CO2Stats

The Kettle Question

An AI image

Energy for generating one image with a text-to-image model. Each figure counts something different: open its details to see what.

The figures at a glance

Energy, highest first. On the scale, each line is ten times the one before.

  1. 11.49 Wh per image generatedLeast efficient open image generation model, Power Hungry Processing
  2. a mean of 2.907 Wh per image generatedMean image generation across open text-to-image models, Power Hungry Processing
  3. 1.223 Wh per image generated (1024 × 1024 pixels, 50 diffusion steps)Stable Diffusion 3 Medium image at 50 diffusion steps (MIT Technology Review)
  4. 0.634 Wh per image generated (1024 × 1024 pixels, standard quality)Stable Diffusion 3 Medium 1024 × 1024 image, estimated total energy (MIT Technology Review)
  5. 0.317 Wh per image generated (1024 × 1024 pixels, standard quality)Stable Diffusion 3 Medium 1024 × 1024 image, GPU energy only (MIT Technology Review)

Least efficient open image generation model, Power Hungry Processing

  • 11.49 Wh per image generated (published as: 11.49 kWh for 1,000 inferences; converted from kWh per 1,000 to Wh per one)

Power Hungry Processing: Watts Driving the Cost of AI Deployment?ACM FAccT '24 (arXiv 2311.16863), peer-reviewed paper, 2024-06

Details

whereas the least efficient image generation model uses as much energy as 522 smartphone charges (11.49 kWh), or around half a charge per image generation

What one action is
1,000 image generations, run sequentially without batching, with the least efficient of the eight image generation models tested
Scope
servers
Counts
GPU, CPU and RAM energy measured with Code Carbon
Model or version
not stated (the sentence does not name the model)
Place
AWS us-west-2 (Oregon)
Method
Each model ran 1,000 inferences on each of three datasets, ten times, on a node of 8 NVIDIA A100-SXM4-80GB GPUs, with energy measured by the Code Carbon package.
Claimed by
Company: Alexandra Sasha Luccioni, Yacine Jernite (Hugging Face) and Emma Strubell (Carnegie Mellon University)

History

  • Added.

Mean image generation across open text-to-image models, Power Hungry Processing

  • a mean of 2.907 Wh per image generated (published as: 2.907 kWh per 1,000 queries (Table 2; the text rounds it to 2.9kWh); converted from kWh per 1,000 to Wh per one)

Power Hungry Processing: Watts Driving the Cost of AI Deployment?ACM FAccT '24 (arXiv 2311.16863), peer-reviewed paper, 2024-06

Details

Table 2. Mean and standard deviation of energy per 1,000 queries for the ten tasks examined in our analysis. ... to the most energy-intensive one, image generation, whose mean consumption is 2.9kWh.

What one action is
1,000 image generations, run sequentially without batching, on prompts from each of three datasets (DiffusionDB, ImageReward, SD Prompts)
Scope
servers
Counts
GPU, CPU and RAM energy measured with Code Carbon
Model or version
Eight text-to-image models (stable-diffusion-v1-4, v1-5, 2-1 and xl-base-1.0, openjourney, dreamlike-photoreal-2.0, bk-sdm-tiny, tiny-sd), per the paper's Table 6
Place
AWS us-west-2 (Oregon)
Method
Each model ran 1,000 inferences on each of three datasets, ten times, on a node of 8 NVIDIA A100-SXM4-80GB GPUs, with energy measured by the Code Carbon package.
Claimed by
Company: Alexandra Sasha Luccioni, Yacine Jernite (Hugging Face) and Emma Strubell (Carnegie Mellon University)

The calculator uses this figure: AI image (mean of eight open models, one at a time on GPU servers, 2024 benchmark).

History

  • Added.

Stable Diffusion 3 Medium 1024 × 1024 image, GPU energy only (MIT Technology Review)

  • 0.317 Wh per image generated (1024 × 1024 pixels, standard quality) (published as: about 1,141 joules of GPU energy; converted from J to Wh)

We did the math on AI’s energy footprint. Here’s the story you haven’t heard.MIT Technology Review, press article, 2025-05-20

Details

Generating a standard-quality image (1024 x 1024 pixels) with Stable Diffusion 3 Medium, the leading open-source image generator, with 2 billion parameters, requires about 1,141 joules of GPU energy.

What one action is
One standard-quality 1024 × 1024 image from Stable Diffusion 3 Medium (2 billion parameters)
Scope
servers
Counts
GPU energy
Leaves out
CPUs, fans and other equipment
Model or version
Stable Diffusion 3 Medium
Method
GPU energy measured by Jae-Won Chung (University of Michigan, ML.Energy) for the article.
Claimed by
Independent: MIT Technology Review (James O'Donnell and Casey Crownhart)
Paid for by
Tarbell Center for AI Journalism (grant)

History

  • Added.

Stable Diffusion 3 Medium 1024 × 1024 image, estimated total energy (MIT Technology Review)

  • 0.634 Wh per image generated (1024 × 1024 pixels, standard quality) (published as: an estimated 2,282 joules total; converted from J to Wh)

We did the math on AI’s energy footprint. Here’s the story you haven’t heard.MIT Technology Review, press article, 2025-05-20

Details

With diffusion models, unlike large language models, there are no estimates of how much GPUs are responsible for the total energy required, but experts suggested we stick with the “doubling” approach we’ve used thus far because the differences are likely subtle. That means an estimated 2,282 joules total.

What one action is
One standard-quality 1024 × 1024 image from Stable Diffusion 3 Medium (2 billion parameters)
Scope
data centre
Counts
GPU energy; an estimate for CPUs, fans and other equipment (GPU energy doubled)
Model or version
Stable Diffusion 3 Medium
Method
GPU energy measured by Jae-Won Chung (University of Michigan, ML.Energy), then doubled to estimate the whole operation, following a 2024 Microsoft paper on LLM inference.
Claimed by
Independent: MIT Technology Review (James O'Donnell and Casey Crownhart)
Paid for by
Tarbell Center for AI Journalism (grant)

History

  • Added.

Stable Diffusion 3 Medium image at 50 diffusion steps (MIT Technology Review)

  • 1.223 Wh per image generated (1024 × 1024 pixels, 50 diffusion steps) (published as: about 4,402 joules; converted from J to Wh)

We did the math on AI’s energy footprint. Here’s the story you haven’t heard.MIT Technology Review, press article, 2025-05-20

Details

Improving the image quality by doubling the number diffusion steps to 50 just about doubles the energy required, to about 4,402 joules.

What one action is
One higher-quality 1024 × 1024 image from Stable Diffusion 3 Medium, with the number of diffusion steps doubled to 50
Scope
data centre
Model or version
Stable Diffusion 3 Medium
Method
Measured by Jae-Won Chung (University of Michigan, ML.Energy) for the article; the sentence follows the doubled 2,282 J total but does not say explicitly whether 4,402 J includes the doubling for non-GPU energy.
Claimed by
Independent: MIT Technology Review (James O'Donnell and Casey Crownhart)
Paid for by
Tarbell Center for AI Journalism (grant)

History

  • Added.
Back to top