Simulacrum Orbita is a never-ending feedback loop between two generative machine learning (AI) models: one generates images, while the other describes them. Like the game “telephone”, a message is passed along between them, becoming slowly distorted over time.
The loop begins with a generated image – such as ‘a red car’. This newly created image is passed along to the machine learning program able to ‘see’ and describe images. Rather than accurately describing that image as a 'red car’, however, this model might describe it as ‘a red building’ instead, perhaps because it mistook the car windows as those of a building. This new description is then used as the prompt for the next image, and so on. In this example, the red building might be described as a red mailbox, and then a fire hydrant, and then perhaps, a red sunset.
.png)
.png)
.png)
.png)
Overtime, random deviations, either in the generated images or in their generated descriptions, transform the content, nature, and trajectory of the image sequence. Upon each subsequent update, each monitor displays, from left to right, a new image created from the generated description of the image before it. Sometimes, the imagery shifts rapidly between different objects, people, and places as they struggle to communicate and understand one another. Other times, the models find themselves lost in forest paths or city streets, or in fields of grass.
.png)
.png)
.png)
.png)
Simulacrum Orbita is powered by the captioning machine learning model BLIP and CC12M V Objective Diffusion.
_____________________________________________
Example run
Below are the images from a one hour uninterrupted run. Zoom in to see the captions, or zoom out to see the overall evolution of imagery!
Also called 'broken telephone' or 'chinese whispers,' although this term connotes non-western language and culture to ideas of confusion and incomprehensibility.
Process
The two models are assembled in a python program -- first, the image generator model creates an image, this image is saved locally, and then passed to the image describer model, which generates a description of the image. The loop then restarts, using this generated image as the prompt for the next image.
the images are displayed using pythons tkinter module.
the monitor unit was cut out using a cnc machine, and assembled with screws (big thank you to Matty and the technicians at woodworking for guiding me through the process!)
