Image Processing: Exposing Latent Space investigates the transformation of images into statistical structures. By applying techniques such as averaging, variance extraction, and Principal Component Analysis to a personal photographic archive, the work operates a shift from photographic traces to data elements within a system of statistical distributions and relations.
The project departs from groups of images and subjects — palms, algae, pebbles, lagoon environments — each generating a distinct series of derived images corresponding to specific statistical operations: the mean (what is common to all images), the variance (where images differ), and the eigenvectors of the covariance matrix, which define the principal directions along which these differences occur. At the same time, eigenvectors can be understood as structures of invariance: they capture what remains stable across variation, revealing the underlying symmetries that organize relations between images.
Together, these operations decompose each corpus into its fundamental components, reducing, separating, and recombining visual information into abstract forms that encode patterns across multiple images.
The eigenvectors define what linear algebra calls an eigenspace — eigen, from the German for “own,” “proper,” or “intrinsic.” An eigenspace is a latent space generated by the internal relations of a dataset: inherently self-referential, it does not describe the world but reorganizes the dataset according to its own internal logic.
This operation resonates with contemporary artificial intelligence. In machine learning systems, images are translated into vectors and organized within high-dimensional latent spaces, where meaning emerges from statistical relations, differences, and correlations rather than from any single image. These spaces are structured through processes of reduction, selection, and normalization that extract stable patterns while discarding others.
Image Processing mirrors this logic through a transparent, linear process, making visible mechanisms that in AI systems often remain opaque. By keeping the construction of the latent space legible, the work exposes processes of reduction, compression, and loss that computational vision typically conceals, revealing the epistemological mechanism of AI.
What emerges is not a single image, but the materialization of a latent structure — a photographic rendering of relations.
As both analysis and a form of resistance, the work exposes how images are transformed into structured data, and how visual culture is increasingly shaped by latent spaces and processes of extraction, reduction, and recombination.
