A Single AI Image Consumes as Much as Charging Your Phone 4 Times

Up to 10 Wh for a single image. That’s what generating an image with GPT-4o in highest quality mode can consume — the same energy as charging your smartphone 4 times. And most people do it without knowing, several times a day.

While a text query to an AI model consumes around 0.3 Wh as a reference, generating an image costs between 3 and 33 times more. And the difference between choosing one model or another can be 46 times.

The consumption map: model by model

Not all image generators consume the same. Bertazzini et al. measured 17 diffusion models on standardized hardware and found brutal differences. Cross-referencing their data with the most reliable estimates available in 2026, here’s the landscape:

• FLUX.2 klein (Black Forest Labs, compact variant): 0.15–0.8 Wh — the most efficient available, designed to run on consumer hardware
• Imagen 3 (Google, on TPU v6): 0.5–2.5 Wh — probably the most efficient among commercial services
• SDXL (Stability, open benchmark on H100): 1.64 Wh — the best actually measured reference point
• Midjourney v7 (draft mode): 1–2 Wh — fast mode saves significantly
• Midjourney v8 Alpha: 2–4 Wh per grid — the new version prioritizes extreme realism over efficiency
• Adobe Firefly Image 4: 0.8–4 Wh — Adobe has generated over 24 billion assets without publishing a single consumption figure
• FLUX.2 dev/full (32B parameters): 1.5–6 Wh — large model, significantly heavier than its klein version
• GPT-4o native image: 0.7–10 Wh — the widest range, depending on selected quality and resolution

1.64 Wh. That is the most solid reference number that exists for generative image: SDXL measured on H100 by the AI Energy Score. Everything else is estimates.

What a single image really costs

To put it in context, a single AI image is equivalent to:

• FLUX.2 klein (0.5 Wh): having an LED bulb on for 3 minutes
• SDXL (1.64 Wh): an LED bulb on for 10 minutes
• Midjourney v8 (3 Wh): an LED bulb on for 18 minutes
• GPT-4o high quality (10 Wh): charging a full smartphone (14 Wh ≈ 1 charge)

It seems like little. But multiply by the number of times you regenerate until you’re happy with the result. If you need 10 iterations to reach the final image with Midjourney v8, you’ve consumed 30 Wh — the equivalent of two smartphone charges.

The real cost of an AI image isn’t generating it once. It’s generating it ten times until you like it.

Three decisions that change consumption

1. Choose the model proportionate to the task

If you need an image for a draft, an internal presentation, or a prototype, a compact model like FLUX.2 klein or Midjourney’s draft mode consumes 10–20 times less than GPT-4o at maximum quality. Save the heavy models for the final output.

2. Reduce regenerations

Every “try” generates a complete image from scratch. Refine the prompt before generating. Use fixed seeds to iterate on variations. A well-defined prompt can save you 5–8 regenerations — and multiply your efficiency by the same factor.

3. Lower the resolution when it doesn’t matter

Resolution scales consumption non-linearly. Generating at 512x512 consumes significantly less than at 1024x1024. If the image is going to a thumbnail, a social post, or a wireframe, maximum resolution is pure energy waste.

What can I do?

• If you’re a regular AI image user: Use our footprint calculator to estimate your monthly consumption. And remember: draft mode or compact models cover 80% of use cases at a fraction of the cost.

• If you lead a creative team: Establish a usage policy: lightweight model for iteration, heavy model only for the final deliverable. This can reduce your team’s consumption by 70–80% without affecting output quality.

• If you’re a developer: Integrate efficient models by default in your pipelines. FLUX.2 klein for previews, larger models only when the user explicitly requests high quality. The user rarely needs 1024x1024 for a first look.