Latent Pollution Model: The Hidden Carbon Footprint in 3D Image Synthesis

TL;DR

This work quantifies the carbon footprint of unconditional 2D and 3D latent diffusion models used in medical imaging, focusing on training and data generation. Using Carbontracker, region-specific carbon intensity, and Power Usage Effectiveness adjustments, the authors vary model size, image dimensions, distributed training, and sampling steps to compute $CO_2e$ for $150{,}000$ training iterations and $10{,}000$ synthesized samples. They find that synthesis dominates emissions, with 3D high-resolution synthesis equating to substantial car travel distances, and that regional and seasonal factors can dramatically shift totals. The study argues for environmentally sustainable practices in AI research, including reducing inference steps, awareness of energy timing, and broader adoption of carbon-emission tracking in medical imaging workflows.

Abstract

Contemporary developments in generative AI are rapidly transforming the field of medical AI. These developments have been predominantly driven by the availability of large datasets and high computing power, which have facilitated a significant increase in model capacity. Despite their considerable potential, these models demand substantially high power, leading to high carbon dioxide (CO2) emissions. Given the harm such models are causing to the environment, there has been little focus on the carbon footprints of such models. This study analyzes carbon emissions from 2D and 3D latent diffusion models (LDMs) during training and data generation phases, revealing a surprising finding: the synthesis of large images contributes most significantly to these emissions. We assess different scenarios including model sizes, image dimensions, distributed training, and data generation steps. Our findings reveal substantial carbon emissions from these models, with training 2D and 3D models comparable to driving a car for 10 km and 90 km, respectively. The process of data generation is even more significant, with CO2 emissions equivalent to driving 160 km for 2D models and driving for up to 3345 km for 3D synthesis. Additionally, we found that the location of the experiment can increase carbon emissions by up to 94 times, and even the time of year can influence emissions by up to 50%. These figures are alarming, considering they represent only a single training and data generation phase for each model. Our results emphasize the urgent need for developing environmentally sustainable strategies in generative AI.

Figures (4)

• Figure 1: $CO_2$ emissions from training and inference of 2D and 3D LDMs. The highest benchmark value represents the annual per capita carbon budget needed for a 50% chance of limiting global warming to 1.5°C, as per the Paris Agreement unfcccparis2015. For 3D synthesis, emissions reach 33.7% of this budget, equivalent to a 1500km flight, such as from Tokyo to Seoul.
• Figure 2: $CO_2$ emissions by model architecture size in Training and Synthesis.
• Figure 3: Regional and temporal differences in Carbon Emissions in a 3D LDM.
• Figure 4: $CO_2$ emissions of different speed-up techniques in 3D LDM.