“This book is the culmination of my last three years of thinking and experimenting with generative AI technology, and especially last year’s sudden acceleration. The images contained are entirely generated with tools accessible to the general public and question the nature of photography and art.
In the future, synthetic images will be fed back into the datasets of the tools that created them, creating a feedback loop where clichés, stereotypes and biases replicate and expand until possible collapse. This book is organized in three chapters / metaphors (The Museum, The Feedback Loop, The Black Hole) to help us think and plan what will become of images and media when trapped in a never ending cycle of nostalgia and self-reference.”
“Don’t you ever wonder where all those images come from? When you type a prompt into one of the text-to-image generators like Midjourney, DALL-E, or Stable Diffusion, and four images appear within seconds, seemingly out of nowhere, don’t you assume that there must be a place where these images are just waiting to be realized, albeit in a preliminary, latent form? Technically, it might not be wrong to say that these images are created from pure noise, gradually shifted, refined, and transformed in a direction that best fits your prompt. But we all know that’s only half the truth. How images and text relate, and thus which visual output best matches a given prompt, is something these models have learned from billions of image-text pairs scraped from all over the web. But the images these models present to you are not mere copies of the images they were trained on, and probably more than just collages, mash-ups, or remixes assembled from recognizable elements of pre-existing content, though some will disagree. Rather, they might be better described as “pattern interpolations” – lingering in the space between the images on which the system was trained, as latent possibilities ultimately derived from patterns extracted from images of the past. AI im-ages, whatever they pretend to show, are ultimately images of images, perhaps even images about images. AI image generation, which we might more accurately call image synthesis, thus depends entirely on the cultural archive: the range of images fed to these models defines the range of their possibilities, which, while seemingly infinite, are in fact limited, especially since in this case the archive, collected from the Internet and filtered for commercial purposes, has a strong Western bias, among other distortions. What is now marketed as “generative AI,” especially in its commercial variants, thus embodies an extractivist view of the cultural archive: every image ever produced, it seems, counts only as a resource to be mined, a source of patterns to be exploited, mere raw material for the endless production of new variants. At the same time, AI image synthesis, as a form of pattern interpolation, is almost nostalgic by design: the past is not only the source from which these im-ages originate, but also the place to which they ultimately seem to return. Each new image, synthesized from the patterns of the past, manifests itself from a mere possibility waiting in-between already existing images and seems to fill a kind of gap in the cul-tural archive – an image that could have been, but never was, and only now becomes visible. But any image that becomes visible through these process-es always remains a mere statistical possibility. It could look different, and indeed a simple command is all that is needed to create another version, another instance in an infinite series of variations that are almost, but never quite, the same: settings are rearranged, faces and poses change, props appear and disappear, and indistinguishable details become sharply defined objects – or vice versa. Underlying AI image synthesis is a fundamental contingency that is the exact opposite of the contingency of the photographic act. Photography is always confronted with the contingency of the world we live in: a situation can drastically shift from one moment to the next, the lighting can change, peo-ple can walk into the frame, or any other unplanned and unfore-seen event can occur. In the photographic act, however, this con-tingency of the world is arrested in an image: the ever changing flow of events comes to a standstill. AI image synthesis, on the other hand, does not give you a frozen image of an ever-changing world, but a statistical snapshot of big data. Every image it pro-duces is just one contingent variation in a never-ending series of statistical probabilities. Still, sometimes we can’t help but look at these images as if they were photographs, and wonder where these synthetic faces came from, what living bodies must have been photographed by a camera in order for us to see their ghostly traces, however distorted, in these images. Somewhere in the training data that feeds these models are photographs of real people, real places, and real events that have somehow, if only statistically, found their way into the image we are looking at. Historical reality is fundamentally absent from these images, but it haunts them nonetheless. In this respect, they are not only nostalgic because they evoke a past that never existed. They can also make us nostalgic for the past they hide from us. The most intriguing images evoked by AI may be those that somehow rec-oncile these two opposing feelings of latent nostalgia.”