Garin Curtis

Beyond Prompts Installation

Artist Statement

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 in a more experimental and interactive way. Through a touch interface, visitors can directly explore the layers of a generative neural network, revealing unexpected behaviours and outcomes.

By making AI's underlying processes accessible, the installation aims to demystify how these systems operate and generate their outputs. It offers a hands-on way to think critically about the mechanics of AI, moving beyond its role as a tool for replication to highlight its potential as a medium for exploration and discovery.

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.

Gallery

The installation was showcased at the UAL Creative Computing Institute’s Winterfestival in December 2024 where it won the industry prize for storytelling.

About

Welcome to Beyond Prompts, my interactive installation exploring the creative potential of generative AI. Developed as part of my master’s thesis at the Creative Computing Institute, University of the Arts London, this project aims to advance real-time, artist-driven interaction with generative AI models. It provides an open-ended platform for exploring the latent space of generative AI, enabling users to exercise more expressive and artistic control over generated outputs.

At the heart of this installation are network-bending techniques, originally pioneered by Terence Broad, which I applied in a novel way to the Stream Diffusion pipeline. These techniques involve applying mathematical transformations to the image tensor within the model’s computational graph during inference, allowing for dynamic and expressive alterations to the generated output.

In the accompanying video, the original prompt is “a portrait of a person.” However, as the user applies and manipulates the network-bending techniques via the tablet, the output evolves into abstract and unexpected visuals. This interactive process provides a hands-on opportunity to explore how AI interprets and generates imagery.

Interactive Features:

Touch Interface:
Users interact with the generative AI by moving their finger across a 2D space displayed on the tablet. This movement is mapped to ten simultaneous transformations applied to the image tensor using another machine-learning algorithm, ensuring smooth and responsive real-time interaction.
Layer Selection:
Users can choose which layers of the neural network the transformations are applied to via controls on the right-hand side of the tablet. This offers deeper customization and a unique way to engage with the neural network’s internal structure.

Technical Setup:

For the installation, I configured the system to perform image-to-image generation, using animated noise as the input. The entire system runs on a plugin I developed for TouchDesigner, enabling seamless integration with real-time visualization workflows. All processes are executed on my laptop, making the installation compact and portable.

Interface Overview:

On the left side of the tablet interface is the XY grid, which allows users to manipulate the transformations applied to the neural network.
On the right side, users can select specific neural network layers to apply the transformations. Each layer corresponds to a distinct stage in the denoising process during image generation, with each stage uniquely shaping the final output and its visual characteristics.

TouchDesigner Plugin

Audiovisual

This audiovisual was created using the same tool as above but in a wide screen format and upscaled to 4K as an example of how the tool could be used in a creative application. All the audio was written and performed by me.

Artworks

A collection of Artworks made using just the tool and upscaled to 1024 x 1024.

Contact

Email: garincurtis@gmail.com
phone: 07748655773