Prompting for products: Investigating design space exploration strategies for text-to-image generative models

Leah Chong, I-Ping Lo, Jude Rayan, Steven Dow, Faez Ahmed, Ioanna Lykourentzou

Introduction

Generative AI presents a significant opportunity for product designers, yet it remains a "double-edged sword." While text-to-image (T2I) models like Stable Diffusion can rapidly generate hundreds of concepts, many results are "hallucinated"—physically impossible to manufacture (infeasible) or derivative (lacking novelty). This research addresses the "semantic gap" where T2I models lack an inherent understanding of engineering constraints. By empirically testing Design Space Exploration (DSE) strategies, this study identifies how designers can move beyond trial-and-error to produce designs that are simultaneously feasible, novel, and aesthetically pleasing.

Methodology

The researchers conducted a controlled experiment with 15 participants, ranging from novices to experienced AI users. The task was to design a bicycle that met three criteria: feasibility, novelty, and aesthetics, within a 30-minute timeframe.

Participants utilized Leonardo.AI, an interface for Stable Diffusion 1.5, which supported two distinct workflows:

  1. Global Editing: Generating entire images from scratch using text prompts.
  2. Local Editing: Refining specific image components through masking, erasing, and targeted "in-painting" prompts.
Overview of Global and Local editing modes in the design tool, showing the transition from initial generation to targeted refinement.

The study analyzed prompt characteristics, including length and goal orientation (mono-criteria vs. multi-criteria). The resulting 18 bike designs were evaluated by crowdsourced raters using a 5-point Likert scale across the three target metrics.

Key Insights: The Global-to-Local Pipeline

A primary finding was the consistent "diverge then converge" pattern adopted by users. Participants typically spent 38.5% of their time on global exploration before "locking in" a base design for local refinement, which occupied the remaining 61.5% of the session. Notably, once users transitioned to local editing, they rarely returned to global generation, suggesting a linear refinement process.

The study also revealed that prompt structure is more critical than time spent or prompt length. While global prompts were significantly more complex, local prompts were concise and targeted.

Average prompt length (number of words) for Global vs. Local editing modes, highlighting the shift toward concise, targeted input during refinement.

Results and Strategic Playbook

The analysis identified specific strategies that correlated with success in different phases:

  • Global Mode Success: Multi-criteria prompts (e.g., "a feasible and modern aerodynamic bike") were significantly more effective during the initial global phase for achieving feasibility (rho=0.6, p=0.02).
  • Local Mode Success: During refinement, mono-criteria prompts focusing on a single attribute (e.g., "sleek carbon fiber frame") were more effective, particularly for enhancing aesthetic ratings.
  • The Aesthetic-Usability Synergy: A strong positive correlation was found between feasibility and aesthetics. When local editing successfully targeted aesthetics, the designs were often perceived as more feasible, suggesting that "cleaner" visual outputs are interpreted as more manufacturable.
  • The Novelty Penalty: Novelty proved to be the most challenging goal. In global editing, prompts targeting novelty were negatively correlated with aesthetics (rho=-0.6, p=0.02), indicating a trade-off where "unique" features often compromised visual appeal in current T2I models.

Conclusion and Future Work

This research demonstrates that strategic prompting—moving from multi-criteria global prompts to mono-criteria local refinements—significantly improves design outcomes. Future work will explore "few-shot" prompting to overcome the novelty penalty and investigate the integration of 2D T2I workflows with 3D modeling tools to bridge the gap between conceptual imagery and physical manufacturing.

Summary updated

BibTeX
@inproceedings{chong2025prompting,
  author = {Leah Chong and I-Ping Lo and Jude Rayan and Steven Dow and Faez Ahmed and Ioanna Lykourentzou},
  title = {{Prompting for products: Investigating design space exploration strategies for text-to-image generative models}},
  booktitle = {Design},
  year = {2025},
  doi = {10.1017/dsj.2024.51},
}
Figure