Garin Curtis

One of several artworks made using the same techniques

This work examines the creative potential of generative AI by exploring its capabilities beyond conventional applications. While most image generative models are trained to replicate patterns from vast datasets using text inputs, this installation invites participants to engage with the system more experimentally and interactively. Through a touch interface, visitors can directly explore the layers of a generative neural network, revealing unexpected behaviours and outcomes.

Using network-bending techniques, which modify and manipulate the internal representation of a stable diffusion model during inference, this work introduces novel interaction with the internal architecture of a deep generative model. The experience prompts visitors to reflect on the assumptions and possibilities of AI, fostering a deeper understanding of its potential role within the creative process.

XY pad on the left and Interactive Map of the internal architecture of the model on the right.

The installation uses network bending techniques – a toolkit developed by Terence Broad – which modify and manipulate the computational graph of a generative model. Originally used for GANs, part of my master’s dissertation project at UAL’s Creative Computing Institute was to implement these techniques into a diffusion model and allow for real-time control over them. The interface I designed for this project acts as a layer of abstraction for these techniques, which often require a deep understanding of the model’s architecture, and lets the user control the model internally using an XY pad. The interface also includes an interactive U-Net map, allowing the user to select different layers in the downsampling block to apply them to.