Better Little People Pictures: Generative Creation of Demographically Diverse Anthropographics

Document Title

Better Little People Pictures: Generative Creation of Demographically Diverse Anthropographics

Document Information

• Subject Area: Data Visualization, Human-Computer Interaction, Generative AI, Social Equity and Bias
• Keywords: Diverse Anthropographics, Generative Models, Demographic Data, Visual Design, Bias and Ethics, Data Visualization, Creative Design, Human-Computer Interaction, Generative AI, Humanistic Computing

Research Background and Issues

Research Background

• Anthropographics are common tools in visualization research and data journalism, used to represent data about populations.
• Previous studies have focused more on the role of anthropographics in evoking empathy and social emotions, but recently, designers and researchers have increasingly emphasized the importance of these graphics in reflecting individual diversity within data.
• The creation of anthropographics is currently labor-intensive and requires significant design resources.

Main Issues and Challenges

1. High Design Barriers: Existing tools are insufficient for quickly creating anthropographics that reflect diversity and accuracy, especially for large-scale representations of complex features like race, gender, or age.
2. Bias Issues: Biases inherent in generative models and designers themselves can be transferred to the generated graphics, potentially reinforcing stereotypes related to race, gender, etc.
3. Ethics and Privacy: Issues such as data agency, privacy, data misuse (e.g., deepfakes), and transparency in generated content require urgent scrutiny.

Research Motivation

• Generative text-to-image models (e.g., Stable Diffusion) have made remarkable progress in recent years, significantly lowering the barriers to creating complex and diverse anthropographics. However, challenges related to generative bias and graphic ethics need to be addressed.
• By leveraging generative models and integrating human design processes, it is possible to elegantly overcome these challenges while achieving richer and more expressive graphical representations.

Solution

Proposed Method and Workflow

• Conceptual Workflow: A "human-AI collaborative" design process is proposed, integrating human designers into the generative AI creation process for active supervision from data preparation to image generation.
  • Data selection and model filtering
  • Image generation and iterative optimization
• Technical Tools:
  • An interactive workflow tool developed in Google Colab based on the Stable Diffusion model (e.g., anthrogen Python library).
  • Public demographic data (e.g., U.S. Census data) is used to drive image generation.

Innovations

• Human-AI Collaboration: Combines manual design with generative models to reduce the risk of bias infiltration.
• Support for Data Diversity: Enables designers to incorporate real or simulated (generated) diverse demographic attributes, addressing issues like beauty bias and gender stereotypes.
• Flexible Visual Styles: Generates anthropographics in diverse visual styles (photo-realistic, cartoonish, simplified symbolic).

Implementation Steps

1. Data Selection and Expansion: Use census data or simulated demographic labels to create diverse datasets.
2. Model and Template Selection: Filter suitable Stable Diffusion variants and design preliminary text templates.
3. Image Generation and Quality Monitoring: Generate images and adjust templates to correct errors or biases in the graphics.
4. Chart Design and Integration: Design data visualization works based on the generated images.

Research Outcomes

Key Results

• A conceptual workflow combining generative models and manual design was proposed to efficiently create diverse anthropographics.
• Algorithms and experiments were designed to improve the scalability of graphic styles, such as cartoonish, hand-drawn, and photo-realistic anthropographics.

Advantages

• Significantly reduced the time and labor costs required to generate anthropographics.
• Enhanced the representation of diverse populations in terms of race, gender, age, etc.
• Proposed clear design steps and solutions to potential risks, supporting transparent presentation of generated content.

Experimental Evaluation

• Case Studies:
  • Abstract anthropographics design for female maternity leave rates: comparing homogeneous and diverse graphics.
  • Exploration of cartoonish anthropographics depicting gender ratios among U.S. college students.
  • Realistic graphics showcasing electoral representative population ratios in two U.S. states.
• Observed Limitations:
  • Persistent beauty bias favoring youth and idealized aesthetics.
  • Generated images of race and gender often lack individual differentiation.
  • Certain social labels may still manifest as stereotypes.

Limitations and Future Directions

1. Model and Data Constraints: Current training data for models is limited and contains biases related to gender, age, and race.
2. Transparency and Trust: Stronger tools are needed to explain the origins and credibility of generated graphics.
3. Future Research Directions:
   • Develop generative models specifically for anthropographics to reduce training data bias and support more flexible visualization features.
   • Build tools to enhance bias detection for designers and audit diversity in outputs.
   • Promote transparency in the generative and design processes, including features that support dialogue with audiences.

☑ Summary: The study provides a structured workflow and tool support for leveraging generative AI models to address data diversity visualization, proposing potential solutions to key ethical and technical challenges, laying the foundation for future research.