« Peux-tu reproduire cette peinture aussi fidèlement et précisément que possible ? » La question semble presque absurdement simple. Pourtant, lorsqu’on la pose à une IA générative, celle-ci ne reproduit pas l’image à l’identique.

Latent Archetypes commence par soumettre une peinture à ChatGPT accompagnée de ce prompt. L’image obtenue est ensuite soumise à nouveau avec exactement la même demande, chaque fois sans aucune mémoire de la séquence précédente.

Au début, presque rien ne semble se produire. Pourtant, d’imperceptibles différences s’accumulent. Lentement, l’image dérive loin d’elle-même : ce qui commence comme une tentative de reproduction parfaite devient un processus de transformation, donnant naissance à une étrange forme d’irreproductibilité technique. À chaque itération, quelque chose se perd, se simplifie, se déplace. L’image devient de plus en plus mystérieuse et, en un sens, plus pauvre que l’original.

L’IA générative semble ramener la reproduction à une condition pré-photographique, dans laquelle toute copie est aussi une reconstruction, et donc une interprétation. Pourtant, cela se produit au sein d’un dispositif capable de générer des images apparemment sans limite. Un dispositif de reproduction illimitée produit une irreproductibilité radicale.

Mais un autre paradoxe apparaît. Après de nombreuses itérations, la dérive peut ralentir et graviter autour de familles de formes persistantes ou métastables. Je les appelle des archétypes latents. Ce ne sont pas des originaux cachés dans la machine, mais des formes qui émergent parce que certaines configurations visuelles semblent plus reproductibles que d’autres par le système génératif. L’irreproductibilité se replie ainsi sur la reproductibilité : l’original ne peut être répété fidèlement, mais les tendances statistiques propres à la machine, elles, semblent pouvoir l’être.

Latent Archetypes part ainsi d’un original qui ne peut être reproduit pour aboutir à des formes où un autre type de reproductibilité semble émerger, sans qu’elles n’aient jamais été des originaux.

Christophe Bruno, 2026

Latent Archetypes (Christophe Bruno, 2026) explores what happens when the failure of generative AI to reproduce an image identically is itself repeated indefinitely. The process first produces drift, but after many iterations it may also give rise to persistent or metastable families of forms. The project asks whether the failure of exact reproducibility can paradoxically reveal another kind of reproducibility.

It all began when I uploaded a painting I had made around the age of seventeen to ChatGPT, with a simple request: “Can you reproduce this painting as faithfully and accurately as possible?” I then fed the resulting image back into the system, again and again, always using the same prompt. Each iteration was carried out in a new conversation, with memory and chat-history reference disabled, and I provided no information from the preceding sequence beyond the image being reconstructed.

The first surprise of this ridiculously simple process was that the image gradually changed, drifting almost imperceptibly away from itself.

As I observed the accumulation of these minute differences, a second and perhaps even more significant surprise emerged: after many iterations the image began to enter phases of relative stabilization.

After repeating the experiment with several paintings and observing comparable processes of drift and periods of relative stabilization, I began to call these emergent liminal forms “latent archetypes”: “latent” because they emerge indirectly from the model’s learned representations rather than being explicitly encoded as images, and “archetypes” because they appear as persistent forms or families of forms around which the trajectories may temporarily organize. This raises the question of whether a latent archetype should be understood as a pre-existing primitive form hidden within the machine, or rather as an emergent property of the trajectory itself.

The resulting image may look extremely close to the original to be copied, but the system does not simply return a pixel-identical duplicate. It generates a new image conditioned by the visual input and the prompt, through representations acquired during training.

It is therefore a reconstruction rather than an identical duplicate. Even when the result appears faithful to the human eye, the image has already undergone subtle modifications. In Latent Archetypes, these minimal differences accumulate from one iteration to the next, gradually transforming apparent reproduction into drift.

On the one hand, generative AI returns the copy to a pre-photographic condition, in which every reproduction is inherently an interpretation and exact identity can no longer be guaranteed. This amounts to a return to technological irreproducibility, reversing Walter Benjamin’s paradigm of technological reproducibility (cf. section II of this FAQ).

On the other hand, the repetition of this very process gives rise to residual forms that may persist, recur, or temporarily stabilize across the trajectory. Irreproducibility thus folds back into reproducibility: not as the faithful repetition of the original, but as the recurring production of configurations that appear statistically robust under repeated reconstruction.

This is the central paradox of Latent Archetypes: the failure to reproduce an image identically eventually reveals another kind of reproducibility—the reproducibility of archetypal tendencies.

The precise mechanism depends on the image-generation model being used. In each case, however, the image is not treated simply as a fixed arrangement of pixels to be duplicated. It is processed through internal mathematical representations learned during training from very large collections of images.

In some generative architectures, part of this internal organization can be described in terms of “latent spaces”: a mathematical space in which visual features—such as forms, colors, textures, objects, styles, and compositions—are encoded through statistical relationships. The exact nature of these representations depends on the architecture and cannot be inferred directly from the present experiment. Rather than functioning as a conventional image database that simply retrieves and returns the uploaded image, the model uses learned statistical relationships to generate a new image conditioned by the visual input and the prompt.

In diffusion models, for instance, image generation involves noise and its progressive removal, guided by patterns learned during training. Other image models may operate differently. For Latent Archetypes, the important point is that each image is subjected to a generative transformation whose output becomes the input of the next iteration, rather than returned as a pixel-by-pixel duplicate. This process introduces a degree of uncertainty and variation, however slight it may initially appear.

For Latent Archetypes, the choice of ChatGPT is not neutral. The project is not only testing a generative image model; it is exploring the visual tendencies of one of the most widely used human–machine interfaces of its time.

Other generative systems could be used, and they would probably produce different trajectories, forms of drift, and latent archetypes. Each model is shaped by its own architecture, training process, datasets, visual conventions, and technical limitations. There is therefore no reason to assume a single latent space or a universal archetype shared by all generative systems.

For now, I carry out the process manually. This makes it slow, repetitive, and almost ritualistic. Automating the iterations through an API or an open-source image model would make larger experiments possible, but it would also transform the conditions of the work.

A comparative study involving several generative systems would be especially revealing. It could show which tendencies are specific to a particular model and which forms recur across different systems. But this remains a future stage of the project. For comparative experiments, recording the exact date, model designation when available, prompt, image dimensions, file format, and all accessible generation settings would be essential.

Yes. The trajectory may change from one version of a system to another, particularly when its image-generation model, training process, interface, or internal parameters are modified.

A newer version may preserve the starting image more faithfully, introduce different kinds of distortion, or follow a different trajectory, enter a different metastable regime, or stabilize around a different family of forms. Even when the prompt and the initial painting remain unchanged, the behavior of the system may no longer be the same.

If a future system were to return a truly exact, pixel-identical copy of the uploaded image, then, for the purposes of this protocol, the operation would become indistinguishable from simple duplication and the observable drift would disappear. As long as the image is regenerated rather than merely copied, however, some degree of displacement may remain—even if it becomes slower, subtler, or almost invisible.

For this reason, Latent Archetypes is historically and technically situated. It does not reveal the archetypes of “AI” in general, but those produced by a particular system at a particular moment. The model version, the date, the interface, and the protocol are therefore not incidental details: they are part of the work itself.

A latent archetype is not simply the last image produced in a sequence. It is a persistent or metastable visual configuration that emerges after a prolonged period of drift, when successive reconstructions begin to preserve the same general composition, forms, colors, and spatial relationships.

Because the generation process remains variable, the archetype does not necessarily correspond to one perfectly fixed image. It may be better understood as a family of closely related images that preserve the same general visual configuration.

At this stage, “latent archetype” is therefore an operational and observational category rather than a mathematically established state of the system. For now, I identify such regimes visually: the drift slows down, the main structure stops undergoing significant transformations, and further iterations produce variations belonging to what appears to be the same family of forms. This does not imply that the trajectory will remain there permanently. More systematic criteria could eventually be developed to measure stabilization, persistence, and escape. A quantitative definition could eventually combine similarity between successive images, the duration of a low-variation regime, and the probability of remaining within or returning to the same visual family.

After many iterations, the drift may slow down and the sequence may remain for a long time within a region of relatively low variation. Certain compositions, forms, colors, and spatial relationships are then reproduced more consistently than others.

However, recent experiments suggest that this stabilization is not necessarily permanent. A sequence may remain visually stable for dozens of iterations and then suddenly undergo a larger transformation. In the language of dynamical systems, such a long-lived yet transient state can be described as metastable.

One possible explanation is that repeated reconstruction gradually removes features that the model preserves less reliably while allowing forms and conventions that it reproduces more consistently to become increasingly dominant. Yet the probabilistic character of generation continues to introduce fluctuations, and some of these may eventually push the trajectory out of one region and into another.

The notion of an attractor therefore remains a useful analogy, but it should not necessarily be understood as a unique fixed endpoint. A latent archetype may instead correspond to a persistent family of forms, a metastable region, or one among several possible regimes through which the trajectory can pass. For now, metastable regime is therefore the more conservative description. “Attractor” remains a useful hypothesis or analogy that would require additional tests — for example, perturbing an image within such a regime and measuring whether repeated reconstruction tends to bring the trajectory back toward the same family of forms.

This remains a conceptual interpretation rather than a mathematical characterization of the observed dynamics.

Not necessarily—and the trajectory may not settle permanently at all. Because each generation includes an element of variation, two sequences beginning with the same image may gradually diverge and follow different trajectories.

They might eventually return to closely related forms, but they could also enter different metastable regions—what might, if confirmed experimentally, be described as different basins of attraction—and produce distinct latent archetypes. In that case, the starting image would not contain a single predetermined destination, but several possible futures within the effective visual landscape of the process.

This question has not yet been tested. Systematically repeating the experiment many times from exactly the same starting image would make it possible to estimate a distribution of possible trajectories, and to determine whether the resulting regimes are consistent, belong to the same family, or regularly bifurcate toward different forms.

Possibly—and this would be one of the strongest indications that latent archetypes are not simply transformed remnants of their starting images.

If paintings with very different origins were repeatedly drawn toward similar compositions or visual structures, this would suggest that they had entered the same or closely related metastable regions within the effective visual landscape of the process. Such convergence would become particularly significant if it occurred repeatedly, beyond the level expected from generic similarities between generated images or from chance clustering. If such convergence proved systematic, these regions might be described in terms of basins of attraction. The resulting form would then appear to belong less to any individual painting than to the statistical organization of the generative process itself.

The convergence might be complete, or it might concern only certain features: a composition, a color structure, a type of figure, a spatial arrangement, or a particular visual atmosphere.

This remains an open experimental question. It would require comparing many trajectories and developing ways to determine when two final forms are genuinely related rather than merely similar to the human eye.

Not all starting images appear to drift at the same rate or to the same degree. In some of my experiments with extremely famous paintings, the transformations have remained relatively close to recognizable features of the original, even though the resulting images are never strictly identical to it.

One possible explanation is that highly familiar images correspond to visual and semantic patterns that the system has learned to reconstruct particularly robustly. Their compositions, subjects, and stylistic features also circulate extensively through reproductions, descriptions, adaptations, and references.

This could contribute to their relative stability, but the present experiment cannot determine how frequently a particular artwork, or related images, actually occurred in the model’s training data. Other factors—composition, recognizability, style, complexity, subject matter, or the model’s general visual knowledge—may also play an important role.

By contrast, when the starting image is obscure, private, or highly singular—as in the case of my early painting—the drift may be much stronger. The reconstruction may then depend more strongly on general visual and semantic regularities than on patterns specifically associated with a highly familiar work. Over successive iterations, its singular features may gradually be replaced by forms that the system reconstructs more consistently.

This remains a hypothesis rather than a demonstrated rule. The degree of drift may also depend on the composition, style, level of detail, prompt, model version, and generation settings. But the experiment suggests that technological reproducibility is uneven: some images seem to resist transformation more effectively than others. This hypothesis could be tested by comparing famous works with unfamiliar images matched as closely as possible for composition, subject matter, style, and visual complexity.

In this sense, the experiment may also probe something resembling a collective visual memory: statistical regularities inherited from the circulation of images through culture and filtered through the model’s training process.

To some extent, yes. The distance between successive images could be tracked using several complementary methods: perceptual similarity, changes in color, composition, edges or textures, and representations produced by an independent vision model. These measurements could indicate whether the amount of change decreases over time, whether the trajectory enters a relatively stable regime, how long it remains there, and whether it later escapes into another one.

It would also be useful to distinguish between two different kinds of change: the variation occurring from one iteration to the next, and the accumulated distance from the original image. A trajectory could, for example, continue to move very little from one step to the next while already having drifted very far from its starting point.

Multiple runs could then be compared to determine whether they remain close together, diverge, revisit similar configurations, or repeatedly enter the same metastable regions. One could also measure how long a trajectory remains within a relatively stable visual family before leaving it, and whether it later returns.

However, no single measurement can fully describe what is changing in an image. A small numerical difference may correspond to an important artistic transformation, while a large pixel-level difference may leave the overall composition almost unchanged. Quantitative analysis would therefore need to combine several complementary measurements with visual and interpretive judgment.

The aim would not necessarily be to prove the existence of latent archetypes in an absolute sense, but to characterize more precisely the dynamics of drift, stabilization, recurrence, and escape through which they appear.

Yes, at least as a first approximation. Each reconstruction does not produce one strictly predetermined result, but one possible result among a range of images that the generative system could produce from the current image and the fixed prompt.

The experiment can therefore be described as a stochastic iterative process: each generated image becomes the current state from which the next reconstruction begins. Under sufficiently controlled conditions, the next result can therefore be modeled as depending primarily on the image currently being reconstructed and on the prompt, rather than on the visible history of the sequence.

In probabilistic terms, this resembles what is called a Markovian process. This does not mean that ChatGPT is internally a Markov chain, nor does it provide a description of the model’s architecture. It is simply a way of describing the observable dynamics of the experimental protocol.

Within such a framework, a latent archetype does not need to be a perfectly fixed image. It could instead correspond to a metastable region: a family of related configurations within which the trajectory has a relatively high probability of remaining for many iterations before eventually escaping.

Repeating the experiment many times would then make it possible to study how often particular visual regimes are entered, how long trajectories remain within them, how frequently they escape or return, and whether different starting images repeatedly reach related regions.

This probabilistic description would therefore provide a way of testing the structure suggested by the project without assuming that latent archetypes already exist as hidden images inside the model.

Only by analogy. In its technical sense, model collapse describes a degenerative process in which successive generative models are trained on data produced by previous models. Over generations of training, information about the original data distribution may be lost, particularly in its less frequent or more weakly represented regions.

Latent Archetypes follows a different process. The model itself is not retrained on its outputs. Instead, the same generative system repeatedly reconstructs the image produced at the previous iteration. The parameters of the model remain approximately fixed while the image changes.

For this reason, the phenomenon observed in Latent Archetypes should not itself be called model collapse.

The comparison nevertheless becomes interesting at the level of recursive information loss. In both cases, repeated generative operations raise the question of whether some features are preserved more robustly than others, and whether singular, unstable, or weakly preserved features tend to disappear while more persistent structures increasingly dominate the result.

In Latent Archetypes, this remains an experimental hypothesis. The observed impoverishment of some trajectories suggests that repeated reconstruction may progressively filter information, but determining whether rare or statistically unusual features are systematically eliminated would require controlled comparisons across many images and repeated trajectories.

The connection with model collapse is therefore not an identity between two mechanisms, but a broader question they share: what kinds of information survive when a generative process is recursively applied to its own outputs?

The “Habsburg AI” effect, a term coined by technology researcher Jathan Sadowski in 2023, describes the progressive loss of diversity that may occur when generative models are increasingly trained on AI-generated material. In this case, recursion takes place across generations of models, progressively reshaping the distribution of possible outputs.

Latent Archetypes involves a different kind of recursion: the model remains fixed, while a single image follows a temporal trajectory through repeated reconstruction. The question is therefore less whether the distribution itself collapses than whether the trajectory can continue to explore it, or becomes trapped within a limited number of highly reproducible regions.

This raises a possible connection with ergodicity: latent archetypes could be interpreted as signs that recursive trajectories do not explore the whole accessible space, but become effectively confined to particular basins of attraction. The Habsburg effect and Latent Archetypes would then describe two different forms of recursive narrowing, one acting on the distribution, the other on trajectories within it.

(A big thank you to Alexandre Houdent for pointing me to the Habsburg effect.)

The most direct “landscape” relevant to Latent Archetypes would be an effective landscape of transition probabilities. Some visual configurations may be easier for the generative system to reconstruct consistently than others; some regions may be frequently entered, revisited, or maintained, while others may be unstable and rapidly left behind.

This should not necessarily be understood as a literal physical landscape or as a single hidden surface inside the model. It is an empirical way of describing differences in the probabilities of moving between families of images under repeated reconstruction.

The concept nevertheless resonates with other kinds of landscapes used in neuroscience, statistical physics, and machine learning.

In predictive-processing and Free Energy Principle frameworks, variational free energy is a formally defined quantity associated with probabilistic inference and the relation between a generative model and observations. A metastable visual configuration in Latent Archetypes may suggest an analogy with a region of relatively stable probability, but the sequence should not be said to literally minimize or descend a variational free-energy landscape. No such quantity is measured in the present experiment.

A related but distinct analogy can be made with machine-learning loss landscapes. During the training of a neural network, the loss is a function of the model’s parameters, and optimization changes those parameters in order to reduce the training objective. Latent Archetypes does not traverse this loss landscape: the model is not retrained at each iteration, and its parameters remain approximately fixed during a given experiment.

The learned parameters of the model nevertheless condition the probabilities of transitions between possible reconstructions. Free-energy and loss landscapes can therefore provide useful conceptual resonances, but neither constitutes a direct mathematical explanation of the observed drift, stabilization, recurrence, or escape.

For the project itself, the more conservative notion is therefore that of an “effective transition landscape” produced by repeated generative reconstruction.

The current protocol follows a single linear trajectory: one image is generated, then used as the source of the next iteration. But the landscape could be explored in many other ways.

Several images could be generated at each iteration, producing a branching structure rather than a single sequence. Repeating the process from the same starting image would reveal whether trajectories tend to remain close, diverge, or converge again. Small changes could also be introduced into the prompt, framing, resolution, colors, or starting image in order to observe how sensitive the trajectory is to its initial conditions and experimental parameters.

A latent archetype could itself become a new starting point. It could be slightly altered and reintroduced into the process to explore the region surrounding it and determine whether the image returns to the same form or escapes toward another metastable region or attractor-like regime.

Trajectories could also be transferred from one generative system to another. An image produced by one model could become the starting point for a different model, allowing us to observe how their respective visual landscapes interact.

Certain parameters could also be varied gradually in order to look for thresholds, bifurcations, or sudden changes of trajectory. A small modification might sometimes push an image from one metastable region into another.

Such experiments could also test more ambitious hypotheses concerning complex basin boundaries or critical regimes. Terms such as fractal boundary, catastrophe, or edge of chaos would require specific quantitative criteria and should therefore be treated, at this stage, as possible directions for investigation rather than descriptions of the observed process.

The project could therefore develop from a series of linear experiments into a cartography of trajectories, metastable regions, bifurcations, and possible passages between latent archetypes.

The trajectory may be difficult to predict over long horizons, especially if it can remain for extended periods in metastable regimes before escaping. This empirical unpredictability should be distinguished from stronger notions such as deterministic chaos or computational irreducibility, which the present protocol does not establish.

It may nevertheless be the case that the practical way to discover a trajectory is to let it unfold.

No. The protocol could be applied to photographs, drawings, advertisements, icons, diagrams, maps, digital images, or images produced entirely by another generative system.

Different categories of images might produce different kinds of drift and stabilization. A diagram may lose information differently from a portrait; a photograph may be reconstructed through familiar visual conventions; an abstract image may be drawn toward recognizable objects or spatial structures.

Painting nevertheless occupies a privileged place in the project. Before photography, reproducing a painting necessarily involved interpretation and transformation. Painting therefore makes the paradox of generative reproduction especially visible: one of the most advanced technologies of image production begins to behave like a copyist rather than a device for exact duplication.

Extending the experiment beyond painting would reveal whether latent archetypes belong to particular artistic traditions or whether they reflect more general tendencies in the way generative systems process images.

Walter Benjamin’s essay The Work of Art in the Age of Its Technological Reproducibility (1935) provides one of the project’s central historical reference points. Benjamin examined how photography and film transformed the status of the artwork by detaching it from its unique presence in a particular place and time and making it available for potentially unlimited reproduction and circulation.

He associated this transformation with the decline of the artwork’s “aura”: its singular authority, distance, and historical presence. Yet Benjamin’s argument was not simply nostalgic. The disappearance of aura also opened new perceptual and political possibilities.

Latent Archetypes begins from a different technological situation. A digital image can ordinarily be copied exactly, pixel for pixel. But when generative AI is asked to reproduce an image, it does not necessarily return an identical copy: it generates a new image that resembles and interprets the original.

Generative AI therefore combines two apparently contradictory conditions: the potentially unlimited production of images and the failure of exact reproduction. In this respect, it returns the copy to something resembling a pre-photographic condition, in which every reproduction introduces interpretation and difference—but it does so from within one of the most advanced apparatuses of technological reproducibility.

The project thus extends Benjamin’s question: what becomes of the artwork, its uniqueness, and its aura when technological reproduction no longer guarantees identity between an original and its copy?

In Benjamin’s account, technological reproducibility does not reveal aura by preserving it, but retrospectively, through its disappearance. Once the original can be endlessly reproduced and circulated, we become aware of what has been lost: its unique presence, historical distance, and non-interchangeability—or, to borrow a contemporary term from the world of NFTs, its non-fungibility ;).

In Latent Archetypes, generative irreproducibility produces a different effect. Because each reconstruction fails to coincide fully with its source, the original once again appears as something that cannot be entirely replaced by its copy. This does not necessarily restore the original aura in Benjamin’s sense. Rather, it reactivates the question of singularity from within the machinery of reproduction itself.

The project therefore asks whether generative AI produces a new auratic effect: not the return of the old aura, but the emergence of a technologically produced non-equivalence between original and copy.

However, when the process enters a relatively stable or metastable configuration, the question becomes even more difficult. An archetype seems to possess a powerful aura: it presents itself as an originary form from which other images might derive. Yet the latent archetype is not an origin. It is a statistical residue produced by repetition, probability, transformation, and loss.

How can an image generated as a statistical outcome or late-stage configuration acquire the authority or appearance of an origin? This unresolved contradiction lies at the heart of the project.

Warhol is an essential precedent for Latent Archetypes. His work explores repetition, seriality, mass culture, and the mechanical circulation of images. Through screenprinting, he repeatedly reproduced images taken from advertising, newspapers, publicity photographs, and consumer packaging. The same source image remains recognizable across a series, while variations in color, alignment, density, and printing make each impression slightly different.

Latent Archetypes, however, operates through recursive rather than simply serial repetition. Its images are not produced as parallel variations based on the same original. Each generated reconstruction becomes the source of the next iteration. The original image therefore becomes progressively more distant, as minute transformations accumulate and repetition turns into a trajectory.

In both Warhol’s work and Latent Archetypes, technological repetition produces difference rather than perfect identity. But Warhol stages the image as a reproducible commodity circulating through mass culture, whereas Latent Archetypes examines what happens when a generative system repeatedly fails to preserve an image and may pass through periods of relative stabilization around statistically persistent forms.

Baudrillard’s interpretation of Warhol provides another point of contact. For Baudrillard, Warhol does not restore the traditional aura of the original. Instead, he pushes the artwork toward simulation and “absolute merchandise,” while suggesting that simulation may acquire an aura of its own.

Latent Archetypes extends this paradox. Aura may first reappear as a question raised by the non-equivalence of the original and its reconstruction. It may then seem to migrate toward the latent archetype itself: a residual form that acquires the appearance and authority of an origin, even though it is actually the product of a process.

Warhol repeats an image in order to expose seriality. Latent Archetypes serializes the failure to repeat an image—and asks where that failure ultimately leads.

Not in the strict Duchampian sense. A readymade is a pre-existing, often mass-produced object that is selected by the artist, removed from its ordinary context, and designated as art. A latent archetype, by contrast, is generated through an iterative process. It is not simply found among the objects of the everyday world.

Yet the comparison remains productive. The artist does not directly design or compose the latent archetype. Instead, he establishes a protocol, allows the process to unfold, recognizes a moment of relative stabilization, and extracts an image from the continuous flow of transformations. In this sense, the archetype is encountered and selected as much as it is produced.

The analogy can be extended through some predictive-processing frameworks, which model perception not as the passive recording of an external world, but as an inferential process shaped by learned expectations and probabilities. This does not prove that latent archetypes are fundamental constituents of reality. It does, however, suggest that what appears immediately “given” may already be the result of a process of prediction, selection, and stabilization.

Latent archetypes might therefore be described as machinic readymades: forms that the artist does not invent directly, but discovers within the statistical possibilities of a trained model. They are not ordinary objects ready-made by industry, but residual images ready-made by the machine.

Not exactly—but the connection is fundamental. Aby Warburg used the term Pathosformel, or “pathos formula,” to describe recurring visual forms—gestures, postures, movements, and expressive configurations—that carry emotional intensity across different historical periods. These forms do not remain unchanged. They migrate, reappear, and acquire new meanings while preserving something of their affective charge. Warburg called this persistence and transformation the Nachleben, or “afterlife,” of images.

Latent archetypes operate differently. They are not necessarily inherited gestures that can be traced through art history, nor are they transmitted from one culture or historical period to another. They emerge through the repeated reconstruction of an image by a generative model.

Yet a large-scale generative model is trained on a visual culture in which images circulate through copies, reproductions, adaptations, stylistic conventions, recurring poses, visual clichés, and cultural memories. Its learned representations may therefore encode statistical traces of some of the survivals and migrations that Warburg sought to map historically.

From this perspective, latent archetypes might be understood as machinic Pathosformeln: residual forms in which long histories of visual repetition have been compressed, recombined, and made statistically operative. The iterative process does not simply preserve these forms; it may reveal which visual configurations the model reproduces most persistently.

Warburg traced the afterlives of images through cultural history. Latent Archetypes asks what becomes of those afterlives once they are absorbed into a generative machine—and whether the machine produces new afterlives of its own.

The latent archetypes often appear poorer than the original image: smoother, more generic, sometimes kitsch, and stripped of many of its singular details. The machine does not necessarily reveal the hidden essence of a painting. It may instead reduce it to the forms and conventions that it can reproduce most consistently.

Yet this apparent impoverishment does not make the resulting images aesthetically insignificant. Latent archetypes can be strangely compelling, precisely because they are both familiar and difficult to identify. They seem to belong to a collective visual memory without referring clearly to any single image, period, or style.

The project does not attempt to decide whether these forms are beautiful. Their beauty, when it appears, may lie less in richness, mastery, or originality than in their ambiguous status: they are at once clichés and apparitions, statistical residues and improbable images.

The project is not simply a critique of artificial intelligence, nor does it oppose painting to technology, the hand to the machine, or human creativity to automation. Instead, it uses generative AI as an experimental apparatus. By asking the system to perform an apparently simple task—reproducing an image as faithfully as possible—the project exposes the assumptions, limitations, and visual tendencies embedded in the act of generation itself.

The question is therefore not whether AI is good or bad at copying. It is what its failure to copy reveals about reproduction, interpretation, memory, probability, and the circulation of images. In this sense, Latent Archetypes does not treat AI primarily as an object of judgment, but as a medium and an experimental apparatus. It works from within the system, using its operations, tendencies, and failures as artistic material.

That remains the open question. What happens to painting when the machinery of reproduction begins to behave as if photography had never happened?

Before photography, the reproduction of a painting necessarily involved interpretation, transformation, and difference. Photography, and later digital technology, seemed to establish another regime: that of exact and potentially unlimited duplication.

Generative AI unsettles this distinction. It belongs to the computational world of infinite image production, yet when it is asked to reproduce a painting, it generates a new interpretation rather than an identical copy.

Latent Archetypes therefore confronts painting with a paradoxical return to a pre-photographic condition—produced from within one of the most advanced technologies of image generation.

The project borrows the term “archetype” from Carl Jung, but significantly transforms its meaning. In Jung’s work, archetypes are recurring structures of the collective unconscious. They are not fixed images in themselves, but underlying patterns that become visible through myths, dreams, symbols, and cultural representations.

In Latent Archetypes, the archetype is not understood as a universal psychological essence. It is a residual statistical form that emerges from the repeated reconstruction of an image by a trained model. It is shaped by datasets, visual conventions, recurring styles, clichés, cultural memories, and the accumulated habits of image production.

The AI does not necessarily possess an unconscious in the human or psychoanalytic sense. Yet its learned representations reflect statistical regularities acquired from large-scale visual and textual culture. The latent archetype may therefore evoke what could metaphorically be called a human–machine unconscious: not a psychic unconscious attributed to the AI, but a statistical residue of human visual culture as filtered through datasets, training procedures, model architecture, and generative constraints.

The project thus displaces the Jungian archetype. What once appeared as an inherited structure of the psyche re-emerges here as an acquired structure of probability.

The project also plays with a form of machinic Platonism. The model’s learned organization can create the appearance that generated images are being drawn toward pre-existing forms.

This recalls Plato’s theory of Forms, according to which the changing objects of the perceptible world participate in stable and intelligible forms. Yet the resemblance is deliberately paradoxical. The latent archetypes of the project are not eternal Ideas, nor are they more perfect or more real than the images from which they emerge.

They are historically and technically produced. They depend on a particular model, its training data, its architecture, and the visual culture from which it has learned. Rather than representing a higher level of perfection, they often appear smoother, poorer, more generic, and closer to forms that the system appears to reconstruct consistently.

Latent Archetypes therefore reverses the Platonic hierarchy. The image does not gradually approach an ideal form of greater truth or richness. It may instead organize temporarily around a residual form produced by repetition and loss—an apparent essence that is actually a statistical reduction.

The Renaissance art of memory was not simply a method for storing information. It organized knowledge through striking images placed within structured mental spaces. In Giordano Bruno’s mnemonic systems, this tradition became increasingly combinatorial and cosmological. Figures, actions, letters, and concepts were arranged within complex rotating wheels, allowing images to be associated, transformed, and recombined.

Memory became less an archive of fixed contents than a generative apparatus for producing relations between images.

Frances Yates interpreted Giordano Bruno’s art of memory as a philosophical and imaginative technique: a means not only of remembering the world, but of internalizing and reconstructing its hidden order. This provides an important precedent for Latent Archetypes. In both cases, images inhabit an organized space of possible relations, and new configurations emerge through repetition and combination. But the direction of the process is different.

Giordano Bruno constructed a system of images intended to reflect an underlying order of the cosmos. Latent Archetypes begins with an existing image and asks what kinds of recurrent structures or effective regularities emerge when a generative model repeatedly reconstructs it.

The learned representational space of an AI is not a Renaissance memory palace, and its statistical representations should not be confused with Giordano Bruno’s metaphysical universe. Yet both Christophe Bruno and Giordano Bruno invite us to think of memory not as passive storage, but as an active space in which images are transformed, connected, and generated.

Biologist Michael Levin uses the expression “Platonic Space” to describe a structured, non-physical space of possible patterns. In his framework, mathematical forms, anatomical structures, behaviors, and perhaps even forms of agency may be understood as patterns that can become embodied in different physical systems (Levin’s Platonic Space is itself a speculative ontological proposal and should not be treated as an established description of biological or computational systems).

Latent Archetypes does not provide evidence for such a space; it creates an experimental situation in which the distinction between generated forms and apparently discovered forms becomes philosophically productive. It resonates with this idea because its images appear to organize around forms that were not explicitly designed by either the artist or the machine during the experiment. These forms can create the impression of existing as possibilities before they become visible, as though the iterative process were discovering rather than inventing them.

However, Levin’s Platonic Space should not be confused with the internal representations of a generative model. A model’s learned representations—sometimes described, depending on the architecture, in terms of a latent space—are technically produced and shaped by its architecture and training data. Platonic Space, by contrast, is proposed as a more general domain of patterns that is not reducible to one particular physical or computational implementation.

Latent Archetypes occupies the tension between these two meanings. Its archetypes are clearly dependent on a specific trained model and are therefore historical, technical, and statistical. Yet their unexpected stabilizations can create the impression that the machine is uncovering forms that somehow precede their individual manifestations.

The project therefore asks whether generative AI merely produces probable images from learned data, or whether it also gives us a new way of imagining a visual space of possible forms—a machinic and historically situated version of Platonic Space.

This is perhaps the most speculative intuition behind the project.

A latent archetype does not seem to pre-exist as a completed image waiting to be uncovered. Each iteration remains partly indeterminate, and several different outcomes are possible. Yet these outcomes are not arbitrary. They are constrained by the source image, the prompt, the trained model, and the recursive protocol.

A physical system may have several indeterminate futures while still possessing objective structures or tendencies that make some outcomes more likely or accessible than others. In this sense, chance does not exclude causality, constraint, or form.

For Latent Archetypes, this suggests a realism of structures and tendencies. Here, “realism” does not refer to realism as a style of painting, although the relation between these two meanings of realism may itself be worth exploring. Nor does it mean that the archetype exists somewhere as a hidden object. It means that the observable transition probabilities may define an effective structure of possibilities. Paths, persistent regions, recurrent configurations, and escape thresholds would then be properties of the dynamics rather than hidden objects contained inside the model.

Some predictive-processing and Free Energy Principle frameworks propose a related idea: perception is constructed through an internal generative model, but this construction remains constrained and corrected by sensory input. In a loosely Kantian sense, what appears is shaped by the conditions through which it appears, without being a pure invention detached from reality.

A latent archetype could therefore be understood as a possible form that has not yet been actualized, but whose emergence is favored by the structure of the system. “Latent” would then acquire a double resonance: a technical one, referring broadly to learned internal representations, and a more speculative one, referring to forms that remain possible without already existing as actual images.

The latent archetype would be real neither as an object nor as an eternal essence, but as a tendency—something that becomes visible through the trajectories it makes possible.

There is a suggestive analogy, although Latent Archetypes should not be understood as a renormalization process in the strict physical sense.

In physics, coarse-graining consists in describing a system at a lower level of resolution by ignoring microscopic distinctions while preserving structures that remain relevant at a larger scale. Renormalization goes further: this operation is repeated across scales, producing a flow in which some features disappear while others persist. Under certain conditions, very different microscopic systems can even flow toward the same large-scale behavior, giving rise to universality classes and fixed points.

Something reminiscent of coarse-graining may occur in Latent Archetypes. At each iteration, an image passes through the model’s process of reconstruction. Features that are difficult to preserve may progressively disappear, while forms, compositions, colors, or visual conventions that the system reconstructs more robustly may survive repeated transformations. The latent archetype could then be understood not as a hidden image already contained in the model, but as a structure that remains relatively stable under repeated reconstruction.

The analogy becomes especially interesting if very different starting images are found to converge toward related metastable configurations. Such a result would evoke the idea of universality: pictorial differences would progressively become irrelevant, while a smaller number of persistent structures would organize the long-term dynamics.

But the comparison has clear limits. The current protocol does not explicitly change spatial scale, integrate out defined degrees of freedom, or implement a renormalization-group transformation. It is therefore more accurate, for now, to speak of progressive information filtering or a coarse-graining-like effect than of renormalization in the strict physical sense.

This raises the possibility that a latent archetype may not be something the machine “contains” in advance, but something that appears through the repeated elimination of differences that its reconstruction process cannot reliably maintain.

From this perspective, Latent Archetypes asks a question close to the logic of renormalization: what survives when transformation is repeated?