---
title: "What is the carbon footprint of an AI-generated image?\nCO2, water and comparisons"
description: "An AI-generated image emits about 1.3 gCO2e, or 9 metres by car; studies range from 0.04 to 5.3 gCO2e. We correct the phone charge claim and estimate water use."
canonical: https://projetcelsius.com/en/blog/ai-image-carbon-footprint/
author: "Guillaume Pakula"
datePublished: 2026-10-05
dateModified: 2026-10-05
language: en
translationOf: https://projetcelsius.com/blog/empreinte-carbone-image-ia/
---

# What is the carbon footprint of an AI-generated image? CO2, water and comparisons

**Author**: Guillaume Pakula
**Published**: 2026-10-05
**Section**: Explained
**Canonical URL**: https://projetcelsius.com/en/blog/ai-image-carbon-footprint/
**French original**: https://projetcelsius.com/blog/empreinte-carbone-image-ia/
**Source**: Projet Celsius (https://projetcelsius.com/en/), a climate strategy consultancy of engineers and PhDs based in Paris and Marseille.

> An AI-generated image emits **about 1.3 gCO2e**, the equivalent of **9 metres by car** and 10 times the emissions of a [written question](https://projetcelsius.com/en/blog/chatgpt-query-energy-consumption/). The energy it uses equals **15% of a phone charge**: the often-quoted "whole charge" was the figure for the most energy-intensive generator. Across all studies, measurements range from **0.04 to 5.3 gCO2e**, and none of these measurements covers ChatGPT, Gemini or Midjourney.

## Key takeaways

- An AI-generated image emits **about 1.3 gCO2e**, the equivalent of 9 metres by car.
- Across all studies, measurements range from **0.04 to 5.3 gCO2e** per image.
- The energy used by an average image equals **15% of a phone charge**.
- **700 million images** would represent about 930 tCO2e if each used 2.9 Wh.

Generating an image with artificial intelligence (AI) emits **about 1.3 gCO2e** (grams of CO2 equivalent) with global average electricity, the equivalent of **9 metres by car**. That is the average in the [most cited study](https://arxiv.org/abs/2311.16863), published in 2024: 2.9 Wh (watt-hours) of energy per image. The **carbon footprint of an AI-generated image** is therefore **10 times** that of a [written question put to ChatGPT](https://projetcelsius.com/en/blog/chatgpt-query-energy-consumption/), while a [5-second video](https://projetcelsius.com/en/blog/ai-video-environmental-impact/) emits as much as **31 images**.

Across all studies, published measurements range from **0.04 to 5.3 gCO2e** per image, and on the same machine the spread between generators is **1 to 46**. OpenAI, Google and Midjourney publish **no measurement** for their services: the values available come from open models, whose code is public, tested in the laboratory between 2023 and 2025.

## How much CO2 does an AI-generated image emit?

### About 1.3 gCO2e per image on average

The most cited measurement comes from 3 researchers at Hugging Face and Carnegie Mellon University. They had several open generators produce 1,000 images and found **[2.9 Wh per image on average (in French)](https://projetcelsius.com/ia/chiffres/electricite-image-generee/)** and **1.35 Wh for the median generator** (0.62 gCO2e), which splits the models tested into 2 halves. The median is quoted because a few very energy-intensive generators pull the average up.

> **The calculation** Multiplying 2.9 Wh per image by 458 gCO2e per kWh, the [average carbon intensity of global electricity in 2025](https://ember-energy.org/latest-insights/global-electricity-review-2026/) according to Ember, an energy think tank, counting the construction of power plants and the extraction of fuels, gives 1.3 gCO2e. A petrol or diesel car emits 142 gCO2e per km according to [Impact CO2](https://impactco2.fr/), the comparison tool of ADEME (the French Agency for Ecological Transition), so 1.3 g corresponds to 9 metres. This Projet Celsius calculation gives an order of magnitude: it counts the electricity used by the servers, as measured in the laboratory, but not their cooling or manufacture, which are covered in [why AI is bad for the environment](https://projetcelsius.com/en/blog/why-ai-is-bad-for-the-environment/).

### From 0.04 to 5.3 gCO2e across all studies

The lowest measurement, **0.04 gCO2e** (0.086 Wh), is for the lightest of the 17 generators tested in 2025 by [researchers at the University of Florence](https://arxiv.org/abs/2506.17016) on a consumer graphics card. The highest, **5.3 gCO2e** (11.5 Wh), is for the least efficient generator in the 2024 study, run on data centre graphics cards. Between these 2 limits, [MIT Technology Review](https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/) estimated in May 2025 that a recent generator, Stable Diffusion 3 Medium, used **0.29 gCO2e** (0.63 Wh). These values come from 3 different protocols: they give **the range of published measurements**, a spread of 1 to 130, and do not rank the generators against each other.

### An image equals 10 written questions, and a 5-second video equals 31 images

According to [Epoch AI's estimate](https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use), [a written question put to an assistant (in French)](https://projetcelsius.com/ia/chiffres/electricite-question-ia/), also known as a query, uses about 0.3 Wh, or **0.14 gCO2e**, so an average image equals 10 questions. [A 5-second video (in French)](https://projetcelsius.com/ia/chiffres/electricite-video-generee/), measured in 2025 on the [most downloaded open generator](https://arxiv.org/abs/2509.19222), uses 90 Wh, or **41 gCO2e**, which equals **31 images, or 300 questions**. The [environmental impact of AI](https://projetcelsius.com/en/blog/environmental-impact-of-ai/) therefore depends above all on what it is asked to produce.

### No measurement published by OpenAI, Google or Midjourney

Since August 2025, Google has published the energy consumption of a [written question put to Gemini](https://arxiv.org/abs/2508.15734), at 0.24 Wh. For images, in early October 2026, there was **no official data** from OpenAI, Google, Midjourney or Black Forest Labs, the developer of Flux. **The gap with commercial services therefore remains unknown**, in either direction.

## Generator, size, passes and electricity: what makes an image's carbon footprint vary

The result depends on 4 factors: **the generator, the size of the image, the number of computation passes** and **the electricity** that powers the servers.

### The generator: a spread of 1 to 46 on the same machine

The University of Florence researchers tested **17 generators on the same machine**, with the same graphics card. The lightest uses **0.086 Wh per image** and the heaviest 4.08 Wh, a spread of **1 to 46**. The lightest is a model designed to produce an image in a few computation passes, the steps through which a generator gradually refines an image of noise.

### Size and number of passes: up to 4.7 times as much energy

An image 1,024 pixels on a side has **4 times as many pixels** as one of 512. In the same study, that change multiplies the energy used **by up to 4.7 depending on the generator**, and one of the models tested is unaffected. Energy use rises in proportion to the number of passes: when MIT Technology Review ran the same model at 25 and then 50 passes, **consumption doubled**, from 0.63 to 1.2 Wh.

### The country of the servers: 11 times less CO2e with French electricity

The same image emits **0.12 gCO2e** if the servers run on French electricity (41 gCO2e per kWh in 2025 according to [Ember data](https://ourworldindata.org/grapher/carbon-intensity-electricity)), which is **11 times lower** than with the global average. With US electricity, 384 gCO2e per kWh according to the same source, the figure is **1.1 gCO2e**.

## How much electricity and water does a generated image use?

### "One image = one phone charge": where the mistake comes from

The comparison comes from the [first version of the most cited study](https://arxiv.org/abs/2311.16863v1), dated November 2023. It stated that the least efficient generator used as much energy as **950 smartphone charges for 1,000 images**, or nearly one charge per image, on the basis of **12 Wh per charge**. That statement concerned **the most energy-intensive of the generators tested**, but the articles that repeated it applied it to every image.

The US Environmental Protection Agency (EPA), the source of this value, made 2 revisions to it in 2024. The published version of the study uses 22 Wh and puts the most energy-intensive generator at about half a charge; the value in force [since October 2024](https://www.epa.gov/energy/greenhouse-gas-equivalencies-calculator-calculations-and-references) is 19 Wh. With this value, that generator accounts for **60% of a charge**, the average image for **15%** and the median image for 7%.

### How much water does an image use? A few millilitres

**No study measures the water consumed per image.** The order of magnitude can be calculated from the water used per kWh: Google states that a question of 0.24 Wh consumes 0.26 mL, or about **1.1 litres per kWh**, to cool its servers. At that rate, an average image uses **3 mL**. If we add the water consumed to produce the electricity, which averages 3.1 litres per kWh in the United States according to [researchers at the University of California, Riverside](https://arxiv.org/abs/2304.03271), our calculation reaches **12 mL**, a little over 2 teaspoons.

### "2 to 5 litres of water per image": a figure with no identified source

This figure has been circulating since at least April 2025. One instance, dated 2 April 2025, already attributes it to the same Riverside team, whose study contains **no per-image figure**: it estimates that 500 mL of water is needed for 10 to 50 written responses from an older model. In June 2025, the Spanish fact-checking site [Maldita](https://maldita.es/malditatecnologia/20250602/consumo-agua-IA-preguntas-respuestas/) found that no scientific study addressed the subject. The low value, 2 litres, is **170 times** our high estimate. Our article on [the water consumed by AI](https://projetcelsius.com/en/blog/ai-water-footprint/) details the scope of these measurements.

## What do millions of generated images add up to?

### 10 images a day for a year: less than a meal with beef

A person who generates **10 images a day for a year** emits 4.9 kgCO2e, a little less than **a meal with beef** (4.97 kgCO2e according to ADEME). Set against the [average carbon footprint of a person in France](https://www.statistiques.developpement-durable.gouv.fr/lempreinte-carbone-de-la-france-de-1990-2024), 8.2 tCO2e in 2024, this amounts to **less than 0.1%**.

### 700 million images in 9 days in spring 2025

On 3 April 2025, Brad Lightcap, OpenAI's chief operating officer, stated that more than 130 million users had generated **more than 700 million images** since the launch of ChatGPT's new image generator on 25 March, [according to TechCrunch](https://techcrunch.com/2025/04/03/chatgpt-users-have-generated-over-700m-images-since-last-week-openai-says/). That is about 80 million a day. On 19 May 2026, Google announced [more than 50 billion images](https://blog.google/innovation-and-ai/sundar-pichai-io-2026/) produced by its generator, up from 5 billion 7 months earlier, or **about 200 million a day** over the period. Neither of the 2 figures has an independent count behind it.

### 2 GWh and 930 tCO2e if every image used 2.9 Wh

If each of these 700 million images had used 2.9 Wh, those 9 days would have required **2 GWh** of electricity and emitted **about 930 tCO2e** with global average electricity: the equivalent of **450 Paris-New York return trips**. For Google's 50 billion images, produced in about 9 months, the same calculation gives **about 150 GWh**, or 0.03% of the electricity France consumes in a year. This Projet Celsius calculation relies on a strong assumption: neither of the 2 services publishes its consumption, and their generators are not the ones the studies measured.

### 3 ways to emit less when generating images

- **Be precise in your request**, or prompt, before running it: every new attempt is one more image.
- **Use a small size for test runs**: going from 512 to 1,024 pixels multiplies the energy used by up to 4.7.
- **Keep the standard quality** where the tool has a setting for it: doubling the number of computation passes doubles the energy used.

We rank 10 actions, including generating fewer images, in our article on [reducing the impact of AI](https://projetcelsius.com/en/blog/reduce-ai-environmental-impact/), and we examine what 4 AI services publish in [which eco-friendly AI to choose](https://projetcelsius.com/en/blog/most-eco-friendly-ai/). A company can count its generated images among its **purchases of digital services**: see our guide to [AI in the Bilan Carbone® (the French carbon accounting method)](https://projetcelsius.com/en/blog/ai-emissions-carbon-accounting-scope-3/) and the [AI carbon footprint calculator](https://projetcelsius.com/en/tools/ai-carbon-footprint-calculator/).

## Key takeaways

- **About 1.3 gCO2e per image on average**, the equivalent of 9 metres by car; 0.62 gCO2e for the median generator.
- **From 0.04 to 5.3 gCO2e** depending on the study, the generator and the size of the image; 11 times lower with French electricity.
- **15% of a phone charge** for an average image, and 3 to 12 mL of water according to our calculation.
- **No published measurement** from the major services: the volume calculations remain assumptions.

Our article on [AI electricity consumption](https://projetcelsius.com/en/blog/ai-electricity-consumption/) gives the orders of magnitude at world scale, and our [comparison of ChatGPT, Claude and Gemini](https://projetcelsius.com/en/blog/chatgpt-vs-claude-vs-gemini-environmental-impact/) sets out, assistant by assistant, what each provider publishes.

## Frequently asked questions

### How much CO2 does an AI image from ChatGPT emit?

OpenAI publishes no measurement per image. The open generators tested in the laboratory use 2.9 Wh per image on average, or about 1.3 gCO2e with global average electricity, the equivalent of 9 metres by car. Published measurements range from 0.04 to 5.3 gCO2e, and nothing shows whether ChatGPT's generator falls within that range.

### Does generating an AI image use as much energy as charging a phone?

No: an average image represents 15% of a charge, or 2.9 Wh against 19 Wh. The comparison comes from the first version of a study, dated 2023, which referred to the most energy-intensive generator, at 11.5 Wh per image, and assumed 12 Wh per charge. With a charge of 19 Wh, that generator accounts for 60% of one.

### How much water does it take to generate an AI image?

Our calculation gives a few millilitres: about 3 mL to cool the servers, or 12 mL when the water used to produce the electricity is counted too. The figure of 2 to 5 litres per image, which has been circulating since at least April 2025, has no identified source.

### Which AI image generator uses the least energy?

No ranking of commercial services is possible, because no data has been published. Among 17 open generators tested in 2025 on the same machine, the heaviest uses 46 times as much energy as the lightest: 4.08 Wh against 0.086 Wh per image. The lightest is a model designed to produce an image in a few computation passes.

### How many AI images are generated every day?

There is no public worldwide count. OpenAI stated that more than 700 million images had been generated in 9 days in spring 2025, or about 80 million a day. Google announced more than 50 billion images in May 2026, up from 5 billion in October 2025, or about 200 million a day over that period.

## Sources

- **Measurements per image**: [Luccioni, Jernite and Strubell, Power Hungry Processing: Watts Driving the Cost of AI Deployment?](https://arxiv.org/abs/2311.16863v3), published version of 15 October 2024, table 2 and section 4.1, and [first version of 28 November 2023](https://arxiv.org/abs/2311.16863v1); [Bertazzini, Albisani, Baracchi, Shullani and Verdecchia, The Hidden Cost of an Image](https://arxiv.org/abs/2506.17016), University of Florence, 20 June 2025; [MIT Technology Review, We did the math on AI's energy footprint](https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/), 20 May 2025; [IEA, Energy and AI](https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf), April 2025, pp. 45 and 46.
- **Conversion to CO2e and equivalents**: [Ember, Global Electricity Review 2026](https://ember-energy.org/latest-insights/global-electricity-review-2026/), and [Our World in Data based on Ember](https://ourworldindata.org/grapher/carbon-intensity-electricity), carbon intensity of electricity in 2025, life cycle (458 gCO2e per kWh worldwide, 41 in France, 384 in the United States); [ADEME, Impact CO2](https://impactco2.fr/), accessed 2 October 2026; [EPA, Greenhouse Gas Equivalencies Calculator](https://www.epa.gov/energy/greenhouse-gas-equivalencies-calculator-calculations-and-references) and [revision history](https://www.epa.gov/energy/greenhouse-gas-equivalencies-calculator-revision-history); [SDES, France's carbon footprint from 1990 to 2024](https://www.statistiques.developpement-durable.gouv.fr/lempreinte-carbone-de-la-france-de-1990-2024) (SDES is the statistical service of the French Ministry of Ecological Transition); [RTE, Bilan électrique 2025](https://assets.rte-france.com/prod/public/2026-02/Bilan-electrique-2025-principaux-resultats.pdf) (RTE is the operator of the French electricity transmission network). The ADEME, SDES and RTE pages are in French.
- **Written question and video**: [Epoch AI, How much energy does ChatGPT use?](https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use), 7 February 2025; [Google, Measuring the environmental impact of delivering AI at Google Scale](https://arxiv.org/abs/2508.15734), August 2025; [Delavande, Pierrard and Luccioni, Video Killed the Energy Budget](https://arxiv.org/abs/2509.19222), 23 September 2025.
- **Water**: [Li, Yang, Islam and Ren, Making AI Less Thirsty](https://arxiv.org/abs/2304.03271), University of California, Riverside and University of Texas at Arlington; [Maldita, questions and answers on the water consumed by AI](https://maldita.es/malditatecnologia/20250602/consumo-agua-IA-preguntas-respuestas/), 2 June 2025; [article citing the figure of 2 to 5 litres](https://misionesonline.net/2025/04/02/cuanta-agua-consume-la-ia/), 2 April 2025. The Maldita and Misiones Online pages are in Spanish.
- **Volumes**: [TechCrunch, statement by Brad Lightcap (OpenAI)](https://techcrunch.com/2025/04/03/chatgpt-users-have-generated-over-700m-images-since-last-week-openai-says/), 3 April 2025; [Google, announcements at the I/O 2026 conference](https://blog.google/innovation-and-ai/sundar-pichai-io-2026/), 19 May 2026; [Google, blog post of 13 October 2025](https://blog.google/technology/ai/nano-banana-google-products/).
- **Images**: Opening photo: Dietmar Rabich, [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Buntstifte_--_2021_--_9164.jpg), CC BY-SA 4.0, cropped. Marseille data centres: [Data Center Dynamics](https://datacenterdynamics.com/en/news/interxion-switches-on-river-cooling-at-marseille-data-centers), 30 November 2021, for the cooling.
- **Method**: Emissions are calculated by Projet Celsius from the energy measured in these studies and the average carbon intensity of global electricity; the manufacture of the servers is excluded. Calculated values are rounded to 2 significant figures.
