OASIS Uncovers: High-Quality T2I Models, Same Old Stereotypes

Sepehr Dehdashtian, Gautam Sreekumar, Vishnu N. Boddeti

ICLR 2025 (Spotlight)

Stereotypes in T2I Models

Stereotypes in T2I Models

Stereotypes in T2I Models

Stereotypes in T2I Models

Stereotypes in T2I Models

OASIS

A toolbox for measuring stereotypes and their origins in the T2I models

OASIS Overview

What Does OASIS Uncover About Stereotypes in T2I Models?

Lower, Yet Significant Stereotypes in Newer T2I Models

Nationality Stereotypes Worsen Existing Gender Stereotypes about Professions

Effective T2I Models Comparison Requires Both Stereotype Score and WAlS

Stereotypical Attributes Emerge in the Early Steps of Generation

T2I Models Internally Associate Concepts with Stereotypes

Summary

• Measures:
  • Level of Stereotype
  • Alignment of Concepts with Attributes
• Discovers:
  • Stereotypical Attributes that a T2I Model Internally Associates with a Concept
• Quantifies:
  • Emergence of Stereotypes During Image Generation