RESEARCH PROJECT
| 2023 · TU/e
AI Style Agent
Human–AI Design Exploration
Helping designers teach, inspect, and refine a personal style space through visual references and semantic language.
1,290
projected posters
Positioned in the style space using model-predicted semantic values.
30
semantic dimensions
Used to describe subjective visual preferences.
73.9/100
avg. Creativity Support Index
Positive perceived support in an exploratory study (n=6).
01
Overview
Turning subjective style preferences into something designers can teach and navigate
Designers often recognise a visual style before they can clearly describe it. AI Style Agent explored whether subjective aesthetic preferences could be externalised into an inspectable style space that designers could teach, navigate, and refine.
Rather than asking AI to generate a finished outcome, I explored a different interaction model: designers provide references and semantic language, the system maps those inputs into a visual representation, and the designer remains responsible for interpretation and correction.
Role
Lead Designer & Researcher
Project
TU/e Industrial Design research project
Scope
Interaction model, semantic framework, data visualisation, interface design, prototyping, study design, analysis
Methods
Interactive machine teaching, semantic differentials, PCA, prototyping, mixed-method evaluation
Project Status
Exploratory research prototype
Duration
February–June 2023
02
Design Opportunity
Designers feel style before they can name it
Traditional inspiration tools assume that designers can already express intent through keywords, categories, or similarity. But aesthetic preference is often recognised before it can be verbalised.
The opportunity was therefore not simply to improve image retrieval. It was to help designers construct a usable personal language for style by connecting abstract impressions with references they could inspect, adjust, and reuse.
03
Model & Interaction Logic
From rated examples to a larger navigable style space
During dataset construction, 156 posters were rated across 30 semantic dimensions, creating the labelled data used to train the style model.
The trained model was then used to predict semantic values for a larger corpus of 1,290 poster references, allowing these previously unrated examples to be projected into the same navigable style space.
This allowed the system to move beyond the original rated examples while preserving the semantic structure learned from human input.
04
Design Evolution
From research visualisation to a designer-facing workflow
Early prototypes helped test spatial navigation, image selection, and nearest-neighbour relationships at different scales. They demonstrated that visual style could be represented spatially, but also revealed that research-oriented tools exposed too much technical structure and too little task meaning.
These findings shifted the design toward a workflow organised around teaching, exploring, and correcting the system.
Research Prototype Based on CAKE
Validated spatial browsing and image selection, but revealed that the controls reflected the research system rather than the designer’s task.
TensorBoard Exploration of 1,290 Posters
Confirmed that larger poster collections could be explored spatially, while revealing the need to explain why particular images were positioned together.
05
Final Experience
A teachable style space for designers
The final prototype translated technical style mapping into a designer-facing experience. Designers could teach the system through references, define meaningful semantic axes, explore related work, and revise the model when its interpretation did not match their judgement.
The interface behaved less like a search engine and more like an editable map of aesthetic intent.
06
Key Interactions
How designers teach and steer the system
07
Exploratory Evaluation
Evaluating AI-supported style sensemaking
I conducted a small mixed-method evaluation to examine whether the interaction model could support style articulation and applied creative work. Six designers used either a personalised model trained on their own references or an average model based on collective data.
The study was exploratory. Its purpose was to identify interaction value and future research questions—not to establish statistically conclusive performance differences.
08
Outcome & Implications
Designing AI to extend judgement rather than replace it
The project suggests that creative AI can be valuable not only by generating outcomes, but by helping designers externalise, inspect, and refine subjective judgement.
The exploratory evaluation supported the interaction model’s potential for style sensemaking and creative support, while also showing that stronger claims about personalisation require longer-term validation.
01 — Make model behaviour inspectable
Designers do not need visibility into every technical operation, but they need enough transparency to understand relationships, recognise mismatches, and intervene when the system’s interpretation does not align with their own.
02 — Treat designer agency as product value
The system should expand the range of references and relationships designers can explore while keeping interpretation, correction, and final judgement in human hands.
03 — Validate personalisation over time
A single-session study can evaluate usability and immediate sensemaking, but longer-term use is needed to understand whether a personalised style model remains coherent, trustworthy, and creatively useful.
The strongest opportunity is therefore not AI that defines a designer’s style, but AI that helps designers examine and evolve it.
