DiverXplorer: An Innovative Tool for Graphic Designers to Explore Stock Images via Diversity Adjustment

DiverXplorer is a prototype image exploration tool designed specifically for graphic designers. It allows users to transition from a broad overview to a focused subset by gradually adjusting diversity. The system uses Determinantal Point Process (DPP) sampling technology to achieve a dynamic balance between diversity and relevance in the CLIP embedding space, helping designers better engage in sense-making and visual pattern discovery during the early design stages.

In graphic design work, designers often need to explore large stock image libraries during open-ended design tasks or early creative stages. However, current mainstream image search tools (such as Adobe Stock, Pixta, Pinterest, etc.) mainly emphasize relevance and similarity sorting. While this design approach can quickly find images matching the search query, it limits designers' ability to overview the entire design space or discover visual patterns. When designers face a library of thousands of images, what they really need is not just a relevance-sorted list, but an exploration tool that helps them understand the overall visual landscape and discover potential creative directions. Traditional search tools often trap designers in a "scroll-click-return" loop, making it difficult to systematically explore the design space.

The research team from CyberAgent AI Lab proposed DiverXplorer, a prototype system designed specifically for stock image exploration. The core concept of DiverXplorer is to divide the image exploration process into multiple steps, where the images displayed in each step are optimized to meet the requirement of a gradual transition from diversity to similarity. Unlike traditional tools, DiverXplorer does not present all results at once; instead, it guides users through an exploration journey from "divergence" to "convergence". In the early steps, the system displays highly diverse images to help designers quickly grasp the full picture of the design space; as the exploration deepens, the system gradually narrows the scope and focuses on subsets of images more similar to the user's selection.

DiverXplorer uses the Determinantal Point Process (DPP) to achieve diversity control. DPP is a probabilistic model that can select a subset from a candidate set that maximizes dispersion. In DiverXplorer, the system calculates the similarity between images in the CLIP embedding space, then uses DPP sampling to select the most representative and diverse combination of images. CLIP (Contrastive Language-Image Pre-training) is a powerful multimodal model that can map images and text into the same embedding space. DiverXplorer uses CLIP's 512-dimensional feature vectors to represent each image, which allows it to calculate not only the visual similarity between images but also the semantic relevance between images and text descriptions.

DiverXplorer's exploration process is designed into multiple steps (usually 4 steps), where the diversity constraint for each step is gradually relaxed according to an exponential decay plan. Specifically:

• Steps 1-3: Diversity exploration phase, using an exponential decay strategy

  • Diversity score = exp(-decay rate × step number)
  • Composite score = diversity weight × diversity + relevance weight × relevance

• Step 4: Similarity convergence phase

  • Focus on images similar to the user's final selection
  • Help users make the final decision

In each step, the user selects an interested image from the displayed image grid (usually a 4×4 layout), and this selection serves as the condition for the next step of exploration. As the steps progress, the diversity constraint is gradually relaxed, and the displayed images become more and more similar, eventually helping the user converge to a satisfactory choice.

The research team conducted a user study with 22 professional graphic designers (16 of whom participated in the main study), comparing DiverXplorer with three baseline tools:

1. Scroll: Traditional scroll browsing sorted by relevance
2. Scroll+div: Scroll browsing with diverse reordering
3. Clustering: Step-by-step similarity refinement via clustering

The study results show:

• DiverXplorer was the top choice for over 60% of designers, standing out among all baseline tools
• Step-by-step diversity control effectively supports sense-making and visual pattern comparison in the early stages
• Diversity helps reduce backtracking behavior, but as exploration deepens, designers' demand for diversity decreases
• Designers particularly value the ability to switch smoothly between divergence (exploration) mode and convergence (comparison) mode, which baseline tools cannot support.

The core value of DiverXplorer lies in its understanding of designers' real needs during the creative exploration process. In the early stages of design, designers often do not know exactly what they want; they need to first understand "what possibilities exist" before gradually narrowing down the scope. Through its structured exploration process, DiverXplorer helps designers:

• Quickly establish an overall understanding of the design space
• Discover potential visual patterns and style trends
• Reduce time wasted on irrelevant options
• Make final choices more confidently.

The research results of DiverXplorer have important implications for the design of image search and recommendation systems. It shows that treating diversity as an interactive dimension rather than a fixed hyperparameter can better support users' exploratory tasks. This interactive design idea of "diversity-similarity trade-off" can be applied to other fields requiring exploratory search, such as music discovery, news reading, product browsing, etc.