Artistic research on Recursive Dynamics in Generative Image Models

In the rapidly evolving landscape of diffusion-based artificial intelligence (AI), feedback processes within generative models like Stable Diffusion (SD) are becoming increasingly significant. This research examines the phenomenon of “self-cannibalizing”-feedback loops, where the recursive use of model outputs as subsequent inputs for the model leads to unique emergent behaviors, biases and hints at a potential model collapse. 

The focus is on the Stable Diffusion pipelines, examining how feedback causes images to transform into limited patterns and concepts after only a few iterations. Instead of focusing on the impact of training data, we explore the algorithmic contributions of various components within the diffusion pipeline. The inner workings of diffusion models embody cybernetic principles, with feedback loops and self-regulating mechanisms shaping generative outcomes. This recursive process reveals the model’s tendencies and uncovers potential structures. We investigate the aesthetic and medium specificity within diffusion model-generated imagery, offering insights into its unique visual language and underlying computational frameworks.

Dates of Event

19^th September 2024 – 20^th September 2024

Poster Talk by Ting-Chun Liu & Leon-Etienne Kühr @ TEDA’24: The Magic Machine, Through the Prism of Art and Science, University of Cambridge