Post-Modern Neuralism:
A Formal Analysis of Early Generative AI Art as a Distinct Art Historical Period

The question of when a technology becomes a medium — and when a medium produces a movement — is one that art history has addressed repeatedly and inconsistently. Photography took decades to achieve institutional legitimacy. Video art was dismissed as novelty for years before its formal properties were systematically analyzed. The pattern is consistent: new image-making technologies arrive, produce work that is initially legible only to those who made it, and are eventually canonized once the moment has passed and the historians have caught up.

We are writing from inside such a moment — or rather, from just after its close. The window of early diffusion model generativity that we term Post-Modern Neuralism lasted, as a coherent aesthetic period, from approximately January 2022 to September 2022. In that time, a small community of practitioners — working primarily with Midjourney v2 and v3 — produced a body of work whose formal characteristics are as distinctive, and as worthy of systematic analysis, as those of Fauvism, Abstract Expressionism, or any other named art historical moment.

The purpose of this paper is to make that case with the rigor it deserves.

The history of art is, in significant part, a history of constraint. The Impressionists did not paint as they did because they preferred blurred edges; they painted as they did because the arrival of photography had made hyper-realism redundant, and because painting in en plein air required fast, approximate application. The constraint produced the aesthetic.

Similarly: early photography's long exposure times produced the ghostly, depopulated cityscapes of Daguerre — not because he wished to capture absence, but because the medium could not yet capture motion. The hallmark of each new medium's early period is a set of formal features that emerge directly from what the technology cannot yet do.

Post-Modern Neuralism follows this pattern precisely. The formal features of early diffusion model art — what we will describe in detail in Section 3 — are not aesthetic choices in any intentional sense. They are the visible structure of the model's incomprehension. And it is precisely this quality that gives them their historical interest.

We note also a precedent from the digital art tradition. Early computer graphics of the 1960s and 70s — the work of Vera Molnár, Charles Csuri, and others — are now canonized not despite their technical limitations but because of them. The pixelated, vector- approximate quality of that work is understood as an authentic record of the computational capacities of its moment. We argue that the same recontextualization is both possible and necessary for the work of the early diffusion model period.

We identify six formal characteristics that, in combination, define the Post-Modern Neuralism aesthetic. These are not all present in every artifact; they are, collectively, the distinguishing features of the period.

The Pop Culture Babies collection — initiated by Topher Welsh in January 2022 and expanded with Eugene Capon as Principal Curator from February 2022 — constitutes the largest systematically archived body of work from the early Midjourney era. With 233 artifacts across 32 thematic series, it provides the statistical and formal basis for the analysis in Section 3.

The collection's particular value as a case study derives from its systematic design. Welsh did not generate images randomly; he generated series — the same set of characters rendered in sequence — which allows for the analysis of the model's learning and consistency within a single prompt domain. The Hulk Generations series (seven iterations) and the Machine Gun Kelly series (six iterations) are particularly useful in this regard: they document the model's iterative refinement of a single subject, making visible the stochastic quality of the generation process.

The collection also benefits from its cultural legibility. The subjects — Marvel characters, DC characters, Star Wars characters, classic television — are among the most thoroughly represented subjects in internet image archives. This means that the model's failures with these subjects are not failures of training data coverage; they are failures of synthesis. The model had seen Tony Stark many thousands of times. Its inability to correctly render his arc reactor is therefore an interesting fact about the limits of pattern recognition, not a fact about data scarcity.

The infant framing of the collection introduces a formal tension that amplifies the hallucination signatures. The model was required to hold two registers simultaneously: the highly specific visual language of superhero aesthetics, and the highly specific visual language of infant portraiture. Its failure to fully synthesize these registers — the suit that becomes a onesie, the weapon that becomes a toy, the power that becomes a birth mark — is not a failure of either register individually. It is a failure of the interface between them. That interface is where the art lives.

The authorship question for generative AI art has been largely unproductive because it has been framed incorrectly. The question "did the AI make this?" implies that authorship requires a single, bounded agent. But the history of collaborative and process-based art — from the workshops of the Renaissance to the conceptual practice of Sol LeWitt — suggests that authorship is better understood as a set of decisive acts than as a single source of causation.

In the production of the Pop Culture Babies collection, the decisive acts were: the selection of subjects, the design of the prompt language, the curation of generations, the documentation of artifacts, and the archival decision to preserve the hallucinations rather than correct them. These acts were performed by human practitioners. The pixel generation was performed by the model.

We propose that the correct title for practitioners in this mode is not "AI artist" — which is ambiguous to the point of uselessness — but Technologist-Curator: a practitioner who designs the conditions of generation and exercises archival judgment over the results. This is a legitimate and historically-grounded creative role, analogous to the photographer who does not grind their own lenses but makes decisive choices about what to point them at.

Post-Modern Neuralism is, by definition, a closed period. The formal features that define it — topological ambiguity, procedural overflow, phantom geometry — were engineering problems, and they were solved. By Midjourney v4 (released October 2022), the finger count was largely correct. By v5 (March 2023), the remaining hallucination signatures had been substantially attenuated. The models had learned to perform anatomical correctness.

This is not a criticism of the models. The improvements were the point; they made the technology more useful. But usefulness and aesthetic interest are not the same thing, and the improvement trajectory of diffusion models represents a specific kind of aesthetic closure: the closure that comes when a medium learns to conceal its own process.

The Post-Modern Neuralism period is interesting precisely because the process was not concealed. Every generation from this period is a primary document: a record of what the model understood, in that specific training state, about the visual world. These records are non-reproducible. You cannot regenerate them, because you cannot un-train a model. You can only preserve what was generated before the training moved on.

The obligation of preservation is therefore absolute.

We address, briefly, the implications of this analysis for museums, galleries, and digital archives.

First: the metadata problem. Early AI generations were often archived without prompt documentation, generation parameters, or model version information. The Pop Culture Babies collection is an exception: the prompt structure is partially recoverable from the filename conventions (tophmagoph_ prefix followed by a truncated prompt description) and the generation dates are documented. Institutions acquiring work from this period should prioritize collections with this level of documentation.

Second: the file integrity problem. PNG and JPEG formats are non-archival by default. Long-term preservation of early AI art requires migration to archival formats (TIFF, preservation-grade PNG) and redundant storage. The physical print medium is, for this purpose, not a substitute for digital preservation — but it is a useful redundancy layer, and one that museums should consider as part of their acquisition protocols.

Third: the access problem. Post-Modern Neuralism material should be publicly accessible, not held in closed institutional collections. The work was produced for mass audiences; its historical significance was demonstrated by mass engagement (16.9 million reach for the Pop Culture Babies collection alone). Institutional preservation should not mean institutional enclosure.

We have argued that the early diffusion model AI art of 2022 constitutes a formally coherent aesthetic period — Post-Modern Neuralism — with specific, analyzable characteristics and a clear historical boundary. We have situated this period within the broader history of technologically-mediated art movements, conducted a formal analysis of its defining features, and argued for its systematic preservation.

The name we have proposed — Post-Modern Neuralism — is intended to be precise rather than clever. It is post-modern in the sense that it operates after the collapse of the idea that images represent reality; it is neural in the sense that it is produced by artificial neural networks; it is a realism in the sense that it represents the model's genuine, unmediated attempt to render the visual world as it understood it.

The hallucinations are not footnotes to this period. They are its primary text.

We invite comment, criticism, and collaboration from practitioners, historians, and institutions working in this space.