Prompt Your Person

Prompt Your Person

AI · Product · Cross-functional Leadership · Communication

Published 2024

Role

Project Lead

Team

Nadia Piet (Creative Lead, Researcher); Fred Wordie (Creative Technologist); Margarita Osipian(The Hmm, Researcher, Exhibition partner) and Lilian Stolk (The Hmm, Researcher, Exhibition partner)

Overview

How would you label someone you love? Prompt Your Person is a generative web app that challenges visitors to write a prompt that would "accurately" generate an image of their favourite person. Do you have the words to describe the people you love? Does this generative image maker see your loved one the way you do?

Made with The Hmm for MU Hybrid Art House, Poetics of Prompting exhibition shown during Dutch Design Week 2024.

Approach

1/ RESEARCH

We used an insights-led approach to inform our hands-on research. In each of three iterative rounds, the research team – Ploipailin Flynn and Nadia Piet from AIxD, and Margarita Osipian and Lilian Stolk from The Hmm – tested hypotheses amongst ourselves. We tested hypotheses with an additional 7 people who had a novice’s understanding of the generative AI. All experiments were conducted using stable-diffusion-xl-1024-v1-0.

2/ DESIGN + BUILD

With the support of The Hmm, the AIxD team led the design and build of the web application. For full details, see outputs for hyperlinks to the web application tech stack and open-source code.

3/ EXHIBITION

Exhibition: Prompting People, Who Are We in the Eyes of the Machine? was exhibited at MU 4 October – 24 November 2024.

Outcomes

Prompt Your Person, web application here

Github link here

Project introduction via Instagram Reel

Research reflections published publicly here

Insights

EFFECTIVE INTRO TO GENERATIVE AI + IT’S LIMITATIONS

The interactive experiential installation proved to be a surprisingly effective tool for fostering critical AI and media literacy.

The strong public response exceeded expectations, with enthusiastic queues indicating that people were eager to understand and experiment with generative AI—many of whom had little prior experience with these tools.

EFFECTIVE “ENTRANCE TICKET” FOR VISITORS

Asking visitors to "describe the people they loved" resonated well, creating an emotional and relatable entry point into the web app. It demonstrated that the choice of their words effected the resulting generated image.

WHAT CONSTITUTES AN “ACCURATE” IMAGE?

When participants prioritized emotionally accurate descriptors over physical ones, the generated images often captured the “vibe” of their muse, even if the physical resemblance was minimal. This sparked an important question: what matters more in AI-generated images of loved ones—emotional accuracy or physical likeness?

WHAT ARE THE PHYSICAL INDICATORS OF “VIBE”

Related, if participants said that their emotionally-laced prompts generated images with the “vibe” of their loved ones, what did they recognise in the image as the “vibe”? What details – big or small, seen consciously or unconsciously – do we associate with different vibes?

THE MODEL BECAME HYPER-FOCUSED ON SPECIFIC DESCRIPTORS

Another unexpected finding was the disproportionate influence of specific words on image outcomes. For instance, descriptors like "Asian woman" consistently produced hyper-sexualized images, regardless of additional context like age or relationship to the prompter, while "Black man with dyed hair" repeatedly generated a singular image of a heavily tattooed and pierced individual.

What is the library of terms that “trigger” this model into certain latent spaces? "Asian woman" was one; "Black man with dyed hair" was another.

Once the model is stuck in one latent space, can we force it back out? Tips & tricks?

Impact

Prompt your Person was exhibited at the Poetics of Prompting exhibition at MU Hybrid Art House in Eindhoven, 4 October – 24 November 2024. You can see the project featured in this video of the exhibition’s opening night here.