smith.nancy•10d agoProlific Poster

MIT study made me rethink my whole stance on AI art energy use

I was reading through some research papers from MIT last week and found out that training a single AI model like Stable Diffusion uses about the same energy as 10 cars over their whole lifetime. That number blew my mind because I always thought the energy complaint was overblown by critics who just don't like the tech. But here's the kicker - the paper also said that using the finished model to generate one image uses way less energy than running a digital camera for the same amount of time. So the training is terrible but the actual use is pretty green compared to traditional photography. I still think AI art has creative value but this made me realize the environmental cost is real and complicated. Has anyone else looked into the energy stats on specific models?

3 comments

3 Comments

pat_stone9d ago

Yeah I remember reading that same MIT study and it really stuck with me how the training cost is so heavy but the actual use is pretty light. I saw another comparison that said training one big model is like driving a gas car across the country 150 times, which put it in perspective for me. I think being able to check model cards on Hugging Face like Aaron mentioned is the way to go, at least we can pick the less wasteful options when we have the choice.

aaronsullivan10d ago

Wait have you looked into the carbon footprint difference between the big models? I did a deep dive on this a few months ago when I was feeling guilty about using Midjourney all the time. What helped me was finding a comparison that showed Stable Diffusion's training emissions were roughly equal to 400 flights from NY to London. That sounds awful until you realize Google's search alone pumps out way more daily. I started checking model cards on Hugging Face before trying new tools and mostly stuck with models that publish their energy data. It's not perfect but at least I can make informed choices now instead of just guessing.

-1

emmam897d ago

Wait has anyone actually run one of these models on renewable energy? My friend Sarah works at a smaller AI lab and she told me they timed their training runs to line up with when the local grid had the most solar power available. She said it cut their carbon impact by like a third or something just by being smart about when they hit the go button. It makes you wonder how many companies are doing that vs just running whatever whenever. I wish there was like a badge or something for models that do that kind of thing on the model cards.