pith. sign in

arxiv: 2606.19957 · v1 · pith:DK22RA3Fnew · submitted 2026-06-18 · 💻 cs.CY

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

Laura U. Marks , Jess MacCormack , Kehui Li This is my paper

Pith reviewed 2026-06-26 15:33 UTC · model grok-4.3

classification 💻 cs.CY
keywords machine learningenvironmental impactimage generationenergy consumptiontrue-cost accountingsustainable computingtiny modelsICT efficiency
0
0 comments X

The pith

Image-generating machine learning's large environmental footprints can be reduced by tiny models, low-precision hardware, and true-cost accounting that questions efficiency driven by shareholder interests.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper surveys electricity consumption during ML training and inference, especially for image generation, and shows how small efficiency gains are outweighed by the carbon, water, and land demands of data centers and devices. It explores technical approaches including inexact computing, tiny models, low-precision hardware architectures, and hardware with limited capacity, plus energy anticipation during design. The authors describe ongoing work on an ethical tiny image generator trained on non-scraped data. They extend the discussion to economic context by calling for true-cost accounting of ML impacts and noting that efficiency criteria reflect a shareholder-capitalist framing of ICT.

Core claim

Machine learning is often touted to improve the efficiency of ICT, but that small gain is overwhelmed by the enormous carbon, water, and land footprints of data centers and ML-ready devices. We survey the electricity consumption of ML applications in training and inference, focusing on electricity-intensive image generation. Our team of a computer engineer, a media scholar, and an artist explore solutions including inexact computing; tiny language models; low-precision hardware architectures; hardware with limited capacity; and anticipating and mitigating energy demands at the design phase. We will sketch our work in progress of an ethical and aesthetically sophisticated tiny image generator

What carries the argument

True-cost accounting for the environmental impact of machine learning, which makes visible costs externalized under current efficiency metrics.

If this is right

  • Tiny models and limited-capacity hardware would lower electricity use in both training and inference phases of image generation.
  • Anticipating energy demands during the design phase would reduce the overall carbon, water, and land footprints of ML systems.
  • An ethical tiny image generator using non-scraped data would offer a lower-impact alternative while maintaining artistic and functional value.
  • True-cost accounting would shift evaluation criteria away from efficiency defined solely by shareholder interests.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar modest approaches could extend to other ML domains such as language or video generation where energy demands are also high.
  • Widespread adoption would likely require changes in data sourcing norms beyond what individual projects can achieve.
  • The artistic framing might encourage community or small-scale deployment models that bypass large data-center infrastructure.
  • Regulatory or incentive structures would need to incorporate true-cost metrics to make the technical solutions competitive.

Load-bearing premise

The proposed technical solutions such as tiny models and low-precision hardware can be implemented at scale while still delivering useful image-generation functionality.

What would settle it

A controlled test measuring image quality, accuracy, and user utility of a tiny non-scraped-data generator against a standard model at matched energy budgets.

read the original abstract

Machine learning is often touted to improve the efficiency of ICT, but that small gain is overwhelmed by the enormous carbon, water, and land footprints of data centers and ML-ready devices. We survey the electricity consumption of ML applications in training and inference, focusing on electricity-intensive image generation. Our team of a computer engineer, a media scholar, and an artist explore solutions including inexact computing; tiny language models; low-precision hardware architectures; hardware with limited capacity; and anticipating and mitigating energy demands at the design phase. We will sketch our work in progress of an ethical and aesthetically sophisticated tiny image generator using non-scraped data. Looking to the economic context, we will propose a true-cost accounting for the environmental impact of machine learning and suggest that the criterion of efficiency is driven by the shareholder-capitalist framing of ICT.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The manuscript surveys electricity consumption of ML applications with a focus on image generation, explores technical solutions including inexact computing, tiny language models, low-precision hardware architectures, hardware with limited capacity, and design-phase energy mitigation, sketches ongoing work on an ethical tiny image generator trained on non-scraped data, and proposes a true-cost accounting framework while arguing that efficiency criteria in ICT are shaped by shareholder-capitalist framing.

Significance. If the sketched solutions can be realized at scale, the work offers an interdisciplinary contribution that combines engineering constraints with media and artistic considerations to address the environmental costs of generative models. The explicit linkage of technical choices to economic framing provides a distinctive angle not commonly found in purely technical surveys of AI sustainability.

major comments (1)
  1. [Abstract and §1] Abstract and §1: the claim to survey electricity consumption of ML training and inference is not supported by any quantitative data, measurements, citations to specific studies, or error analysis in the manuscript; this is load-bearing because the scale of the problem is used to motivate all subsequent technical and economic proposals.
minor comments (1)
  1. The repeated use of forward-looking phrasing ('we will propose', 'we will sketch') should be accompanied by a clear statement of what portions of the work are currently complete versus planned, to help readers assess the status of the tiny image generator and true-cost accounting.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback. We will address the concern about the lack of quantitative support for the survey claims by incorporating appropriate literature references and data in the revised manuscript.

read point-by-point responses

  1. Referee: [Abstract and §1] Abstract and §1: the claim to survey electricity consumption of ML training and inference is not supported by any quantitative data, measurements, citations to specific studies, or error analysis in the manuscript; this is load-bearing because the scale of the problem is used to motivate all subsequent technical and economic proposals.

    Authors: We agree that the current version does not provide quantitative data, measurements, or specific citations to support the survey claim in the abstract and §1. This is a substantive gap. In revision we will add citations to established studies on ML energy use (including training and inference for image-generation models), include representative quantitative figures drawn from the literature, and briefly note methodological considerations such as variability in reported estimates. These additions will be placed in a new or expanded subsection of §1 to better motivate the technical and economic proposals that follow. revision: yes

Circularity Check

0 steps flagged

No derivations, equations, or fitted predictions present

full rationale

The paper is an exploratory survey and position piece focused on environmental impacts of ML image generation, proposing directions such as tiny models and true-cost accounting. It contains no mathematical derivations, equations, parameter fits, or quantitative predictions that could reduce to inputs by construction. All claims are forward-looking proposals without self-referential fitting or self-citation chains that bear the central argument. The absence of any load-bearing technical steps makes circularity analysis inapplicable; the work is self-contained as a qualitative discussion.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

No mathematical content or derivations; the paper rests on the domain assumption that data-center footprints dominate ML energy use and that efficiency metrics are shaped by capitalist priorities, but these are stated rather than derived.

axioms (1)
  • domain assumption Data centers and ML devices have enormous carbon, water, and land footprints that overwhelm efficiency gains.
    Stated in the opening of the abstract as the premise for the survey.

pith-pipeline@v0.9.1-grok · 5673 in / 1081 out tokens · 15725 ms · 2026-06-26T15:33:57.809351+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 10 canonical work pages · 2 internal anchors

  1. [1]

    Anders S. G. Andrae. 2023. From an Environmental Viewpoint Large ICT Net- works Infrastructure Equipment must not be Reused.WSEAS Transactions on Environment and Development19 (April 2023), 375–382. doi:10.37394/232015. 2023.19.34

    work page doi:10.37394/232015 2023

  2. [2]

    Hrishav Bakul Barua and Kartick Chandra Mondal. 2019. Approximate Com- puting: A Survey of Recent Trends—Bringing Greenness to Computing and Communication.J. Inst. Eng. India100, 6 (Dec. 2019), 619–626

    2019

  3. [3]

    2020.Cisco Annual Internet Report (2018–2023)

    Cisco Systems. 2020.Cisco Annual Internet Report (2018–2023). White paper

    2020

  4. [4]

    Undeclared Preferences

    Washington Cunha, Vıtor Mangaravite, Christian Gomes, Sérgio Canuto, Elaine Resende, Cecilia Nascimento, Felipe Viegas, Celso França, Wellington Santos Martins, Jussara M. Almeida, Thierson Rosa, Leonardo Rocha, and Marcos André Gonçalves. 2021. On the cost-effectiveness of neural and non-neural approaches and representations for text classification: A co...

    work page doi:10.1016/j.ipm 2021

  5. [5]

    Zhaoxia Deng, Jongsoo Park, Ping Tak Peter Tang, Haixin Liu, Jie Yang, Hector Yuen, Jianyu Huang, Daya Khudia, Xiaohan Wei, Ellie Wen, Dhruv Choudhary, Raghuraman Krishnamoorthi, Carole-Jean Wu, Satish Nadathur, Changkyu Kim, Maxim Naumov, Sam Naghshineh, and Mikhail Smelyanskiy. 2021. Low-Precision Hardware Architectures Meet Recommendation Model Inferen...

    work page doi:10.1109/mm.2021.3081981 2021

  6. [6]

    Cem Dilmegani. 2025. Top 20 AI Chip Makers: NVIDIA & Its Competitors in

    2025

  7. [7]

    https://aimultiple.com/ai-chip-makers#5-edge-ai-chips

    AI Multiple. https://aimultiple.com/ai-chip-makers#5-edge-ai-chips

  8. [8]

    A. Etzioni. 1998. Voluntary simplicity: characterization, select psychological implications, and societal consequences.Journal of Economic Psychology19 (1998), 619–643

    1998

  9. [9]

    Ahmad Faiz, Sotaro Kaneda, Ruhan Wang, Rita Osi, Prateek Sharma, Fan Chen, and Lei Jiang. 2024. LLMCarbon: Modeling the End-to-End Carbon Footprint of Large Language Models. InProceedings of the International Conference on Learning Representations (ICLR)

    2024

  10. [10]

    Ploipailin Flynn, Nadia Piet, Sofia Vieira, Ana Marques, Kaashvi Kothari, Nadia Nadesan, Gabriella Garcia, Arnab Chakravarty, et al. 2025. AIxDESIGN: Slow AI. Online project. https://aixdesign.co/projects/slow-ai

    2025

  11. [11]

    Daniel Geiger, Ellen Thomas, and Alastair Barr. 2023. Data centers are sprouting up as a result of the AI boom, minting fortunes, sucking up energy, and changing rural America.Business Insider(13 October 2023)

    2023

  12. [12]

    D. Gope, G. Dasika, and M. Mattina. 2019. Ternary Hybrid Neural-Tree Networks for Highly Constrained IoT Applications.arXiv preprintarXiv:1903.01531 (2019)

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  13. [13]

    2020.Less is More: How Degrowth Will Save the World

    Jason Hickel. 2020.Less is More: How Degrowth Will Save the World. William Heinemann, London

    2020

  14. [14]

    Lorenz M. Hilty. 2015. Computing Efficiency, Sufficiency, and Self-sufficiency: A Model for Sustainability?. InProceedings of the First Workshop on Computing within Limits (LIMITS’15). Irvine, California. http://www.limits2015.org/papers/ limits2015-hilty.pdf

    2015

  15. [15]

    Hilty and Bernard Aebischer

    Lorenz M. Hilty and Bernard Aebischer. 2015. ICT for Sustainability: An Emerging Research Field. InICT for Sustainability, Lorenz M. Hilty and Bernard Aebischer (Eds.). Springer, Zurich, 3–36

    2015

  16. [16]

    2025.Energy and AI: World Energy Outlook Special Report

    International Energy Agency (IEA). 2025.Energy and AI: World Energy Outlook Special Report. White paper. IEA

    2025

  17. [17]

    2024.Impacts of AI Technologies on Palestinian Lives and Narra- tives

    Ameera Kawash. 2024.Impacts of AI Technologies on Palestinian Lives and Narra- tives. Position paper. 7amleh: The Arab Center for the Advancement of Social Me- dia. https://7amleh.org/storage/AI%20&%20Racism/7amleh%20-AI%20english1- 1.pdf

    2024

  18. [18]

    Michael Kwet. 2024. How US Big Tech supports Israel’s AI-powered genocide and apartheid. https://www.aljazeera.com/opinions/2024/5/12/how-us-big-tech- supports-israels-ai-powered-genocide-and-apartheid

    2024

  19. [19]

    Sophie Lambert, Margot Deruyck, Ward Van Heddeghem, Bart Lannoo, Wout Joseph, Didier Colle, Mario Pickavet, and Piet Demeester. 2015. Post-peak ICT: Graceful degradation for communication networks in an energy constrained future.IEEE Communications Magazine53, 11 (November 2015), 166–174

    2015

  20. [20]

    John Lanchester. 2025. King of Cannibal Island.The London Review of Books47, 23-24 (25 December 2025), 23–30

    2025

  21. [21]

    Steffen Lange, Johanna Pohl, and Tilman Santarius. 2020. Digitalization and Energy Consumption: Does ICT Reduce Energy Demand?Ecological Economics 176 (2020)

    2020

  22. [22]

    Pengfei Li et al . 2024. Making AI Less ’Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models.arXiv preprintarXiv:2304.03271 (2024). https://arxiv.org/abs/2304.03271

    work page arXiv 2024

  23. [23]

    Alexandra Luccioni et al . 2024. Power Hungry Processing: ’Watts’ Driving the Cost of AI Deployment?. InProceedings of the ACM Conference on Fairness, Accountability, and Transparency (ACM FAccT ’24). Rio de Janeiro, Brazil

    2024

  24. [24]

    McKinsey and Company. 2023. Investing in the rising data center econ- omy. https://www.mckinsey.com/industries/technology-media-and- telecommunications/our-insights/investing-in-the-rising-data-center- economy

    2023

  25. [25]

    Massimo Merenda, Carlo Poccaro, and Demetrio Iero. 2020. Edge Machine Learning for AI-Enabled IoT Devices: A Review.Sensors20, 9 (2020), 2533. doi:10.3390/s20092533

    work page doi:10.3390/s20092533 2020

  26. [26]

    Janine Morley, Kelly Widdicks, and Mike Hazas. 2018. Digitalisation, energy and data demand: The impact of Internet traffic on overall and peak electricity consumption.Energy Research & Social Science38 (2018), 128–137

    2018

  27. [27]

    remains astronomical

    Sebastian Moss. 2025. JPMorgan: Global data center and AI infra spend to hit $5 trillion, demand for compute "remains astronomical".Data Center Dynamics(12 November 2025). https://www.datacenterdynamics.com/en/news/jpmorgan- global-data-center-and-ai-infra-spend-to-hit-5-trillion-demand-for-compute- remains-astronomical/

    2025

  28. [28]

    Patterson, Jay Chen, Daniel Pargman, Barath Raghavan, and Birgit Penzenstadler

    Bonnie Nardi, Bill Tomlinson, Donald J. Patterson, Jay Chen, Daniel Pargman, Barath Raghavan, and Birgit Penzenstadler. 2018. Computing Within Limits. Commun. ACM61, 10 (October 2018), 86–93

    2018

  29. [29]

    Thomas Olsson et al. 2023. A User-Centered Lens into Digital Excess: Exploring the Superfluity and Environmental Burden of the Digital World. InLIMITS ’23: Proceedings of the Workshop on Computing within Limits

    2023

  30. [30]

    Ilia Pavlov. 2025. Controlling the image generation process with parametric activation functions.arXiv preprintarXiv:2510.15778v1 (2025). https://arxiv.org/ abs/2510.15778v1

    work page arXiv 2025

  31. [31]

    James Pierce. 2012. Undesigning technology: Considering the negation of design by design. InProceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI). 957–966

    2012

  32. [32]

    Hanna Pihkola, Mikko Hongisto, Olli Apilo, and Mika Lasanen. 2018. Evaluating the Energy Consumption of Mobile Data Transfer—From Technology Develop- ment to Consumer Behaviour and Life Cycle Thinking.Sustainability10 (2018), 2494–2510

    2018

  33. [33]

    Chris Preist, Daniel Schien, and Eli Blevis. 2016. Understanding and Mitigating the Effects of Device and Cloud Service Design Decisions on The Environmental Footprint of Digital Infrastructure. InProceedings of the 2016 CHI Conference on Human Factors in Computing Systems. ACM, 1324–1337

    2016

  34. [34]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis With Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 10684–10695

    2022

  35. [35]

    Tilman Santarius, Jan C. T. Bieser, Vivian Frick, Mattias Höjer, Maike Gossen, Lorenz M. Hilty, Eva Kern, Johanna Pohl, Friederike Rohde, and Steffen Lange

  36. [36]

    Annals of Telecommunications78 (2023), 277–295

    Digital sufficiency: conceptual considerations for ICTs on a finite planet. Annals of Telecommunications78 (2023), 277–295

    2023

  37. [37]

    Timo Schick and Hinrik Schütze. 2021. Small Language Models Are Also Few- Shot Learners. InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. https://arxiv.org/abs/2009.07118v2

    work page arXiv 2021

  38. [38]

    Sgouri Sgouridis, Denes Csala, and Ugo Bardi. 2016. The sower’s way: quantifying the narrowing net-energy pathways to a global energy transition.Environmental Research Letters11, 9 (2016). 9 pages

    2016

  39. [39]

    Smith, Alex Hubbard, Alex Newkirk, Nuoa Lei, Md Abu Bakar Siddik, Billie Holecek, Jonathan Koomey, Eric Masanet, and Dale Sartor

    Arman Shehabi, Sarah J. Smith, Alex Hubbard, Alex Newkirk, Nuoa Lei, Md Abu Bakar Siddik, Billie Holecek, Jonathan Koomey, Eric Masanet, and Dale Sartor. 2024.US data center energy Usage Report. Technical Report. Berkeley Lab Energy Analysis & Environmental Impacts Division, Berkeley

    2024

  40. [40]

    Small File Media Festival. 2026. Small File Media Festival. https://smallfile.ca

    2026

  41. [41]

    M.S. Smith. 2025. The hidden behemoth behind every AI answer.IEEE Spectrum (1 October 2025)

    2025

  42. [42]

    So, Chen Liang, and Quoc V

    David R. So, Chen Liang, and Quoc V. Le. 2019. The evolved transformer. In Proceedings of the 36th International Conference on Machine Learning (ICML)

    2019

  43. [43]

    Hito Steyerl. 2023. Mean Images.New Left Review140-141 (March-June 2023), 17

    2023

  44. [44]

    Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and Policy Considerations for Deep Learning in NLP.arXiv preprintarXiv:1906.02243v1 (2019). https://arxiv.org/abs/1906.02243v1

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  45. [45]

    Rosemary Thomas. 2023. Small but Powerful: A Deep Dive into Small Language Models (SLMs). Medium. https://medium.com/version-1/small-but-powerful-a- deep-dive-into-small-language-models-slms-b793bdb002f2

    2023

  46. [46]

    Neil Thompson, Kristjan Greenewald, Keeheon Lee, and Gabriel F. Manso. 2021. The Computational Limits of Deep Learning. InProceedings of the 7th Workshop on Computing within Limits (LIMITS ’21)

    2021

  47. [47]

    Patterson, M

    Bill Tomlinson, Eli Blevis, Bonnie Nardi, Donald J. Patterson, M. Six Silberman, and Yue Pan. 2013. Collapse Barua and Practice: Theory, Method, and Design. ACM Transactions on Computer-Human Interaction20, 4 (September 2013), Article 24

    2013

  48. [48]

    United States Environmental Protection Agency (EPA). [n. d.]. Greenhouse Gas Equivalencies Calculator. Online calculator. https://www.epa.gov/energy/ greenhouse-gas-equivalencies-calculator

  49. [49]

    Andrea Vetter. 2018. The Matrix of Convivial Technology: Assessing technologies for degrowth.Journal of Cleaner Production197 (2018), 1778–1786

    2018

  50. [50]

    2024.Small language models (SLMs): A cheaper, greener route into AI

    Libing Wang and Tianchong Wang. 2024.Small language models (SLMs): A cheaper, greener route into AI. Technical Report. UNESCO. Modest, artistic, and radical solutions to the environmental impact of image-generating machine learning LIMITS ’26, June 23–25, 2026, Online

    2024

  51. [51]

    Blair, Marina Jirotka, Bran Knowles, Steven Sorrell, Miriam Börjesson Rivera, Caroline Cook, Vlad C

    Kelly Widdicks, Federica Lucivero, Gabrielle Samuel, Lucas Somavilla Croxatto, Marcia Tavares Smith, Carolyn Ten Holter, Mike Berners-Lee, Gordon S. Blair, Marina Jirotka, Bran Knowles, Steven Sorrell, Miriam Börjesson Rivera, Caroline Cook, Vlad C. Coroama, Timothy J. Foxon, Jeffrey Hardy, Lorenz M. Hilty, Simon Hinterholzer, and Birgit Penzenstadler. 20...

    2023

  52. [52]

    Kelly Widdicks and Daniel Pargman. 2019. Breaking the Cornucopian Paradigm: Towards Moderate Internet Use in Everyday Life. InProceedings of the Fifth Workshop on Computing within Limits (LIMITS ’19). ACM, 1–8. doi:10.1145/ 3338103.3338105 Article 2

    work page arXiv 2019

  53. [53]

    Robin Wigglesworth. 2025. The global data centre and AI build-out will be an extraordinary and sustained capital markets event: Inside JPMorgan’s big data centre report.Financial Times(11 November 2025). https://www.ft.com/content/ 59133fa3-3071-47b4-8761-c6922d07c34e

    2025

  54. [54]

    Yuzhuang Xu et al. 2024. OneBit: Towards Extremely Low-bit Large Language Models. InProceedings of the 38th Conference on Neural Information Processing Systems (NeurIPS 2024)

    2024