pith. sign in

arxiv: 2606.19151 · v1 · pith:NRYML25Wnew · submitted 2026-06-17 · 💻 cs.CY · cs.CV

The Market in the Model: Latent Diffusion as Neural Economy

Eryk Salvaggio This is my paper

Pith reviewed 2026-06-26 19:03 UTC · model grok-4.3

classification 💻 cs.CY cs.CV
keywords latent diffusionneural economyvisual cultureplatform economyattention economygenerative image modelssocial exchangecommodification
0
0 comments X

The pith

Latent diffusion models function as neural economies that abstract social communication into vectors for sale.

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

The paper examines the mechanisms of latent diffusion models by tracing the engineering problems each component was built to solve and the theory of vision they embed. It claims this ensemble creates a contained symbolic system that turns social relations into commensurable vectors, transferring them into parcels available for exchange. A reader would care because this moves critique beyond datasets to the model's internal operations that reinforce platform and attention economies. The analysis argues that copyright-focused defenses alone risk strengthening the very commodification the model produces. It calls instead for centering how social exchange is displaced at every stage of training and generation.

Core claim

The model operates as a neural economy: a contained symbolic system that abstracts social communication into commensurable vectors as it transfers the social sphere into parcels for sale. Tracing the training and generation pipelines component by component reveals what each operation displaces, and how it further entrenches the logics of platform and attention economies over social communication.

What carries the argument

The neural economy, a symbolic system that uses model components to convert social communication into exchangeable vectors.

If this is right

  • Critiques limited to copyright and commodity defenses risk reaffirming the fetishism the model produces.
  • Understanding requires examining what each training and generation operation displaces.
  • The system inscribes platform and attention economy logics into every generated image through its component decisions.
  • Analysis must shift focus to how social exchange is abstracted rather than treating the model as a neutral black box.

Where Pith is reading between the lines

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

  • The same component-tracing approach could be applied to text or video generative models to test for similar abstraction of social data.
  • Platform companies may face pressure to redesign architectures if the displacement of social exchange becomes a regulatory concern.
  • Educational uses of these models could be evaluated for how they teach users to treat social relations as vectorized commodities.

Load-bearing premise

That histories of the model's components and the vision theory they inscribe allow tracing pipelines to show how each step displaces social communication and entrenches platform logics.

What would settle it

Evidence that component-level decisions in latent diffusion do not systematically abstract social elements into vectors or prioritize attention-economy outcomes in generated outputs.

Figures

Figures reproduced from arXiv: 2606.19151 by Eryk Salvaggio.

Figure 01
Figure 01. Figure 01: Images produced by prompting Stable Diffusion 1.5 with “Gaussian Noise.” [PITH_FULL_IMAGE:figures/full_fig_p007_01.png] view at source ↗
read the original abstract

Valuable critique of generative image models within visual culture and the humanities has emphasized the role of datasets in shaping the images they produce. Yet, close studies of the ideological positions embedded into the mechanism of the models have been neglected, leaving them imagined as "black boxes." In a bid to expand, rather than replace, dataset critique, this paper examines the mechanisms of the latent diffusion model in terms of the problems they were brought in to solve on behalf of computer vision engineers, and the decisions each component was tasked with automating. I interpret that ensemble through the histories of its parts and the theory of vision the system inscribes into every generated image. Drawing on Impett and Offert's notion of neural exchange value, I offer this analysis to argue that the model operates as a neural economy: a contained symbolic system that abstracts social communication into commensurable vectors as it transfers the social sphere into parcels for sale. Tracing the training and generation pipelines component by component reveals what each operation displaces, and how it further entrenches the logics of platform and attention economies over social communication. The paper warns that any critique fixated exclusively on copyright and commodity defenses risks reaffirming the very fetishism the model produces, and argues instead for centering social exchange.

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

2 major / 2 minor

Summary. The paper claims that latent diffusion models operate as a 'neural economy': a contained symbolic system that abstracts social communication into commensurable vectors. By tracing the historical problems each model component was designed to solve and the theory of vision inscribed in the system, the argument holds that training and generation pipelines transfer the social sphere into parcels for sale, entrenching platform and attention economy logics. It extends Impett and Offert's neural exchange value concept and advocates shifting critique from copyright and datasets toward centering social exchange.

Significance. If the interpretive framework holds, the paper would contribute to critical AI studies and visual culture by extending dataset critiques to the internal mechanisms and design decisions of generative models. It offers a conceptual lens for how model components embed economic logics, which could inform interdisciplinary work on the societal mediation of communication by AI systems. The theoretical synthesis is a potential strength for humanities-oriented analysis in the cs.CY domain.

major comments (2)
  1. [Abstract] Abstract: The central claim that 'tracing the training and generation pipelines component by component reveals what each operation displaces' is load-bearing for the neural economy interpretation and the conclusion that platform logics are entrenched, yet the provided text offers no specific, evidence-based instances of displacement or independent benchmarks against which the tracing can be assessed; the argument therefore remains dependent on the initial framing.
  2. [Abstract] Abstract: The definition of the model as a 'neural economy' that 'transfers the social sphere into parcels for sale' is introduced via the author's synthesis of component histories without external grounding or falsifiable predictions, rendering the conclusion circular with respect to the ad-hoc axiom that such tracing necessarily reveals the claimed displacements.
minor comments (2)
  1. The abstract is conceptually dense; a brief roadmap paragraph outlining the sequence of pipeline components to be analyzed would improve readability across disciplinary boundaries.
  2. Consider adding a dedicated section or table listing the core latent diffusion components (e.g., VAE encoder, denoising U-Net, scheduler) with one-sentence historical notes on the engineering problem each was introduced to address; this would make the subsequent interpretive tracing more traceable for readers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their engagement with the manuscript and for highlighting areas where the abstract's claims require clarification. We respond to each major comment below, maintaining the interpretive character of the work in critical AI studies.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that 'tracing the training and generation pipelines component by component reveals what each operation displaces' is load-bearing for the neural economy interpretation and the conclusion that platform logics are entrenched, yet the provided text offers no specific, evidence-based instances of displacement or independent benchmarks against which the tracing can be assessed; the argument therefore remains dependent on the initial framing.

    Authors: The abstract summarizes an argument developed at length in the body of the manuscript, where each component (VAE encoder, U-Net denoiser, scheduler, and text conditioning) is examined through the engineering problems it was introduced to solve and the representational reductions it enacts. These reductions are evidenced by reference to the original technical literature on each module rather than new quantitative benchmarks. The paper is situated in visual culture and critical theory; it does not claim or perform empirical validation against independent metrics. The interpretive claim therefore rests on historical and conceptual tracing, not on the absence of benchmarks. revision: no

  2. Referee: [Abstract] Abstract: The definition of the model as a 'neural economy' that 'transfers the social sphere into parcels for sale' is introduced via the author's synthesis of component histories without external grounding or falsifiable predictions, rendering the conclusion circular with respect to the ad-hoc axiom that such tracing necessarily reveals the claimed displacements.

    Authors: The neural-economy framing is explicitly positioned as an extension of Impett and Offert's established concept of neural exchange value, supplying external grounding. The component histories are drawn from peer-reviewed computer-vision literature and are independent of the interpretive conclusion; the synthesis applies that prior concept to the specific architecture and training regime of latent diffusion. The paper does not present falsifiable predictions because its contribution is a theoretical lens for humanities-oriented analysis rather than a positivist hypothesis test. The argument is therefore not circular but deductive from documented technical decisions. revision: no

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper advances an interpretive argument in visual culture and critical theory, framing latent diffusion models as a 'neural economy' by tracing component histories and drawing on external notions such as Impett and Offert's neural exchange value. No equations, fitted parameters, predictions, or first-principles derivations are present that could reduce to inputs by construction. The central claim rests on historical framing and component-wise analysis rather than self-referential definitions or self-citation chains. This is a self-contained interpretive essay without the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The paper relies on interpretive domain assumptions about model components encoding economic logics and introduces the new concept of neural economy without independent evidence or falsifiable handles; analysis is based on abstract only.

axioms (2)
  • domain assumption The ensemble of components in latent diffusion models inscribes a specific theory of vision into every generated image.
    Invoked as the basis for interpreting the model through histories of its parts.
  • ad hoc to paper Tracing the training and generation pipelines component by component reveals displacements that entrench platform and attention economies.
    Central interpretive step with no independent verification provided.
invented entities (1)
  • neural economy no independent evidence
    purpose: To frame the latent diffusion model as a contained symbolic system that abstracts social communication into vectors for sale.
    New concept introduced to organize the critique; no falsifiable handle or external evidence supplied.

pith-pipeline@v0.9.1-grok · 5742 in / 1309 out tokens · 39762 ms · 2026-06-26T19:03:11.874790+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

32 extracted references · 24 canonical work pages · 8 internal anchors

  1. [1]

    Evaluating CLIP: Towards Characterization of Broader Capabilities and Downstream Implications

    “Evaluating CLIP: Towards Characterization of Broader Capabilities and Downstream Implications.” arXiv preprint.http://arxiv.org/abs/2108.02818. 15 Alenichev, Arsenii, Patricia Kingori, Jonathan Shaffer, and Koen Peeters Grietens

    work page arXiv

  2. [2]

    The Elephant in the Room: Reflecting on Text-to-Image Generative AI and Global Health Images

    “The Elephant in the Room: Reflecting on Text-to-Image Generative AI and Global Health Images.” BMJ Global Health 9 (4): e015601.https://doi.org/10.1136/bmjgh-2024-015601. Amoore, Louise, S. J. Bennett, Alexander Campolo, Benjamin Jacobsen, and Ludovico Rella

    work page doi:10.1136/bmjgh-2024-015601 2024

  3. [3]

    Vatai, E., Drozd, A., Ivanov, I

    “Politics of the Prompt: Government in the Age of Generative AI.” Economy and Society 54 (3): 573–96.https: //doi.org/10.1080/03085147.2025.2560177. Baack, Stefan

    work page doi:10.1080/03085147.2025.2560177 2025

  4. [4]

    A Critical Analysis of the Largest Source for Generative AI Training Data: Common Crawl

    “A Critical Analysis of the Largest Source for Generative AI Training Data: Common Crawl.” In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency. New York, NY, USA: ACM. Bianchi, Federico, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan

    2024

  5. [5]

    Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale

    “Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale.” In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, 1493–1504. New York, NY, USA: ACM. Birhane, Abeba, Vinay Prabhu, Sang Han, Vishnu Naresh Boddeti, and Alexandra Sasha Luccioni

    2023

  6. [6]

    Into the LAIONs Den: Investigating Hate in Multimodal Datasets

    “Into the LAIONs Den: Investigating Hate in Multimodal Datasets.” arXiv preprint.http://arxiv. org/abs/2311.03449. Birhane, Abeba, Vinay Uday Prabhu, and Emmanuel Kahembwe

    work page arXiv

  7. [7]

    Multimodal Datasets: Misogyny, Pornography, and Malignant Stereotypes

    “Multimodal Datasets: Misogyny, Pornography, and Malignant Stereotypes.” arXiv preprint.http://arxiv.org/abs/2110.01963. Bourdieu, Pierre

    work page arXiv

  8. [8]

    Excavating AI: The Politics of Images in Machine Learning Training Sets

    “Excavating AI: The Politics of Images in Machine Learning Training Sets.” AI & Society.https://doi.org/10.1007/s00146-021-01162-8. Demir, Ugur, and Gozde Unal

    work page doi:10.1007/s00146-021-01162-8

  9. [9]

    Patch-Based Image Inpainting with Generative Adversarial Networks

    “Patch-Based Image Inpainting with Generative Adversarial Networks.” arXiv preprint.http://arxiv.org/abs/1803.07422. Deng, Jia, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    ImageNet: A Large-Scale Hierarchical Image Database

    “ImageNet: A Large-Scale Hierarchical Image Database.” In Proceedings of the 2009 IEEE Conference on Computer Vision and Pattern Recognition, 248–55. IEEE. Denton, Remi, Mark Díaz, Ian Kivlichan, Vinodkumar Prabhakaran, and Rachel Rosen

    2009

  11. [11]

    Whose Ground Truth? Accounting for Individual and Collective Identities Underlying Dataset Annotation

    “Whose Ground Truth? Accounting for Individual and Collective Identities Underlying Dataset Annotation.” arXiv preprint.http://arxiv.org/abs/2112.04554. Devlin, Jacob, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova

    work page arXiv

  12. [12]

    Taming Transformers for High-Resolution Image Synthesis

    “Taming Transformers for High-Resolution Image Synthesis.” arXiv preprint.http://arxiv.org/abs/2012.09841. Fazi, M. Beatrice

    work page arXiv 2012

  13. [13]

    The TESCREAL Bundle: Eugenics and the Promise of Utopia through Artificial General Intelligence

    “The TESCREAL Bundle: Eugenics and the Promise of Utopia through Artificial General Intelligence.” First Monday.https://doi.org/10.5210/fm.v29i4.13636. Ghosh, Sourojit, Pranav Narayanan Venkit, Sanjana Gautam, Shomir Wilson, and Aylin Caliskan

    work page doi:10.5210/fm.v29i4.13636

  14. [14]

    Do Generative AI Models Output Harm While Representing Non-Western Cultures: Evidence from A Community-Centered Approach

    “Do Generative AI Models Output Harm While Representing Non-Western Cultures: Evidence from A Community-Centered Approach.” arXiv [Cs.CY].https://doi.org/10.48550/arXiv.2407.14779. 16 Hanna, Alex, and Tina M. Park

    work page doi:10.48550/arxiv.2407.14779

  15. [15]

    Against Scale: Provocations and Resistances to Scale Thinking

    “Against Scale: Provocations and Resistances to Scale Thinking.” arXiv preprint.http://arxiv.org/abs/2010.08850. Ho, Jonathan, and Tim Salimans

    work page arXiv 2010

  16. [16]

    Classifier-Free Diffusion Guidance

    “Classifier-Free Diffusion Guidance.” arXiv preprint.http: //arxiv.org/abs/2207.12598. Impett, Leonardo, and Fabian Offert

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    Image-to-Image Translation with Conditional Adversarial Networks

    “Image-to-Image Translation with Conditional Adversarial Networks.” arXiv preprint.http://arxiv.org/abs/1611.07004. Latour, Bruno

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Where Have We Been? Where Are We Going?

    “Where Have We Been? Where Are We Going?” Presentation at the Beyond ImageNet Large Scale Visual Recognition Challenge Workshop, CVPR 2017.https://image-net.org /static_files/files/imagenet_ilsvrc2017_v1.0.pdf. Luccioni, Alexandra Sasha, and Joseph D. Viviano

    2017

  19. [19]

    The New Value of the Archive: AI Image Generation and the Visual Economy of ‘Style

    “The New Value of the Archive: AI Image Generation and the Visual Economy of ‘Style.”’ IMAGE. Zeitschrift Für Interdisziplinäre Bildwissenschaft. 19 (1): 100–111.https://doi.or g/10.25969/MEDIAREP/22314. Pasquinelli, Matteo

    work page doi:10.25969/mediarep/22314

  20. [20]

    Scalable Diffusion Models with Transformers

    “Scalable Diffusion Models with Transformers.” arXiv [Cs.CV]. https://doi.org/10.48550/arXiv.2212.09748. Qadri, Rida, Renee Shelby, Cynthia L. Bennett, and Remi Denton

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2212.09748

  21. [21]

    AI’s Regimes of Representation: A Community-Centered Study of Text-to-Image Models in South Asia

    “AI’s Regimes of Representation: A Community-Centered Study of Text-to-Image Models in South Asia.” arXiv [Cs.CY].http://arxiv. org/abs/2305.11844. Radford, Alec, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever

    work page arXiv

  22. [22]

    Learning Transferable Visual Models From Natural Language Supervision

    “Learning Transferable Visual Models from Natural Language Supervision.” arXiv preprint.http: //arxiv.org/abs/2103.00020. Reisner, Alex

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    The Company Quietly Funneling Paywalled Articles to AI Developers

    “The Company Quietly Funneling Paywalled Articles to AI Developers.” The Atlantic, November 4, 2025.https://www.theatlantic.com/technology/2025/11/common-crawl-ai-train ing-data/684567/. Rombach, Robin, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer

    2025

  24. [24]

    High-Resolution Image Synthesis with Latent Diffusion Models

    “High- Resolution Image Synthesis with Latent Diffusion Models.” arXiv preprint.http://arxiv.org/abs/ 2112.10752. Ronneberger, Olaf, Philipp Fischer, and Thomas Brox

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    U-Net: Convolutional Networks for Biomedical Image Segmentation

    “U-Net: Convolutional Networks for Biomedi- cal Image Segmentation.” arXiv preprint.http://arxiv.org/abs/1505.04597. Salvaggio, Eryk

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [26]

    How to Read an AI Image

    “How to Read an AI Image.” Cybernetic Forests. October 2, 2022.https: //mail.cyberneticforests.com/how-to-read-an-ai-image/. ———. n.d. “The Fixed-Explosive: Language Models and the Blurry Subject.” Manuscript under review. Shoemaker, Tyler

    2022

  27. [27]

    – American Behavioral Scientist 43 (3): 377–391

    “The Ethnography of Infrastructure.” American Behavioral Scientist 43 (3): 377–91. https://doi.org/10.1177/00027649921955326. Sterne, Jonathan

    work page doi:10.1177/00027649921955326

  28. [28]

    Mean Images

    “Mean Images.” New Left Review, no. 140: 82–97.https://doi.org/10.64590/uhm. Terranova, Tiziana

    work page doi:10.64590/uhm

  29. [29]

    Free Labor

    “Free Labor.” Social Text 18 (2): 33–58.https://doi.org/10.1215/01642472 -18-2_63-33. Turk, Victoria

    work page doi:10.1215/01642472

  30. [30]

    Valéry, Paul

    https://restofworld.org/2023/ai-image-stereotypes/. Valéry, Paul

    2023

  31. [31]

    Attention Is All You Need

    “Attention Is All You Need.” arXiv [Cs.CL].https://doi.org/10.48550 /arXiv.1706.03762. Zhang, Richard, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang

    work page internal anchor Pith review Pith/arXiv arXiv

  32. [32]

    The Unreason- able Effectiveness of Deep Features as a Perceptual Metric

    “The Unreason- able Effectiveness of Deep Features as a Perceptual Metric.” In Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 586–95. IEEE. Žižek, Slavoj

    2018