Iconix: Controlling Semantics and Style in Progressive Icon Grids Generation

Paper Title

Iconix: Controlling Semantics and Style in Progressive Icon Grids Generation

Publication Info

• Topic area: Human-AI co-creative systems for icon design.
• Keywords: Icon design, semantic richness, visual complexity, generative AI, progressive abstraction, human-AI collaboration, semantic scaffolding, style consistency, creative workflows, visual abstraction.

Background and Problem

• Problem / challenge: Existing generative AI tools struggle to balance semantic control and stylistic consistency, particularly for abstract concepts. Designers face a tedious trial-and-error process to create cohesive icon sets with varying levels of detail and abstraction.
• Significance: Icons are essential for user interfaces, requiring adaptability across devices, contexts, and user needs. Streamlining the design process while maintaining creative control can enhance productivity and innovation in visual communication.
• Motivation and related work: Prior systems focus on either semantic composition or stylistic application but fail to integrate both dimensions effectively. Generative models often entangle semantics and style, limiting their utility for structured icon design workflows. This paper addresses these gaps by introducing a method for generating structured icon grids that balance semantic richness and visual complexity.

Solution

• Proposed approach: Iconix, a human-AI co-creative system that generates progressive icon grids by organizing design along two axes: semantic richness and visual complexity.
• Novelty:
  1. A computational method for generating progressive icon grids, coupling semantic scaffolding with progressive visual simplification.
  2. A dual-axis grid interface that enables structured exploration of design trade-offs.
  3. Empirical validation demonstrating reduced cognitive workload and enhanced creative exploration.
• Procedure and key techniques:
  1. Semantic Expansion: Uses large language models and knowledge bases to generate related concepts and rank them by attributes like concreteness and imageability.
  2. Semantic Scaffolding: Constructs a structured knowledge representation with taxonomic, constitutive, and associative dimensions to guide exemplar generation.
  3. Progressive Simplification: Employs Score Distillation Sampling (SDS) and semantic segmentation to create a continuum of icon abstractions.
  4. Style Refinement: Fine-tunes generative models to produce stylistically consistent icons in outline, filled, and colorized formats.

Results

• Concrete findings:
  • Iconix achieved a usability score of 79.53 (vs. 63.59 for the baseline).
  • Reduced cognitive load across all NASA-TLX dimensions, including mental demand (2.41 vs. 3.41) and effort (2.84 vs. 4.19).
  • Higher design satisfaction, with significant improvements in semantic richness (5.84 vs. 3.75), visual complexity (5.59 vs. 3.94), and overall satisfaction (5.81 vs. 4.53).
• Advantage over baselines:
  • Iconix outperformed ChatGPT in usability, cognitive load reduction, creativity support, and design satisfaction.
  • Produced clearer semantic progression, better stylistic consistency, and more diverse and creative outcomes.
• Experiments / evaluation:
  • A within-subjects study with 32 participants compared Iconix to ChatGPT across two tasks: designing icons for a concrete concept ("Hamburger") and an abstract concept ("Hope").
  • Metrics included usability (SUS), cognitive workload (NASA-TLX), creativity support (CSI), and design satisfaction.
• Limitations and future work:
  • Small participant sample limits generalizability.
  • Dependence on segmentation accuracy; future iterations could incorporate human-in-the-loop corrections.
  • Comparison focused on general-purpose AI (ChatGPT) rather than professional design tools like Adobe Illustrator.

Summary

Iconix is a human-AI co-creative system that enables designers to generate progressive icon grids by balancing semantic richness and visual complexity. Through semantic scaffolding, progressive simplification, and style refinement, Iconix supports structured exploration and reduces cognitive workload. A study with 32 participants demonstrated that Iconix outperforms a baseline workflow in usability, creativity support, and design satisfaction. This work advances the field of generative AI by operationalizing abstraction as a controllable parameter, bridging human intent with machine capability for scalable and cohesive icon design.