ECO-AI – Energy-Conscious Optimization for AI Training

The rapid advancement of large-scale generative AI models, including diffusion-based image generators and multimodal text-to-image systems, has led to unprecedented capabilities but also substantial energy demands and carbon emissions. In this paper, we systematically analyze the energy footprint of training diffusion-based generative models at varying scales (200 million to 3 billion parameters) while evaluating training duration (48, 96, 240 hours), precision modes (FP32 vs. FP16-mixed), optimizer choices (AdamW vs. Lion), and batch sizes across two distinct energy grids–70% renewable and 30% renewable. We also measure model performance metrics such as FID and Inception Score to verify that adopting these energy-saving strategies does not significantly compromise generative quality. Our findings demonstrate that mixed-precision training, strategic optimizer selection, and renewable-aware scheduling can reduce carbon footprints by over 30% with minimal loss in performance. We further discuss inference overheads, highlighting that while training energy can be substantial, inference for high-volume deployments also poses a major sustainability challenge. This work underscores the need for sustainable AI practices across the entire lifecycle of generative models and provides actionable insights for optimizing both energy efficiency and model quality.