site keeps 0.1 Wh as a deliberately generous upper bound: it favours the conventional action, and the comparison still tips the other way.
How long must you look at the image before the screen catches up with the compute?
Fourteen seconds for a loaded photo, twenty-two minutes for a generated one. The working uses the 25 Wh a laptop draws per hour of use (ADEME). For the photo: 0.1 ÷ 25 = 0.004 hours, or 14 seconds. For the generated image: 9 ÷ 25 = 0.36 hours, or 22 minutes. A photo you scroll past has already cost more in screen time than in transfer before you finish the caption. A generated image stays dominated by its own compute for a good twenty minutes. For a text prompt the same crossover lands at 43 seconds. Image generation is the one everyday AI action where the data centre outweighs the device showing the result.
Related equivalents
Sources
- Luccioni, Jernite & Strubell : « Power Hungry Processing: Watts Driving the Cost of AI Deployment? » (arXiv 2311.16863, nov. 2023 ; ACM FAccT, juin 2024) : 1,38 kWh pour 1 000 images avec Stable Diffusion v1.5, 11,49 kWh avec Stable Diffusion XL, le modèle le moins efficace du panel
- « The Hidden Cost of an Image » (arXiv 2506.17016)
- Hugging Face : AI Energy Score / S. Luccioni (énergie image & vidéo)
- Aslan et al. (2018), Journal of Industrial Ecology : « Electricity Intensity of Internet Data Transmission », 0,06 kWh par gigaoctet pour les réseaux fixes en 2015, intensité divisée par deux tous les deux ans
- Mytton (2024), Journal of Industrial Ecology : « Network energy use not directly proportional to data volume », l'énergie d'un réseau ne se divise pas par gigaoctet transporté
- ADEME : consommation des appareils ménagers
Frequently asked
How much energy does an AI-generated image use?+
Roughly 9 Wh under this site's assumptions, or 30 text prompts. The only published per-model measurements run from 1.4 Wh for Stable Diffusion v1.5 to 11.5 Wh for Stable Diffusion XL (Luccioni et al., FAccT 2024). No commercial service publishes its own figure.
Does generating an image use as much energy as charging a phone?+
About half as much. A phone charge costs 19 Wh (ADEME); an image costs roughly 9 Wh. The headline that made the rounds came from the hungriest model in the 2024 study: 11.5 Wh against the 22 Wh charge its authors assumed. Half a charge, not a full one.
Is reusing a photo better than generating one?+
On energy, yes, by a factor of 10 to 100 depending on the model. On a single action the gap stays tiny: generating one image costs about what an LED bulb burns in 54 minutes. What moves the needle is the number of images, never the image.
A number does not tell you whether to generate the image. An illustration made in thirty seconds that replaces a photo shoot is not the same call as one regenerated six times to pick a favourite. That judgement belongs to a different trade: the GHIS studio advises small and mid-sized companies on how they use AI.