Abstract
Graphic designers explore large stock image collections during open-ended or early-stage design tasks, yet common tools emphasize relevance and similarity, limiting designers' ability to overview the design space or discover visual patterns. We present an image exploration prototype that enables stepwise adjustment of diversity, allowing users to transition from diverse overviews to increasingly focused subsets during exploration. Our approach implements diversity control via determinantal point process (DPP)-based sampling and exposes diversity-similarity tradeoffs through interaction rather than static ranking. We report findings from a pilot study with professional graphic designers comparing our technique to baselines inspired by current tools in open-ended image selection tasks. Results suggest that stepwise diversity control supports early-stage sensemaking and comparison of visual patterns, while revealing important tradeoffs: diversity aids discovery and reduces backtracking, but becomes less desirable as exploration progresses. We aim to provide a novel perspective on how to implement transitions between diversity and similarity.
Introduction
When graphic designers initiate a project, their creative goals are often "underspecified." They require a broad exploration of visual possibilities before committing to a specific aesthetic direction. However, current commercial stock image platforms, such as Adobe Stock or Pinterest, typically prioritize relevance and similarity. These tools often present results in infinite scrolling grids that can lead to "creative fixation"—a state where designers become stuck on a single concept because they are not exposed to sufficiently diverse alternative styles.
There exists a fundamental tension in the design process: high diversity is essential for early-stage brainstorming and orientation, while visual similarity is required for the final selection and fine-tuning. DiverXplorer addresses this gap by transforming the diversity-similarity tradeoff into a controllable, interactive experience, moving away from static rankings toward a dynamic exploration workflow.
System Design and Methodology
The core innovation of DiverXplorer is its stepwise interaction technique, which utilizes Determinantal Point Processes (DPP) to sample images. Unlike standard clustering or relevance-based ranking, DPP is a probabilistic model designed to select a subset of items that are "maximally spread out" within a feature space, ensuring high visual and semantic variety.
Stepwise Refinement
To manage the transition from exploration to selection, DiverXplorer implements a three-step workflow:
- Step 1 (Diversity Focus): The system maximizes diversity (div:sim = 1:0) to provide a comprehensive overview of the design space.
- Step 2 (Transition): As the user selects an image, the system begins to relax the diversity constraint (div:sim = 0.25:0.75), introducing more similarity to the chosen image.
- Step 3 (Similarity Focus): The final step prioritizes similarity (div:sim = 0:1) to the user's previous selections, allowing for fine-grained comparison and final choice.
Implementation
The system leverages CLIP (Contrastive Language-Image Pre-training) embeddings to represent images mathematically. These embeddings capture both visual features and semantic meaning, allowing the DPP algorithm to calculate the "distance" between images effectively. The interface presents a simple grid where users can progress by clicking "See more" or backtrack to explore different branches of the design space.
Evaluation and Results
The researchers conducted a user study with 22 professional graphic designers to evaluate DiverXplorer against three common baselines: a standard relevance-ranked grid (Scroll), a diversity-ranked grid without stepwise progression (Scroll+div), and a folder-based grouping tool (Clustering). The study utilized a dataset of 2,400 stock images from Pixabay across four open-ended design tasks.
Key Findings
- User Preference: DiverXplorer was the most preferred tool, ranked #1 by over 60% of the participants.
- Performance Metrics: The tool outperformed the standard "Scroll" baseline with statistical significance in several areas: Usefulness (p < .01), Smoothness (p < .001), Satisfaction (p < .05), and Time Saving (p < .01).
- Qualitative Impact: Designers reported that the tool helped them "clarify intent" and discover unexpected styles that they would have otherwise overlooked.
Discussion and Insights
The study revealed that diversity is a phase-dependent resource. While highly valuable for orientation in the early stages of design, it can become obstructive "noise" once a designer has committed to a specific style. DiverXplorer’s stepwise approach acts as a "forcing function," encouraging designers to pause and reflect rather than mindlessly scrolling.
Furthermore, the researchers noted that while clustering is a common organizational method, it often collapses diversity too quickly. If a designer selects a specific image, clustering tools might immediately hide all other styles (e.g., hiding illustrations if a photo is picked). DiverXplorer’s DPP-based approach maintains a controlled level of diversity longer, facilitating discovery without premature commitment.
Conclusion and Future Work
DiverXplorer demonstrates that interactive diversity control can significantly enhance the early-stage design process. While the current implementation faces computational challenges in scaling to millions of images due to the sequential nature of DPP sampling, it provides a robust framework for future exploration tools. Future research will focus on optimizing the algorithm for larger datasets, exploring alternative feature representations like DINOv2, and investigating how these selection tools integrate into broader design workflows across different cultural contexts.
Summary updated