QuerySwitch: Supporting the Design Process by Balancing Vagueness through Large Language Models

Paper Title

QuerySwitch: Supporting the Design Process by Balancing Vagueness through Large Language Models

Publication Info

• Topic area: Human-AI collaboration in creative design processes
• Keywords: vagueness, large language models, fashion design, divergent-convergent process, creativity support tools, hierarchical keywords, combinational images, query-output modes, structured exploration, human-computer interaction

Background and Problem

• Problem / challenge: Existing LLM-based tools struggle to balance vagueness across the iterative divergent-convergent design process, leading to issues like excessive vagueness or fixation. Current systems lack mechanisms to support nuanced, context-sensitive assistance for managing vagueness.
• Significance: Balancing vagueness is critical in creative workflows to foster diverse exploration while maintaining coherence. Addressing this gap can improve the usability of LLMs in design tasks and enhance creative outcomes.
• Motivation and related work: Prior research has focused on generating creative outputs with LLMs but has largely treated vagueness as a starting point, neglecting its role throughout the design cycle. This paper builds on the need for systems that integrate vagueness management into LLM-based human-AI collaboration.

Solution

• Proposed approach: QuerySwitch, an interactive system enabling designers to dynamically switch between two query modes—Abstract Query-Hierarchical Keywords and Parallel Query-Combination Images—to balance vagueness during the design process.
• Novelty:
  1. Introduces hierarchical keywords to maintain coherence while supporting diverse exploration during divergence.
  2. Implements combinational images to prevent fixation and enable flexible idea synthesis during convergence.
  3. Supports abstract and parallel queries to align with the vagueness-driven nature of creative workflows.
  4. Provides a structured framework for iterative divergent-convergent cycles in design.
• Procedure and key techniques:
  1. Designers input abstract concepts or upload moodboards in the divergent phase, generating hierarchical keywords for exploration.
  2. In the convergent phase, designers use parallel queries to combine detailed elements, producing combinational images.
  3. The system allows switching between query modes across moodboard and sketch stages, ensuring alignment with the iterative design process.

Results

• Concrete findings:
  • Hierarchical keywords improved query complexity (t = -5.17, p < 0.001) and semantic consistency (U = 3007.50, p < 0.001) compared to the baseline.
  • Combinational images showed higher semantic similarity (U = 7.50, p = 0.001) and complexity (U = 3007.50, p < 0.001) than the baseline.
  • Participants rated QuerySwitch higher on usability (e.g., effectiveness t = 2.75, p = 0.022) and self-perceived experience (e.g., controllability t = 2.41, p = 0.039).
  • NASA-TLX scores indicated lower workload for QuerySwitch (overall score t = -2.55, p = 0.031).
• Advantage over baselines:
  • Reduced query adjustment effort and higher satisfaction with outputs compared to ChatGPT with DALL-E 3.
  • Enabled consistent alignment with design concepts and greater diversity in outputs.
• Experiments / evaluation:
  • Conducted a 90-minute within-subjects study with 10 professional fashion designers.
  • Measured usability (USE questionnaire), workload (NASA-TLX), and output quality (consistency and complexity metrics).
  • Compared QuerySwitch to a baseline system using the same underlying models (GPT-4-turbo and DALL-E 3).
• Limitations and future work:
  • Limited reflection of real-world fashion trends and cultural contexts in generated images.
  • Small participant pool (n=10) and short study duration may not capture the full diversity of design practices.
  • Future research should explore integration with trend databases, longitudinal studies, and multimodal interaction methods.

Summary

This paper introduces QuerySwitch, a system designed to balance vagueness in the creative design process by enabling dynamic switching between hierarchical and combinational query modes. The system effectively supports the iterative divergent-convergent cycle, enhancing creative exploration and preventing fixation. User studies with professional fashion designers demonstrated significant improvements in usability, output quality, and workflow efficiency compared to baseline systems. While limitations include the need for trend integration and broader participant sampling, QuerySwitch offers a promising framework for LLM-based creativity support tools across various design domains.