InspirationGraph for Progressive Design Space Exploration

Paper Title

InspirationGraph for Progressive Design Space Exploration

Publication Info

• Topic area: Enhancing early-stage design ideation using AI-assisted tools.
• Keywords: Text-to-image models, design space exploration, divergent thinking, novice designers, cognitive load, prompt engineering, visualization, human-AI collaboration, creativity support tools.

Background and Problem

• Problem / challenge: Existing Text-to-Image (T2I) tools primarily support the convergent stages of design, focusing on refinement and visualization rather than early-stage exploratory ideation. These tools rely on linear, one-shot interactions and demand precise prompt engineering, which can be cognitively taxing and limit open-ended exploration.
• Significance: Early-stage design requires divergent thinking and systematic exploration of ideas. Current T2I tools fail to support this phase effectively, particularly for novice designers, who face challenges such as design fixation and difficulty articulating vague creative intentions.
• Motivation and related work: Prior research has highlighted the importance of design space exploration and the potential of AI as a generative partner in creativity. However, existing tools lack structured mechanisms for iterative exploration and fail to balance guidance with flexibility. This paper addresses the gap by focusing on early-stage ideation with a novel interaction paradigm.

Solution

• Proposed approach: InspirationGraph, a prototyping tool based on a dimension–attribute dictionary and a tree-based visualization framework, supports progressive design space exploration for novice designers.
• Novelty:
  1. A new paradigm for multi-dimensional, iterative design exploration using semantic attributes.
  2. Development of InspirationGraph, which simplifies prompt engineering and visualizes exploration paths interactively.
  3. Empirical insights into novice designers’ cognitive needs and interaction patterns with AI-assisted tools.
• Procedure and key techniques:
  • Introduce a dimension–attribute dictionary to guide prompt construction incrementally.
  • Use a tree-based visualization to organize and trace exploration paths.
  • Combine auto and manual iteration modes: auto mode suggests new dimensions and attributes via LLM, while manual mode allows user-specified inputs.
  • Generate images through prompt inheritance, ensuring semantic consistency across iterations.

Results

• Concrete findings:
  • InspirationGraph significantly improved Creativity Support Index (CSI) scores (M = 69.11, SD = 7.64) compared to the baseline (M = 58.59, SD = 7.91; p < .001).
  • Reduced cognitive load, as shown by lower NASA-TLX scores (M = 37.89, SD = 10.38) versus the baseline (M = 52.15, SD = 9.39; p < .001).
  • Users generated more images (M = 19.79) with fewer prompts (M = 14.46) compared to the baseline (images: M = 10.71, prompts: M = 59.63; p < .001).
• Advantage over baselines:
  • Enhanced support for divergent exploration and goal orientation.
  • Reduced effort in prompt engineering and improved usability.
  • Visualized exploration paths facilitated reflection and comparison.
• Experiments / evaluation:
  • Conducted a user study with 24 novice designers, comparing InspirationGraph to a baseline tool.
  • Tasks involved designing lounge and dining chairs, with participants alternating tools and tasks.
  • Evaluated using CSI, NASA-TLX, and a custom Progressive Exploration Support Scale (PESS).
• Limitations and future work:
  • Current T2I models struggle with fine-grained control when multiple dimensions are specified.
  • The semantic dimensions used may not generalize across design domains.
  • Future work should refine dimension–attribute selection and improve model capabilities for multi-dimensional compositional control.

Summary

This paper introduces InspirationGraph, a tool designed to support early-stage design ideation by combining structured semantic control with a tree-based visualization framework. The system reduces cognitive load, enhances divergent thinking, and aligns with novice designers’ needs by simplifying prompt engineering and visualizing exploration paths. A user study demonstrated significant improvements in creativity support and usability compared to a baseline tool. While limitations in T2I model precision and generalizability remain, this work lays a foundation for extending AI-assisted design tools to other creative domains.