The piece presents 5 minutes 35 seconds of neural network-rendered video, at 6k resolution, exploring image_histories--as--continuous_spaces. It begins sampling from Chicago art & design history (O'keeffe, Paschke, Marshall, etc) and ends with internet vernacular (cats, cheeseburgers, celebrity faces). The hope is to synthesize incredible work coming out of AI research into an independent, standalone artwork requiring little explanation.
This projection-mapped video (with audio track) was commissioned by the Terra Foundation for the inauguration of a public art program called Art on theMART. It was projected on the Merchandise Mart's 2.5 acre south wall in downtown Chicago from Sept 29, 2018 to Dec 31, 2018.
Tech notes: Almost the entirety of the piece was rendered by an in-house created neural network architecture. We combine a progressively-trained GAN with a high resolution autoencoder. The autoencoder provides a Z-vector prior from its bottleneck which is also used as a source of reconstruction loss. We call our network yGAN. Additionally, we built a versatile system (called Genmo) to decompose input video/animation into varied tilings, render those tiles into a new domain with our trained models, and recompose them. In this work, all rendered imagery is derived from carefully determined input video & animation. We ultimately trained 17 distinct models to render it.