Caryatids

An evolving geometric composition driven by a figure arrayed through space and time. The work draws inspiration from the early chronophotography studies of Étienne-Jules Marey and Eadweard Muybridge and the sculptural abstraction of the Russian constructivists. The piece is mediated by a bespoke machine learning model that transforms the movement of human silhouettes into an undulating architecture, the ‘caryatids’ supporting it fleetingly recognisable in the negative space.

AI/machine learning is used as a tool of transformation in this work where its output is combined with other computer graphics techniques to create the final work. The starting point though is very human – the silhouette and movement of a single dancer drives the composition. Abstraction is introduced by feeding an array of these silhouettes, extracted over time, into a custom neural network (pix2pix) which transforms the silhouettes into an evolving three-dimensional structure of rectangular shapes. A bespoke dataset dictates this ‘shape language’. It was carefully developed to ‘teach’ the machine how to translate round, organic surfaces into rectilinear shapes, enforcing a sort of AI-mediated rectangular abstraction. With a well-trained model, the ‘glitch’ is mostly held at bay – evident only as small noise in the geometry at the top, bottom and middle of the frame. However at certain points, glitch overwhelms the composition when the machine learning model is asked to translate too much. At these points the model fails and the output degrades noisily or breaks completely. The final sequence, though, still handles this gracefully as the degraded outputs continue to be rendered with the same lighting and shading quality as the rest of the sequence. The final piece retains a coherent aesthetic quality with the glitch adding intermittent compositional uncertainty and interest to the work.