Carbon in Motion: Characterizing Open-Sora on the Sustainability of Generative AI for Video Generation

Baolin Li (Northeastern University); Yankai Jiang (Northeastern University); Devesh Tiwari (Northeastern University)

Abstract

The rapid rise of generative AI (GenAI) technologies has brought innovative video generation models like OpenAI’s Sora to the forefront, but these advancements come with significant sustainability challenges due to their high carbon footprint. This paper presents a carbon-centric case study on video generation, providing the first systematic investigation into the environmental impact of this technology. By analyzing Open-Sora, an open-source text-to-video model inspired by OpenAI Sora, we identify the iterative diffusion denoising process as the primary source of carbon emissions. Our findings reveal that video generation applications are significantly more carbon-demanding than text-based GenAI models and that their carbon footprint is largely dictated by denoising step number, video resolution, and duration. To promote sustainability, we propose integrating carbon-aware credit systems and encouraging offline generation during high carbon intensity periods, offering a foundation for environmentally friendly practices in GenAI.