The Carbon Intensity of Generative Design: Emissions Analysis of Training and Sampling From Generative Models

Abstract This research examines the environmental impact of artificial intelligence (AI) in manufacturing, focusing on the energy consumption and emissions from AI technologies. While computer vision can enhance production efficiency and minimize waste by analyzing data, its deployment is not without environmental costs, primarily due to the energy-intensive processes of training and operating AI systems. Existing methodologies for evaluating AI’s carbon footprint predominantly cover emissions during model training, neglecting the significant emissions from sampling — generating data from trained models. Our study bridges this gap by developing a carbon accounting method for training and sampling emissions, utilizing the carbon emissions of diffusion models as a surrogate for those of computer vision in the manufacturing sector. We introduce novel contributions by characterizing the emissions of generative models in terms of fixed and variable costs, akin to manufacturing terminologies. Our findings reveal that image-based diffusion models produce 772x more carbon emissions per sample compared to tabular models with a similar number of parameters. Moreover, for small models, emissions are dominated by sampling, highlighting the necessity of quantifying sampling emissions for sustainable applications of AI. This comprehensive approach to quantifying AI’s environmental footprint is crucial for ensuring that the benefits of AI in manufacturing substantially outweigh its energy and resource costs.