AI Dreams Interior Without Material Effects

Through the exploration of training a StyleGAN, the ephemeral and dreamlike qualities surfaced in a different way. Again, without features within the images that were fed into and produced by the StyleGAN, the AI script created physical features without context. This method of recursion generated completely unexpected surface treatments that produced an AI-generated space. I began this process by studying existing drawings of Turrell's Ganzfeld. Through the published sketches, images of exhibits, and trial and error in using a physical model, I recreated a very small version of the LED gradient effect seen in Turrell's work. The model was a quick study to understand the color theory and placement of LED lights to create a smooth gradients seen in Ganzfeld. I used insights from the making process to model and further test the interior space in Rhino.

The turning point of the project was creating a recursion loop using a StyleGAN. For a brief overview, a GAN (generative adversarial network) is a type of generative model. Its goal is to take a small piece of information and produce a complex output. Put simply, its a generative model that creates something new. To train a successful GAN there needs to be two sub-models: a generator model that creates something, and a discriminator model that checks the “real-ness” of the created thing. Typically this approach is used to generate a single image given a style reference and a structural reference. However, in AI Dreams the output is reintroduced as secondary input repeatedly - a recursion loop. A later test produced the following sequence showing progressive changes as the model was trained.

Recursion Loop Test | Outputs Used in Digital Surface Tests

Later in the experiment, I tested a hypothesis that a recursion loop within a StyleGAN creates new objects. The test used images of color and no features: cyan as content and magenta as style. I looped the inputs and outputs, and after image 50 the network produced new form. This result drew parallels between Ganzfeld and AI Dreams, both of which "saw" features previously not there.

AI Dreams Interior With Material Effects

The final result is an interior space that blends the Ganzfeld LED lights with material and surface effects produced from StyleGAN explorations. This research has continued into a years long exploration of projects that lie in the intersection of spatial design, artificial intelligence, and human-computer interactions. I hope this continued research speaks to the growing role of AI in our lives and the need for designers and engineers to consider its creative possibilities.