Reproduce “Mitigating stereotypical biases in text to image generative systems” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Piero Esposito, Parmida Atighehchian, Anastasis Germanidis, Deepti Ghadiyaram

Researchers at Runway developed a practical approach to measure and mitigate stereotypical biases in text-to-image (TTI) models, demonstrating substantial improvements in fairness metrics for perceived gender and skin tone. Their method involves targeted finetuning with a synthetically generated diverse dataset, which effectively reduces bias without compromising image quality or prompt alignment.

Key insights

  • Targeted finetuning with synthetically generated diverse datasets can substantially improve fairness in Text-to-Image (TTI) models.
  • The composition of training data, particularly a uniform distribution of demographic attributes, is crucial for building fairer AI systems.
  • Bias mitigation can be achieved without compromising image generation quality or prompt alignment in TTI models.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.