Paper Title: The Compositional Gap Between Photography and Artificial Intelligence Images: An Aesthetic Study of the Viewer’s Shifting Gaze
Author: Seokwon Yoon
DOI: https://doi.org/10.47294/KSBDA.27.3.27

Seokwon Yoon’s “The Compositional Gap Between Photography and Artificial Intelligence Images: An Aesthetic Study of the Viewer’s Shifting Gaze” does not distinguish photography from AI-generated imagery solely through the forms visible on the screen. The “compositional gap” discussed in the paper lies deeper than differences in subject placement, perspective, or color arrangement. When viewing a photograph, the viewer’s gaze passes through the depicted subject and moves toward an event in the past. When viewing an AI-generated image, by contrast, the gaze turns toward the data, algorithms, prompts, and selection processes that exist outside the image. The more similar the two images appear, the farther apart the destinations of the viewer’s gaze become.

Photography Never Simply Transferred Reality

The paper begins by dismantling the familiar opposition that “photography is real, whereas AI imagery is fictional.” To do so, Yoon does not begin with AI images but returns to the history of photography. This move is important because, before questioning the authenticity of AI images, it asks when and to what extent photography itself has ever functioned as evidence of reality.

Nineteenth-century portrait photography preserved the actual appearance of its subjects, but it also manufactured class and social status through curtains, costumes, furniture, lighting, and poses. The person in the photograph truly existed, but the studio constructed how that person would be perceived. Photography’s indexicality and its staged quality did not become combined only after the image was taken. They were already operating together before the shutter was released.

John Heartfield’s photomontages combined fragments of different realities to create scenes that had never actually existed. Yet these scenes did not say less about reality; they exposed the political conditions of their time more aggressively. Man Ray’s rayographs likewise produced photographic images without a camera, using only the contact between light and objects. Such examples make it difficult to restrict the identity of photography exclusively to the camera or to realistic representation.

Yoon invokes these cases not to force AI imagery into the category of photography, but to demonstrate that photography itself has always been a site where traces of reality collide with the construction of meaning. There is undoubtedly a physical difference between photography and AI-generated imagery. However, this difference cannot immediately be translated into a boundary between reality and fiction. Photography has never passively transcribed reality, and AI does not extract images from nothing.

Mise-en-Scène Has Moved from the Set into the Algorithm

The paper’s central concept, “algorithmic mise-en-scène,” allows the production of AI images to be understood as a system of direction and staging. Whereas photographic mise-en-scène involved arranging subjects, backgrounds, lighting, lenses, and camera positions, the mise-en-scène of AI imagery consists of data selection, model training, prompt writing, parameter adjustment, and the selection of outputs.

Direction has not disappeared. It has moved from the visible space of the photographic set into the invisible procedures that precede the generation of the image. The photographer organizes the space in front of the camera, whereas the AI image-maker organizes the conditions under which probability operates. In the former, the positions of objects matter. In the latter, what matters is how data and language push the image’s possibilities in particular directions.

Yoon calls this “calculated visual composition.” Here, calculation does not imply a perfectly controlled result. Diffusion models operate probabilistically and generate different results even from identical prompts. Yet these variations do not emerge from an empty field of randomness. The distribution of the training data, the model’s weights, and the connections established between language and image shape in advance the range within which results can appear.

The composition of an AI image therefore does not begin within the frame. Before a figure is placed at the center or light enters from a particular direction, the conditions of composition have already been determined by which images were included in the training data, how often the beauty standards of a particular culture were repeated, and which visual conventions became strongly connected within the model. Composition expands from the arrangement of objects to the arrangement of probabilities.

From this perspective, bias in AI imagery is not an error that happens to appear in the final output. It forms part of the mise-en-scène that determines what appears beautiful, natural, or realistic. If photographic studios once arranged signs of class, generative models arrange the accumulated norms of visual culture around probabilistic centers.

The Viewer Begins to See the Conditions of the Image Rather Than Its Subject

Before a photograph, viewers commonly ask, “What was photographed?” or “When did this happen?” The subject within the photograph becomes a passage leading toward a past outside the image. To borrow Roland Barthes’s formulation, photography pushes the gaze toward the sense that “this has been.”

The direction changes with AI-generated imagery. Even when viewers ask whether the figure in the image actually existed, they cannot find the answer within the image itself. They naturally begin to ask, “How was it made?”, “What was the model trained on?”, and “Which prompts and choices were involved?” The gaze moves away from the represented subject and toward the structure of generation.

The case of Boris Eldagsen, who refused an award after winning at the Sony World Photography Awards with an AI-generated image, clearly illustrates this shift. The work looked like a photograph and evoked emotion like a photograph. Yet once its method of production became known, the appearance of the work remained unchanged while the direction of its interpretation shifted. Viewers no longer stopped at reading the expressions and relationship of the two women. They began to consider what data and calculations had produced those expressions.

The meaning of an image has never been contained exclusively in its appearance. AI imagery, however, makes it difficult to define the very mode of an artwork’s existence without knowing its conditions of production. In the past, production context was additional information used to interpret an image. It has now become essential information for determining what kind of image one is looking at.

Yoon describes this transformation as a shift in the viewer’s role from “recipient of evidence” to “interpreter of the process of construction.” Viewers no longer simply accept the sense of reality presented by an image. They trace what the image selected and excluded in order to produce that sense of reality. Seeing thus expands from recognizing the subject to analyzing its conditions.

Chance Moves from the Photographic Moment into Probabilistic Space

One of the paper’s most intriguing sections reinterprets photographic chance and the punctum within an AI environment. In traditional photography, chance occurs at the site of the shoot. A person who unexpectedly enters the frame, an unforeseen facial expression, or a piece of clothing momentarily caught by the wind remains in the photograph as a trace that exceeds the photographer’s intention. Barthes’s punctum is connected to the experience of being suddenly pierced by such a detail.

AI-generated images also contain forms that exceed the maker’s expectations. Their chance elements, however, are not traces of events that intruded from the physical world. They emerge through the interaction of ambiguous prompts, probabilistic sampling, combinations within latent space, and the judgment of the person selecting the final output. Yoon understands this as an “algorithmic rupture.”

If photographic chance is a remainder left by the world on the camera, chance in AI imagery is a trace of the computational system deviating from human language and intention. Both obstruct complete control, but one enters from outside the photographic apparatus, while the other emerges from within the model. This difference also changes the viewer’s emotional response. An unfamiliar detail in a photograph leaves behind the shock that “this was really there at that moment.” An unfamiliar detail in an AI image provokes an investigation into why the model produced such a form.

The meaning of the decisive moment also changes. Whereas the photographer waited for a moment to capture in reality, the AI image-maker selects one image from numerous generated outputs. The decisive moment is relocated from a singular event passing through time to the moment at which one result is fixed within a set of possible images.

From a Photograph of One Moment to an Image of Overlapping Times

A photograph captures the light of a particular moment. Even when long exposure or compositing is used, photographic time fundamentally retains a relationship with the time of an event. To look at a photograph is to look at the remainder of a moment that has passed. An AI-generated image has no comparable single original time. Instead, it contains overlapping temporalities: the time during which data accumulated, the time of model training, the time when the prompt was written, the time of repeated image generation, and the time when the final result was selected. The screen remains still, but its origin does not converge upon a single moment.

For this reason, an AI-generated image resembles a statistical surface upon which visual conventions from multiple periods have been compressed, rather than a trace of one particular past. A depicted person may never have existed, but the faces, lighting, clothing, and photographic grammar used to construct that person were collected from different times and places. The claim that AI imagery “does not pass through reality” is only half correct. It does not come into direct contact with an individual subject, but it passes indirectly through records of countless real subjects.

Through this temporal structure, Yoon expands photography from a medium that preserves “what has been” into a generative medium that explores “what may be constructed.” The temporal axis of photography thus shifts from the past toward potentiality. The shutter closes around an event that has already occurred and leaves it as an image, whereas the prompt opens the conditions under which a scene that does not yet exist may appear.

Artistic Intention Does Not Disappear but Becomes Distributed

The automatic generation of AI images can create the impression that the artist’s role has diminished. The paper approaches this issue through the concept of “distributed intention.” In photography, intention is relatively concentrated in the choice of camera position, exposure, framing, and the moment of releasing the shutter. In AI imagery, intention is distributed across prompts, reference images, model selection, repeated generation, curation, correction, and exhibition context.

The creative subject is no longer defined primarily as the hand that directly produces the completed image, but as the agent that organizes generative conditions and processes of selection. The algorithm, likewise, is less an independent artist than a collaborative apparatus in which human visual culture has been accumulated as data and reorganized through technical rules.

A paradox emerges here. The production process becomes automated, but the structure of responsibility surrounding the image becomes more complicated. Even when the result cannot be explained by the intention of a single person, the composition of the data, the design of the model, the direction of the prompt, and the decision to adopt a particular output continue to determine the meaning of the image. The distribution of intention across multiple stages does not mean that intention has disappeared. Rather, viewers must examine more carefully what value judgments are concealed at each stage.

Reality Moves from the Image’s Origin to the Event of Interpretation

The paper’s most significant conceptual transition concerns the changing criterion of reality. If the reality of photography derived from physical contact between subject and image, the reality of AI imagery is formed through the perceptual and interpretive effects that occur during viewing. An AI-generated image may have no original event that was photographed. Nevertheless, viewers may experience anxiety, compassion, déjà vu, or a historical atmosphere through the image. Such emotions do not immediately disappear once the viewer knows that the image is fictional. The persistence of a sense of reality without a physical origin makes it difficult to explain reality exclusively as a property of the trace.

Yoon regards reality not as a certificate of authenticity fixed within the image, but as an event that occurs when image, viewer, and social context meet. From this perspective, the “recovery of reality” does not mean restoring the evidentiary authority that photography possessed in the past. It means identifying what operates as real, who is led to believe it, and what social effects that belief produces.

The question also changes. “Is this image real?” becomes “What kind of reality does this image set into motion?” The former examines the image’s origin, whereas the latter traces how the image organizes the viewer’s emotions and social judgments. The politics and ethics of AI imagery emerge precisely within this process of operation.

The Expansion of the Question of Photography, Rather Than Its End

In this paper, the shutter and the prompt are not markers dividing an old medium from a new one. They are two apparatuses that organize reality in different ways. The shutter cuts a moment out of the world, while the prompt establishes the conditions under which certain possibilities may emerge from a latent visual field. Photography selects traces; AI imagery selects possibilities.

This does not mean that the differences between the two media disappear. Photography involves physical contact between a particular subject and light, while AI-generated imagery lacks such contact. The achievement of the paper lies not in denying this distinction, but in preventing a physical difference from being immediately transformed into an aesthetic hierarchy or an absolute boundary between reality and fiction.

Viewing photography and AI imagery along a continuum also increases the responsibility of the viewer. Images no longer prove by themselves where they came from. Viewers must read their processes of production, the genealogies of their data, their structures of circulation, and the power relations surrounding them. The gaze no longer remains on the surface of the screen but moves toward the system that made the screen possible.

Ultimately, what AI changes is less the outward appearance of photography than the way photography is believed and interpreted. Photography once gained its authority as a trace left by the world. Images now encompass not only traces of the world, but also the possibilities through which a world may be constructed, the technologies that calculate those possibilities, and the perceptions through which viewers accept them as real. As the boundary of photography becomes blurred, what disappears is not photography itself, but the long-standing belief that a photograph can guarantee its own truth.