2026
generative video
Dirty Window continues untitled [ ], working with Google Street View. The sites are places held in memory. A memory names a place, and that place has been photographed for years by a machine that was not there for the memory and knows nothing of it. I train an AI on that record to look back onto the site that trained it.
The model is a generative adversarial network, trained from scratch on a closed set of Street View images from one memory-site, so that it has no visual reference outside that site’s own record. I use Street View because it is already there: pervasive, held on the cloud, covering nearly every place, returned to and re-photographed over years. From it I gather the site’s images and approve them one by one, deciding which belong to the place as I hold it, and what I keep becomes the entire training set.
Through training, the model learns the site from those images alone. It does not keep them; it holds what recurs across them: its light, its angles, its weather, the surfaces the camera passed, the way that site appears across its own accumulated images. What forms is a texture of what the site is like, saturated with the concurrency of times the shared camera has accumulated.
Source images are refracted through this texture, appearing only through the pattern the site itself has produced. The site’s detail and the camera’s flaws are learned together, so the blurs, seams, and repetitions of automated capture return as part of the place itself. The window is dirty because it is made of the site’s own residue.
still from generative video
