HoloFair: Unified T2I Fairness Evaluation and Fair-GRPO Debiasing
Researchers have introduced HoloFair, a new framework for evaluating and mitigating biases in text-to-image generation models. This framework includes a large-scale dataset and a metric called the Multi-attribute, Group-wise Bias Index (MGBI) to assess various demographic biases. Additionally, they developed Fair-GRPO, a reinforcement learning method that uses a multi-objective reward function to improve fairness without sacrificing image quality, as demonstrated on the SD3.5-Medium model. AI
IMPACT Introduces a new benchmark and debiasing technique to address fairness issues in generative AI, potentially leading to more equitable AI systems.