Let's talk about something we don't often discuss in the AI art community: this technology has an environmental footprint.
Every image you generate requires compute. That compute requires energy. That energy, depending on where it's generated, might come from fossil fuels. It's worth understanding the scale.
The Numbers
A single AI image generation uses somewhere between 0.1 and 1 kWh of electricity — roughly equivalent to:
- Charging a smartphone 20-200 times
- Running a refrigerator for a few hours
- Driving about 0.5-5 miles in an electric car
Multiply this by millions of images generated daily, and the numbers add up.
Large AI companies estimate their AI services produce hundreds of thousands of tons of CO2 annually. That's comparable to small countries.
Why It Uses So Much Energy
AI image generation uses "diffusion models," which work through iterative steps. A single image might go through 20-50 refinement steps. Each step runs the entire neural network.
Compare this to older approaches likeGANs, which might produce an image in one pass. Diffusion produces better results — but at greater computational cost.
The industry is aware of this. There's active research into more efficient models, "knowledge distillation" (smaller models that approximate larger ones), and better hardware.
What Makes It Better or Worse
Efficiency varies enormously. A well-optimized model on efficient hardware uses a fraction of the energy of an unoptimized one running on old hardware.
Batch processing is better. Generating 10 images in one batch uses less energy than generating 10 one at a time, thanks to computational sharing.
Server location matters. Servers powered by renewable energy (hydro, solar, wind) have a fraction of the carbon footprint of those powered by coal.
User behavior counts too. Generating 100 images to find one good one uses 100x the energy of being more mindful with your prompts.
What ArtFelt Does
We think about this:
Efficient infrastructure. We use optimized models and efficient hardware where possible.
Donated compute. Many of our GPU resources come from community donations. This is inherently more sustainable than massive dedicated server farms.
We're transparent. We could hide this. We choose not to.
We're improving. The technology gets more efficient over time. What required massive compute last year might be achievable on a laptop next year.
What You Can Do
Practical steps:
- Be intentional. Don't generate hundreds of images on autopilot. Think about what you want, craft your prompt, generate fewer, iterate thoughtfully.
- Use the results. An image you never use represents wasted compute. Generate with purpose.
- Support sustainable platforms. Ask questions. Know where your images come from.
- Offset if you want. Some platforms offer carbon offsets. It's not a perfect solution, but it's something.
The Bigger Picture
Technology has always had environmental costs. The internet, streaming video, cryptocurrency — all consume significant energy. The question isn't whether to use technology, but how to make it sustainable.
AI image generation will likely become dramatically more efficient. The energy per image has already dropped 10x in the past few years. That trend will continue.
In the meantime, being mindful matters. Small choices add up when millions of people make them.
Create with intention at ArtFelt. Good art comes from thoughtful creators.