Greener Pixels: Reducing Carbon Emissions in AI Image Generation

This repository contains all the scripts, datasets, and analysis for the research project "Greener Pixels", which investigates how to reduce carbon emissions during the inference stage of AI image generation. The study evaluates the environmental impact of popular text-to-image models and explores practical strategies—such as spatial and temporal shifting—to minimize emissions without compromising image quality.

---

Overview

AI image generation models are being used to create over 34 million images per day, resulting in significant carbon emissions. While most prior studies focus on emissions from training, this research focuses on the inference stage, which can contribute up to 90% of total emissions over a model's lifecycle.

We analyze the energy usage of five open-source models across two GPU types (NVIDIA A100 and H100) and propose two key mitigation strategies:

• Spatial Shifting: Relocating computation to low-carbon regions.
• Temporal Shifting: Running jobs during cleaner energy hours.

---

Related Work Summary

• Training vs. Inference: While model training has been widely studied, inference-related emissions are under-examined despite their larger cumulative footprint.
• Shifting Techniques: Existing studies suggest the promise of spatial/temporal shifting but lack real-world quantitative analysis.
• Carbon Intensity Data: Electricity Maps provide region- and time-specific carbon intensity data, enabling precise emission tracking.

---

Methodology

1. Data Collection

   • Carbon intensity data was gathered for 228 global regions.
   • AI model performance was benchmarked using 10 prompts across 5 models and 2 GPU types.

2. Energy Measurement

   • Scripts using pyNVML recorded energy consumption during image generation.
   • Image quality was scored on a 1–5 rubric using standardized prompts.

3. Spatial Shifting Simulation

   • Simulated 34 million image requests per day based on regional interest in generative AI.
   • Requests were routed to either (a) the lowest carbon regions globally or (b) the lowest within the same continent.

4. Temporal Shifting Simulation

   • Carbon intensity patterns were analyzed over 24-hour cycles.
   • Workloads were shifted to midday hours when solar energy is most available.

---

Key Findings

Model & Hardware Impact

• DreamShaper used 95% less energy than the largest model, while maintaining excellent image quality.
• Upgrading from A100 to H100 GPUs reduced energy usage by up to 18%.

Spatial Shifting

• Global spatial shifting reduced emissions by 97%, routing tasks to low-carbon regions like Sweden.
• Regional shifting achieved a 39% reduction while maintaining feasibility under data sovereignty laws.

Temporal Shifting

• In regions like California and Texas, shifting to midday hours cut emissions by up to 77%.
• Effectiveness varied by location; New York showed minimal benefit due to a flat carbon intensity curve.

---

Recommendations

To reduce the carbon footprint of AI image generation:

1. Use smaller, energy-efficient models (e.g., DreamShaper).
2. Deploy on efficient GPUs, like the H100.
3. Place servers in low-carbon regions, such as Sweden or France.
4. Schedule workloads during clean energy hours, especially in solar-rich regions like California.