Generating an image is the most polluting task an AI tool can do. The process can be as polluting as fully charging a smartphone. The total emissions vary between different AI models, and Stable Diffusion XL, in particular, performs poorly in this regard.
Previous studies already proved that training an AI model requires a lot of energy. A new study now looks at how this situation evolves when the model is made available in an AI product. The study shows that training is only a fraction of the problem.
AI models emit the most CO2 when used in a product. Generating an image also generates the highest pollution, according to Code Carbon. Code Carbon is the tool the researchers developed to measure CO2 emissions while running a prompt. All the prompts used in the research were created based on the 10 tasks Hugging Face‘s AI tools had to perform the most. The same prompts passed 88 different models.
Image generating pollutes
Generating 1,000 images is the equivalent of driving a gasoline-powered car for 6.4 kilometres if you deploy Stable Diffusion XL. This model of Stability AI was found to have the highest emissions. The least polluting model allows you to drive only 0.966 meters for this amount of images.
The research should help anyone who wants to work with AI tools to minimize their carbon emissions. This would be possible by always looking for AI tools specifically designed for the situation in which the user wants to use the tool. For example, when writing a summary, a general model will also think about new text and a possible summarization of the text. As a result, the model wastes more energy than necessary. This explanation does, however, not apply to image generation, where models have already been developed for this specific task.
Also read: ChatGPT’s water consumption is astronomical: 40 prompts demand one litre
1 day ago
