Pictorial composition, understood as the arrangement of graphical elements on the plane, is typically associated with qualitative rules and heuristics. Although informative for artists and their practice, these norms and guidelines only act as external constraints on the canvas. I believe that artworks are able to fix more fundamental compositional features in their pictorial matter. I therefore develop a paradigm in which every artwork of an artist represents a partial view of a higher-dimensional representation, aggregating intrinsic compositional regularities. I choose to materialize this theoretical hyper-compositional object by a continuous, vectorial and probabilistic space. In addition, the sequential non-stationary nature of the compositional process, together with the complex and evolving definitions of its underlying functional units, coalesce into a perceptual phenomenon that cannot be readily modeled through pixel-based deep learning models, such as CNNs. I adopt a different strategy, constructed around a parametric definition of stroke execution and hierarchically nested RNN-VAEs. This allows the model to tackle pictorial material by aligning its behavior with the artistic gesture. These artificial neural networks are then trained on more than 5000 personal abstract compositions, and seek to form a compositional space that is compact, cohesive and expressive. Next, I check the perceptual homogeneity of the latent density of this space with experiments of similarity judgment between generated compositions. The resulting algorithms allow me to explore the dynamical interaction of graphical elements in accordance not only with my own compositional regularities, but also with the perceptual regularities of the visitors. Finally, the concept of hyper-composition is intimately associated with the multitude, so any attempt of restitution on paper is confronted by continuity, diversity, and therefore infinity. The dynamics captured by the model are also located in transitions, interpolations and series, that only a digital pen-plotter is able to reproduce.

Surrounding proposals are the result of two-dimensional spherical interpolations carried out with a very high sampling density. The purpose of these mesmerizing mappings is to show the delicate compositional changes captured along the hidden dimensions of the model.

Under the density of a thick stroke, we do not know the internal transition from one border to another. For Transition plots, I operate spherical interpolations between two local samples and draw all intermediary lines. Compressions and expansions within each graphical element produce a diluted ink effect, which can also give the feeling of shallow relief, or enfolding surfaces.

With Trace, I explore a different visual principle based on perceptually corrected interpolations. Precisely, I record the trace left along constant horizontal, vertical or circular shifts. This way, compositional aspects are completely disrupted, but a calm and natural volume appears, evoking the passage of vanished forms, where only dynamics remain.