Monthly Archive: April 2023

If we keep hearing that AI cannot know more than what it has been trained to know via algorithms, then most people already know that this is not true. The almost incomprehensible number of vectors, which is approaching the trillion mark, and also the almost infinite variety of possible combinations do not correspond to the idea that the system has been given concrete learning content. And yet this cannot be doubted, at least for the image generators of the AI: fed with millions and millions of images and their labels, the generated images do not get out of the model of a complex combinatorics. However, this is so complex that one cannot grasp with human consciousness all factors which have led to the respective combination. Thinking in combinations, however, the assumption of a per se eclectic modus operandi, have a direct effect on perception. Like a counterfeit coin, we hold the offer sideways against the light, as it were, and scan the surface – only we can’t bite on it yet. The result is subject to the examining gaze – at least for the time being. Then, however, an interaction can occur between the object of investigation and the investigator: no cheap copy, no dull imitation, no rigid superimposition is revealed. What emerges instead is a free-acting play of very different materials, modestly associable with sources about whose true existence we know nothing: why does the city of Paris wrap itself in the modification sequence of an old Kodachrome photograph of anglers on the Seine like Christo’s Reichstag? Why do images emerge that resemble illustrations of “Waiting for Godot”? With the dried-up rivulet in the cartoon pastel at the end, is concern about climate change already inherent in AI? You don’t have to be an AI expert to know that all these associations are ultimately nonsense. But why does our imagination not allow to deny an imaginer the imagination? Because then it is already indicated that our imaginations and the AI have no connection, but produce this in the best case only in the cooperation.

Starting with a 1970’s tourist shot of an elderly couple having a break with their limousine in a desert-like landscape, the AI sequencing scans all the motif elements in the sequential episodes: the luxury sedan, the woman in the desert, High Noon, water towers amid deadly drought, until finally, in the emptiness of the desert, first Batman and finally Superman appear. “What was the AI thinking?” is a question that we have long known to be empty. By means of language – desert, woman, man, limousine – associations and variations appear that seem to come from the spectrum of topics, i.e. also from the linguistic: Water, danger, fiction, encounter. As light and clichéd as these “associated” elements initially appear, they continue to develop in a straightforward and at the same time multifaceted manner: vectors point the way. They are vectors of meaning that orient themselves on linguistic meaning and always enter into new constellations and create motifs; they are at the same time formal vectors that orient themselves on colors, proportions, surfaces, and much more. No one can think or know what connections they will finally make in the “deep learning” process, just as someone who gives a direction thereby does not know and cannot know what will appear in the dimension of temporality on the further way in this direction. In this respect, the semi-conscious exploratory course of the user and his careful or also spontaneous micro-decisions are directly reflected in the process of the vector constellations – but show substantially a different quality. It is, so to speak, an analogous process.

It’s fascinating how a single image can undergo so many rapid transformations, resulting in completely different versions each time. The first version with pastel-colored geometric shapes provides a clear image of the character discovering an old tin can on the Bosphorus beach, but it lacks vibrancy, making it somewhat dull. The second version, with a diffuse landscape of shards and debris, uses colors that remind one of 3-D glasses, adding vibrancy and depth to the image.

However, the third iteration takes things to a completely different level by dissolving the original motif and transforming the metal can into a giant tube in a comic style. The Bosphorus now becomes a sunny fantasy landscape, and a figure appears but ultimately fades away into the diffuse background.

With each new generation of pictorial material, the diffuseness takes refuge in surfaces, faces, colors, and diffusely structured crystallization points, eroding any impression of unambiguous meaning. This erosion of meaning is fascinating and irritating in the same time.

Perception is a term that has limitations when applied to AI functions. It refers to the processes of the human brain, whereas “deep learning” imitates neural processes but is not the same as the neural process of the human brain. When “deep learning” AI image generators present image sequences or offer variations, the images are based on algorithms and vectors predetermined by programmers. However, the results can defy their assumptions and correspond to their own machine logic. The interaction between the generated images and human perception can result in a flow that goes beyond visual correspondence and content logic. For example, tracing the modification sequences of an accident event (such as a person falling or jumping off a bridge) in unpredictable ways can reveal the quality of this machine logic as a simulation of perception.

A sequence that begins with the same or similar images as another can take a completely different direction in terms of meaning, solely due to gradually different structure. It becomes difficult to differentiate whether the decisive image contains meaning or simply shows visual patterns.

In a fast-paced selection process, driven by a constant stream of visualizations, novel patterns of meaning no longer arise solely through logical or formal criteria, but rather through a process of negotiation. The user links their reflective process directly with the visualization and concurrently adapts to ongoing suggestions for modification. These modifications facilitate bisociation, similar to Kekule’s famous bisociation process in the invention of the benzene ring. Reflection processes and visualizations intersect, resulting in a unique imaginative quality.

At CHROMALAND STUDIO, artists are using AI-generated images to delve into the interplay between conscious and subconscious perception in flow sequences. They are creating visual series through “deep learning” AI processes and observing the patterns of human cognition that emerge. The AI-based image production is done in parallel with the observation and documentation of these patterns.