top of page

Latent Letters

Continuing my study of human-AI collaboration in creative spaces, I found an interest in how building AI models can become a medium for artistic expression. At the heart of all artificial intelligence systems lies data. As an artist, I am in control of what data I want to link together and how it can share the same algorithmic space. I design the thinking space, or rather the latent space, of my model through the datasets I gather, clean, and construct. I design how that information gets associated inside of the AI and how I want that model to respond to a given input. Latent Letters is my first exercise in using data as a medium and AI as a glue to drive unique, interactive experiences.

Design Breakdown

The Experience

Latent Letters is a study of collaboration between artist, participant, and AI. I build the space and decide what it can hold, the model finds the connections within it, and the participant’s words set those connections in motion. None of us could create the result alone, and none of us fully controls it. So when the colors fill the wall, whose answer is it?

​

In this installation, I invite my viewer to sit down at a desk and type a letter to their Past, Present, or Future self. I decided on this prompt because writing to any version of yourself gives you a moment to reflect on where you are in life. It guides the participant to use casual, non descriptive language and to create a personal connection to the visuals. When they press “Enter,” their letter is sent through my AI model, where their words are weighed on how energetic and pleasant they are. Those weights are then matched to color palettes made from photos of real flower petals. I chose to draw my color from flowers to ground and limit my model to something natural. Every color my participant sees was pulled from nature, not generated. The visual composition becomes a lava lamp of moving bubbles of color and projected it on the wall in front of the participant. The interaction ends with their letter transforming into shifting color-language possibilities derived from the color of flowers and the energy and pleasantness of the participant’s letter.

​

Through early testing of the system, I noticed a common confusion: “Why is it this color?” Because I chose to use chroma and not hue to define the associations, there was a natural instinct to question the result. This showed that the system was working; it was showing the moment people stop accepting AI responses and start interrogating them. This is a key part of the system’s design. I am asking viewers to be open to new interpretations of color-language association. They will question it, and it may not be what they expect, and that is okay. But I did want to offer insight into why the result mapped the way it did. I implemented two features. The first, on an iPad, visualizes the latent space of my model, showing a physical mapping of the color. The second goes deeper into how the words themselves are weighted and how that affected the result. Neither of these is key to the interaction; rather, they offer a moment to learn more about the inner workings of the system.

​

​

Installation 
Barnes Ogden Gallery - Louisiana State University
iPad display
ScreenRecording_11-05-2026 09-18-48_1.mov
00:00/00:00
Phone interface
computer vis test.mp4
00:00/00:00
phone interface.mp4
00:00/00:00
The Color

The color comes from two datasets, one that I sourced online ( the Oxford 102 Flower Dataset), and one I constructed from my own phots, and photos given to me from my family, friends, and classmates. I chose flowers as the color source to ground and limit my model to something natural. Each image was segmented into the flower petals and organized into palettes.  

Custom flower dataset

Custom flower dataset

00:00/00:00
The Language

To ensure my model could cover the wide range of linguisitc inputs I used three language datasets: Empathetic Dialogues, EmoBank, and GoEmotions. Together these datasets cover Blog posts, news articles, online forum posts, and crowd sourced sentences. 

Cleaning and Scoring

all datasets were cleaned using a mix of manual and script based methods. I then utilized the NRC VAD Lexicon to score the data according to Valence(pleasantness), Arousal(intensity), and Dominance(control). The shared scores enable the AI to draw connections between words and the color palettes.

Vectorized flower palettes with scores

After seeing matches between the palettes and the language with simple score matches I saw how the color language associations will differ from typical Western association. For example in one test I input the word "furious" and the palette that was retrieved was filled with very deep purples instead of reds or oranges. This is because of the scoring method. Furious has a low pleasure rating but a high arousal rating. That places it in the cooler color palette range and associates it with heavy saturation.

The Model

The model is a small contrastive CLIP model. It doesn't generate anything, rather it matches the viewers words to the closest colors in its shared space. This shared space is called Latent Space.

First Full Model Interaction

First Full Model Interaction

00:00/00:00
Why Latent Space?

Usually, an interaction with AI is very straightforward. You prompt it, and you are given your answer. I wanted to experement with latent space to visualize the possibilities living inside of my model. The model pulls 12 palettes from the area near its "answer" and sends them to TouchDesigner where they become a shifting composition of light and color.

Model Interaction with TouchDesigner

Model Interaction with TouchDesigner

00:00/00:00
TouchDesigner

Primary node-based visual programming environment for real-time interactive playback.

Claude Code
NRC VAD Lexicon
Oxford 102 Flowers Dataset

Used for collaborative coding

bottom of page