Truace
Subscribe
52/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Energy per AI image varies up to 46x across 17 models, a lab study finds

Study title: The hidden cost of an image: Quantifying the energy consumption of Artificial Intelligence image generation

Evidence type
Benchmark or lab evaluation
Sources
1 sourcePeer-reviewed
Published
2 min read
Truvace record
TRV-2026-2533cite or check it

Study title: The hidden cost of an image: Quantifying the energy consumption of Artificial Intelligence image generation

Researchers at the University of Florence measured the energy used to generate images with 17 open diffusion models, running more than 9,000 measurements. Energy per image differed by up to 46x between the most and least power-hungry models, with Lumina highest and LCM_SSD_1B lowest.

Read the study abstract

With the growing adoption of Artificial Intelligence (AI) image generation, in conjunction with the ever-increasing environmental resources demanded by AI, we are urged to answer a fundamental question: What is the environmental impact hidden behind each image we generate? In this research, we present a comprehensive empirical experiment designed to assess the energy consumption of AI image generation through diffusion models. Our experiment compares 17 state-of-the-art image generative models by considering mul…

The problem

In a test of 17 diffusion image models, energy per image differed by up to 46x between models, and int8 quantization raised energy use for most models, by up to 64.54%.

In brief

Resolution and compression did not behave as expected. Going from 512 × 512 to 1024 × 1024 pixels raised energy use by 1.3x to 4.8x depending on the model, and int8 quantization, a technique meant to cut memory use, raised energy use for most models, by up to 64.54%. Prompt length and content made no statistically significant difference.

The team also scored image quality with FID, precision and recall. Energy and quality did not always trade off: Flash_SD3, Hyper_SD and SD_3 combined high quality with low energy use, while Lumina ranked lowest on both. The work covers only models free for academic use, so commercial services such as DALL-E and Midjourney were not tested.

Main points

  1. The study compared 17 diffusion models, 10 U-Net-based and 7 Transformer-based, all from Hugging Face, in more than 9,000 measurements.
  2. Median energy per image differed by up to 46x between Lumina (highest) and LCM_SSD_1B (lowest); U-Net-based models generally used less than Transformer-based ones.
  3. int8 quantization raised energy use for most models, by up to 64.54%; only SDXL_Turbo, Hyper_SD, Flash_SD3 and Flux_1 saw modest savings of up to 12.13%.
  4. Quadrupling pixel count (512 × 512 to 1024 × 1024) raised energy by about 1.3x for SDXL_Turbo and Hyper_SD, 3.85x for SD_1.5 and 4.75x for Lumina, while PixArt models stayed near 1.0x.
  5. PixArt models used more energy for non-square formats such as 768 × 1024; the authors guess this reflects square-only training data.
  6. Prompt length (short, medium, long) had no statistically significant effect on energy use for any model, per Kruskal-Wallis tests.
  7. Flash_SD3 reached the lowest FID of all models while keeping energy use low; Lumina ranked worst on both quality and energy.

The full story

Every model ran on one workstation with an AMD Ryzen 9 7950X, 128 GB of RAM and an Nvidia GeForce RTX 4090. Energy was logged with CodeCarbon as the sum of CPU, GPU and RAM use. Models ran with default settings, at two quantization settings, three resolutions and three prompt lengths, using 30 prompts drawn from CIFAR-100 class names.

Short prompts were one or two words, long prompts averaged 33 words (expanded with the Phi-3-Mini model) and medium prompts averaged 14 words. A preliminary test of the 100 CIFAR-100 classes, each run 10 times, found no link between prompt content and energy use.

Image quality was scored against 1,100 ImageNet-1k validation images, with 396 generated images per model. Lumina's poor FID came largely from weak 512 × 512 output, since it is tuned for 1024 × 1024 and above.

The authors say they hope the results help reduce the environmental cost of each generated image. The paper is an author version of an arXiv paper, published in Engineering Applications of Artificial Intelligence.

Sources

1 on file
  1. Engineering Applications of Artificial Intelligencedoi:10.1016/j.engappai.2026.116340
    Peer-reviewed2026-10-06

Talk it over

Four questions to start onWho carries this cost, and what would have to change to prevent it?

Start with a question

Energy per AI image varies up to 46x across 17 models, a lab study findshttps://truvace.com/item/the-hidden-cost-of-an-image-quantifying-the-energy-consumption-of-artificial-int

  1. Who carries this cost, and what would have to change to prevent it?

    The problem In a test of 17 diffusion image models, energy per image differed by up to 46x between models, and int8 quantization raised energy use for most models, by up to 64.54%.

    Answer in the thread

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on a benchmark or lab test, from Engineering Applications of Artificial Intelligence.

    Answer in the thread

  3. What does it leave open? What else would you want to know before acting on it?

    The record’s own caveat Only models free for academic use and runnable offline were tested, so subscription services such as DALL-E and Midjourney were excluded, and all runs used one workstation with an RTX 4090.

    Answer in the thread

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

    Answer in the thread

The questions are the same on every record. The details under them are taken from this record as written.

What readers say

More from Climate