Gallery Entry

Erosion

By Valie Messini · AT

*Erosion* began from an unease with the way algorithmically mediated image cultures, particularly social media, produce increasingly recognisable forms of self-presentation. Subjects become visible through recurring personas—the sexy one, the sporty one, the mother, the intellectual—each accompanied by familiar poses, expressions, framings and gestures. What interested me was not only the homogenisation of self-presentation, but also the fragmentation of subjecthood into multiple, already recognisable forms. Rather than presenting a continuous and contradictory mixture of traits, the self shifts between discrete scripts. Fragmentation and homogenisation are therefore not opposites: the subject fragments into multiple personas, but the personas themselves are standardised. I understand this as a form of erosion: a gradual disappearance of particularity through repeated translation into established forms of visibility and legibility. In the work, erosion operates in two related senses. The first concerns subjectivity: what happens when subjects are repeatedly translated into forms that computational systems can recognise? The second is computational: what transformations occur as an image passes through successive operations? Both concern a pressure towards legibility. Technically, the work was produced with a pix2pix generative adversarial network trained on a dataset of synthetic faces from *This Person Does Not Exist* and tessellated versions of those images. I procedurally constructed each pair by transforming a face into a geometrically reduced representation; the model then learned to translate this reduced input back towards a facial image. Training across many synthetic faces echoed the fragmentation at the centre of the work: instead of beginning with one stable subject, the model learned from a multiplicity of already artificial faces that had themselves emerged through an earlier computational process. I deliberately stopped training after 100 epochs, preserving an unstable state of translation rather than treating technical convergence as the goal. For the video, I confronted this model with reduced representations derived from a selfie video of one specific face: my own. Contours and tessellations became inputs through which the model produced transformed faces. Variations in these inputs modulated the transformation: thicker contour lines generated increasingly abstract results, while tessellated inputs retained more spatial information and enabled more precise reconstruction. As the inputs departed from the structure used during training, the outputs became unstable. The work therefore tests how far a learned mapping can be stretched before recognition begins to break down. I subsequently reworked the generated video through levels, clarify, cartoon, chromatic and glowing-edge effects and overlaid it with the original video, processed through the same sequence. The apparent distance of my face from the camera controls the balance between the two layers. At a greater distance, the generated image dominates; as the face approaches, the original increasingly leaks through, although the transformation never disappears completely. The tension between erosion and legibility thus becomes spatial. As my face recedes, it dissolves into an abstract, roughly painted figure. As it moves closer, it becomes more recognisable. The movement suggests a return towards individual subjecthood, but one that cannot be completed: the face approaches the surface yet remains inside the image. The screen becomes the boundary of that return. Even at the point of greatest recognisability, the subject remains mediated.

Category

VIDEO

Approved

11 October 2026