HoloFair: A New Benchmark for T2I Fairness and Fair-GRPO Debiasing
A team of researchers has unveiled HoloFair, an all-encompassing benchmark framework aimed at analyzing multidimensional demographic bias in text-to-image (T2I) models. This framework is founded on a large-scale dataset focused on fairness and utilizes the SpaFreq (Spatial-Frequency) attribute classifier. HoloFair introduces the Multi-attribute, Group-wise Bias Index (MGBI) metric, which evaluates both intrinsic diversity and conditional biases, overcoming the shortcomings of current methods that concentrate on single-dimensional biases. Furthermore, they have developed Fair-GRPO, a reinforcement-learning-based debiasing technique that modifies generative model distributions via a multi-objective reward function. Experiments conducted with the SD3.5-Medium model confirm the effectiveness of this method. The findings are published in a paper on arXiv (2605.24687).
Key facts
- HoloFair is a benchmark for multidimensional demographic bias analysis in T2I models.
- It uses a large-scale fairness-oriented dataset and the SpaFreq attribute classifier.
- The MGBI metric assesses intrinsic diversity and conditional biases.
- Fair-GRPO is a reinforcement-learning-based debiasing method.
- Fair-GRPO uses a multi-objective reward function to alter generative model distributions.
- Experiments were conducted on the SD3.5-Medium model.
- Existing evaluation methods typically address only single-dimensional biases.
- The paper is available on arXiv with ID 2605.24687.
Entities
Institutions
- arXiv