How much energy does a Stable Diffusion image use?
An image generated with Stable Diffusion XL uses 11.49 Wh. Version 1.5, released a year earlier, uses 1.38, eight times less, for the same kind of image and the same test rig, dated November 28, 2023. It's the only study that compares image generators model by model: no commercial service, not Midjourney, not DALL-E, not Imagen, publishes its own figures.
Figures published on November 28, 2023, measured in a lab on a single dedicated GPU (not a commercial service) · 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 · ADEME : consommation des appareils ménagers
Where do Stable Diffusion XL's 11.49 Wh come from?
From an independent study, not a Stability AI figure. Luccioni, Jernite and Strubell ran a thousand image generations on eight open-weight models, each on the same single GPU, and published the energy bill for each one (“Power Hungry Processing,” submitted November 28, 2023, published at ACM FAccT in June 2024). Stable Diffusion XL (stabilityai/stable-diffusion-xl-base-1.0) is the least efficient of the eight, at 11.49 kWh for a thousand images, or 11.49 Wh per image. Set against the 25 Wh a laptop draws in an hour (ADEME), a single XL image is worth about 28 minutes of screen time.
What about version 1.5?
1.38 kWh for a thousand images, or 1.38 Wh each: eight times less than XL, measured on the same rig. Runwayml/stable-diffusion-v1-5 is still one of the most widely deployed models on the market, despite its age. On the same screen-time math, version 1.5 costs only about 3 minutes of laptop use. The other six models in the panel range from 0.76 Wh (segmind/tiny-sd, a lightweight build) to 3.39 Wh (dreamlike-art/dreamlike-photoreal-2.0): model size matters more than the simple fact of generating an image instead of text.
Why isn't there a page for Midjourney or DALL-E?
Because neither has published a figure. This measurement only exists for open-weight models, ones you can download and time on your own hardware. Midjourney and DALL-E are closed services: nobody outside OpenAI or Midjourney Inc. can measure what they actually use. It's the same gap Google is the only one to have closed for text, with Gemini: a provider publishing its own fleet-measured figures is still the exception, not the rule.
Does this figure reflect a real commercial service?
No, and probably not in the direction you'd expect. The measurement runs on a single dedicated GPU, with none of the idle servers, data-center losses, or request queue of a real production service. Those costs only add up, never subtract: a commercial model's production footprint is likely higher than its lab measurement, not lower. That's why this site uses a higher figure as its working assumption for a generic AI-generated image when no model is named: a figure placed near the top of the range measured here, not in the middle.
A Stable Diffusion prompt in everyday actions
The figures published for Stable Diffusion, converted with the site's reference actions (every action value is sourced, see the Sources page):
- 1 hour of LED light = 0.87 Stable Diffusion prompts
- 1 hour of TV = 6.1 Stable Diffusion prompts
A model's number doesn't say which one to pick for a given job, or whether generating the image is worth it at all. That's a different craft: the GHIS studio helps companies scope their AI use.