Two claims about AI image energy circulate, and both are lazy. One says a single image burns as much power as a small town. The other says local generation is free because the GPU is already on your desk. Neither comes with a measurement.
There is a measurement. It is peer reviewed, it is specific, and it gives you a way to work out your own number instead of borrowing someone else's adjective.
What Was Actually Measured
Alexandra Sasha Luccioni, Yacine Jernite and Emma Strubell measured the energy of AI inference across tasks in Power Hungry Processing: Watts Driving the Cost of AI Deployment?, published at ACM FAccT 2024. Their headline finding is that multi-purpose generative architectures are orders of magnitude more expensive per inference than task-specific systems.
For image generation specifically, the full paper reports:
- A mean of 2.907 kWh per 1,000 images across the image generation models tested, with a standard deviation of 3.31.
- The least efficient image model used 11.49 kWh per 1,000 images, which the authors equate to 522 smartphone charges, roughly half a charge per image.
- All runs used a node of eight NVIDIA A100-SXM4-80GB GPUs on AWS. Each run used one GPU, but the idle power of the other seven is included in the totals.
That last line matters. The figures include the overhead of a shared server, which is part of why a cloud image costs more energy than the GPU math alone suggests.
Your Desk Is a Different Machine
A local render is a different setup: one GPU, one job, no idle neighbours. You can estimate it from two numbers you can actually find.
The ceiling is the card's rated power. NVIDIA lists the RTX 4090 at 450 W total graphics power. That is a maximum, not a measured draw, so treat it as the pessimistic case. The price comes from the U.S. Energy Information Administration, which puts the average residential rate at 18.31 cents per kWh for July 2026.
Energy is power multiplied by time. Here is the GPU at its full 450 W rating, for a thousand images, at three render times:
| Seconds per image | kWh per 1,000 images (450 W) | Cost at 18.31 cents per kWh |
|---|---|---|
| 5 | 0.63 | about $0.11 |
| 10 | 1.25 | about $0.23 |
| 30 | 3.75 | about $0.69 |
These are worst-case figures for the card alone. The rest of the computer draws power too, so the wall number will be higher than the GPU number. The real draw during a render is usually below the rated ceiling. The two errors pull in opposite directions, which is exactly why you should measure rather than trust either.
What the Numbers Say, and What They Do Not
Three things are true at once.
- Local generation is not free. At ten seconds per image, a thousand images is a little over a kilowatt-hour on the card. It is a real cost, and anyone claiming zero is not counting.
- It is small next to a subscription. At those rates the electricity for a thousand images costs well under a dollar. A Midjourney Standard plan is $30 a month before you generate anything.
- The biggest lever is not where you run. It is how much you run. Every step, every discarded variation and every re-roll is energy. The cheapest image is the one you did not have to generate twice.
What a Heavy Year Looks Like
Scale makes the comparison clearer. Take a heavy user who keeps 200 images a day, every day, for a year: 73,000 images. Using the same worst-case card-only assumptions as the table above, 450 W and ten seconds per image, that year is about 91 kWh at the GPU, or roughly $17 at 18.31 cents per kWh. The same number of images at the paper's 2.907 kWh per thousand would be about 212 kWh.
Those two figures are not a like-for-like race. One is a ceiling estimate for a single consumer card, and the other is a measured mean across several models on shared data-centre hardware, idle GPUs included. What they do show is the order of magnitude. A year of heavy local generation is tens of dollars of electricity, not hundreds, and the same year on a Standard subscription is $360 before any per-image cost exists.
Your render time is the variable that moves this most. Double the seconds per image and you double the energy, which is why the sections below spend more time on wasted renders than on hardware.
Waste Is the Real Bill
That last point is where workflow beats hardware. Energy per image falls when fewer images are thrown away, and most are thrown away because the first try did not land: the composition was wrong, the steps were set too high, or a change could not be traced back to its cause.
Some of the Vision Studio design is aimed directly at that. Every model starts from the same 25 steps at CFG 7.5 rather than an inflated default. ControlNet and reference layers fix composition before the render instead of after twenty retries. The iteration tree records the settings difference between every render and its parent, so a good result can be branched from rather than rediscovered. Live step previews let you stop a render that is going wrong before it finishes.
Measure It Yourself
- Buy a plug-in power meter. They are inexpensive and they measure the whole machine at the wall, which is what you pay for.
- Read idle, then read during a batch. The difference is the cost of generating.
- Time a batch of fifty. Divide the energy by fifty and you have your own per-image number to replace every estimate in this post.
- Multiply by your tariff, not the national average. Rates vary widely by state and by time of day.
Numbers Over Adjectives
AI energy deserves scrutiny, and the research community has given us the tools to apply it honestly. The cloud figure is measured and published. The local figure is yours to measure, in an afternoon, with a meter. Whichever way your number comes out, it will be more useful than either the scare headline or the free-lunch slogan.
Vision Studio is free, open-source (MIT), and runs locally on your GPU. Download it at vision-studio-x.com/download or read the source at github.com/Git-Rocky-Stack/vision-studio.
-Rocky
#AIEnergy #Sustainability #LocalAI #GPU #CreatorEconomy #OpenSource #VisionStudio