Fashioning Creative Expertise with Generative AI: Graphical Interfaces for Design Space Exploration Better Support Ideation Than Text Prompts

Title of the Paper

Fashioning Creative Expertise with Generative AI: Graphical Interfaces for GAN-Based Design Space Exploration Better Support Ideation Than Text Prompts for Diffusion Models

Paper Information

• Research Area: Human-Computer Interaction, Generative Artificial Intelligence, Creativity Support Tools
• Keywords: Generative AI, Deep Generative Models, Creativity Support Tools (CST), Creativity, Design Space Exploration, Fashion Design, Divergent Thinking, Convergent Thinking, Idea Generation

Research Background and Problem

• What problems or challenges did the authors identify?
  Current generative modeling tools lack critical features, limiting their potential to support creative professionals in exploring design spaces. Specifically, text-prompt-based diffusion models fail to effectively support both divergent and convergent creative thinking.

• Why is this problem important?
  Generative AI has been widely applied in creative fields, but its insufficient support for design space exploration hinders the realization of creative potential. Redesigning tools to better support creative processes, particularly divergent and convergent thinking, is crucial for enhancing creativity in professional design.

• Research Motivation and Related Work:
  This study investigates the impact of deep generative models on creative work and proposes a novel tool, generative.fashion, aimed at supporting divergent and convergent design space exploration through a graphical user interface. The authors also reviewed related research on generative AI and "design space exploration tools," finding that existing tools fail to comprehensively cover the entire design space.

Solution

• What methods or solutions did the authors propose?
  The authors proposed a GAN (Generative Adversarial Network)-powered tool, generative.fashion, which optimizes creativity support through a graphical user interface. Compared to text-prompt-based diffusion models, this tool better supports design space exploration through innovative interactive features.

• What are the innovative aspects of this solution?

  1. Support for Divergent Thinking: Randomly generates designs and explores unknown regions of the design space.
  2. Support for Convergent Thinking: Allows users to refine details and explore specific areas and subtle variations in designs.
  3. Unique Techniques: Enables visual exploration of the design space, such as style mixing in the GAN latent space and drag-and-drop functionality on a 2D design canvas.

• What are the implementation steps and key technologies used?

  1. Trained a generator using the StyleGAN2-ADA model and applied PCA (Principal Component Analysis) to identify semantically meaningful directions in the latent space.
  2. Developed a new user interface featuring interactive functionalities such as random generation, style mixing, and a 2D design canvas.
  3. Conducted qualitative and quantitative experiments to compare the generative.fashion tool with text-prompt-based diffusion models (Stable Diffusion) and Google Images.

Research Outcomes

• What specific outcomes were achieved?

  1. Qualitative studies revealed that generative.fashion better supports both divergent and convergent thinking, providing stronger creative support than text-prompt-based models.
  2. In quantitative experiments, the Creativity Support Index (CSI) score of generative.fashion was significantly higher than that of Google Images and Stable Diffusion.
  3. Compared to other tools, users rated generative.fashion higher in terms of satisfaction and usefulness.

• What advantages does it have over existing solutions?

  1. The interactive features of generative.fashion align more closely with theoretical principles of design space exploration.
  2. Compared to text-prompt-based models, it offers greater user control and better support for regional exploration.
  3. It not only facilitates precise adjustments to design details but also inspires users' creative ideas.

• What were the experimental or evaluation results?
  Through qualitative and quantitative studies, the authors demonstrated that generative.fashion is more innovative and practical compared to Stable Diffusion and Google Images, offering significant advantages in supporting creative design. User preferences and statistical results consistently supported the authors' hypotheses.

• Limitations and Future Directions

  1. The output quality of generative.fashion is inferior to diffusion models, particularly in generating accurate details.
  2. This study did not directly explore the potential of new interaction modes for diffusion models in supporting creativity.
  3. Future work could focus on improving generation quality and leveraging user interaction data to further investigate the impact of interaction on creative thinking.

This study highlights the importance of targeted graphical user interfaces in unleashing the creative potential of generative AI and provides actionable guidance for developing tools that support design space exploration.