Modest, artistic, and radical solutions to the environmental impact of image-generating machine learning

Technical Summary: Modest, Artistic, and Radical Solutions to the Environmental Impact of Image-Generating Machine Learning

Problem Statement
The paper argues that the environmental impact of machine learning (ML), specifically image generation, is vastly underestimated by the prevailing narrative of efficiency. While ML is often touted for optimizing Information and Communication Technology (ICT), this "small gain" is overwhelmed by the massive carbon, water, and land footprints of data centers and ML-ready devices. The authors highlight a "rebound effect" where increased efficiency leads to expanded usage and infrastructure, rather than conservation. They cite projections that AI data centers could consume 35 gigawatts annually by 2030 and note that a single image generation produces approximately 300 grams of CO2e—750 times that of text classification. The current trajectory, driven by shareholder capitalism and the demand for speed and accuracy, creates a vicious cycle where infrastructure expansion justifies further demand, threatening to exceed planetary boundaries.

Methodology
The authors employ a multidisciplinary approach combining computer engineering, media scholarship, and critical art practice. Their methodology involves:

1. Literature Survey: Analyzing electricity consumption in ML training and inference, focusing on the proliferation of GPUs, the shift toward edge computing, and the disproportionate energy cost of inference compared to training.
2. Critical Analysis of Efficiency: Deconstructing the corporate definition of "efficiency" (speed and accuracy) and contrasting it with "digital sufficiency" and "degrowth" frameworks.
3. Conceptual Art Project (Work in Progress): The team is developing a "tiny image generator" as a conceptual artwork. This project serves as a technical testbed for ethical and low-energy ML. Key technical strategies for this prototype include:
   • Hardware Constraints: Training and running models on limited-capacity hardware (e.g., Raspberry Pi) rather than massive server farms.
   • Data Ethics: Rejecting scraped datasets (like LAION-5B) in favor of ethically sourced, non-scraped data, including contributions from participants and specific archives (e.g., textiles and Islamic calligraphy from the Victoria & Albert Museum).
   • Model Architecture: Utilizing "Tiny Machine Learning" (TinyML) techniques, including:
     • Quantization: Reducing precision (e.g., 8-bit or 4-bit) to shrink model size.
     • Few/One-Shot Learning: Training on minimal examples to reduce data requirements.
     • Noise Injection: Using noise rather than statistical averages to generate novel, non-clichéd outputs, thereby avoiding the "mean images" typical of large models.
     • Dataset Weighting: Explicitly assigning higher importance to specific ethical datasets (e.g., Islamic art) to counteract pre-trained biases toward photorealism and Western art.
   • Political Integration: Mandating the inclusion of specific political references (e.g., Palestine) in outputs to counteract hegemonic AI narratives.

Key Contributions

• Reframing Efficiency: The paper challenges the industry standard that prioritizes speed and accuracy, arguing that these metrics drive unsustainable energy consumption. It proposes "digital sufficiency" and "true-cost accounting" (including environmental externalities) as necessary alternatives.
• Technical Feasibility of "Tiny" AI: The authors demonstrate that image generation can be adapted to run on limited hardware using small language models (SLMs), low-precision architectures, and non-neural or sparse approaches, provided the goal shifts from general intelligence to specific, modest tasks.
• Ethical Data Practices: The paper proposes a workflow where metadata and participant contributions are treated as integral parts of the artwork, contrasting with the opaque data scraping of commercial models.
• Critical AI Art: By integrating political mandates (such as referencing Palestine) and aesthetic constraints (austerity, novelty over verisimilitude), the project offers a model for "decolonial" AI that resists the dehumanizing tendencies of mainstream generative models.

Results
As the paper describes a "work in progress," it does not present final quantitative experimental results (e.g., specific watt-hour savings or final model accuracy metrics). However, the authors report on the theoretical and design-phase outcomes:

• They successfully identified that current industry goals (e.g., 100ms latency) are driven by corporate addiction models rather than necessity, and that relaxing these constraints allows for significantly smaller models.
• They determined that using a "small" variant of StableDiffusion (SD 2.1 unclip-small) still carries a significant carbon footprint due to the parent model's training; thus, their proposed approach requires starting from scratch or fine-tuning with extreme constraints to be truly "tiny."
• They established that dataset weighting is a viable technical method to override pre-trained biases and prioritize specific cultural aesthetics without requiring massive compute resources.

Significance and Claims
The paper claims that the only sustainable path forward for ICT is "degrowth" and the "undesign" of resource-intensive technologies. It asserts that:

• Efficiency is not Conservation: In a shareholder-capitalist framework, efficiency merely enables more consumption. True environmental mitigation requires "graceful degradation" and "collapse informatics."
• Modesty as a Radical Act: Creating modest, artistically sophisticated, and politically engaged AI models is a radical alternative to the "accelerationist" goals of large corporations.
• Policy Implications: The authors suggest that if governments priced electricity according to its environmental cost, corporations would be forced to adopt genuine efficiency. Until then, they advocate for voluntary simplicity and the creation of "convivial tools" that operate within strict energy and data limits.
• Artistic Intervention: The proposed tiny image generator serves not just as a tool, but as a critique of the "cornucopian paradigm," demonstrating that useful, creative AI can exist without the massive environmental footprint of current Large Language Models (LLMs) and generative imaging systems.