Urban Resolution

Images · 2023-01 · Houston, TX

In 2018, the UN produced a list of the 81 cities with populations larger than five million people. The scale of these cities is impossible to imagine, even for many of us living in them. Similarly, many AI models have a blindspot when thinking about scale: trained on fairly low resolution images, many models are unable to produce coherent results beyond their trained resolution. In the latent space of these large language models, 1024 x 1024 pixels is the only scale that makes sense. Unlike rendering where larger resolutions are simply a matter of more time, generative AI models simply can’t coherently understand a larger dimension. Using the Stable Diffusion SDXL model set at a generative resolution of 3072 x 3072 with the prompt “A Street in x” where “x” is iteratively replaced with each of the 81 megacities, the model produces dense fields of visual cliches. Tokyo is a wall of advertising, Istanbul is composed of hanging rugs, palm trees float across a Los Angeles blue sky. Any small section of the images makes some sense, but as a whole these urban visions are chaotic and heavily biased colleges. At the same time, they offer architectural densities that hint at possible futures.