FAIntbench: A Holistic and Precise Benchmark for Bias Evaluation in Text-to-Image Models

TL;DR

FAIntbench introduces a holistic bias benchmark for Text-to-Image models by formalizing a four-dimension bias definition system (manifestation, visibility, acquired attribute, protected attribute), a 2,654-prompt dataset, and 18 automated evaluation metrics. It combines CLIP-based alignment with implicit and explicit bias scores ($S_{sum,im}$, $S_{sum,ex}$) and a manifestation factor $oldsymbol{ au}$ to separate ignorance from discrimination, enabling fine-grained bias analysis. Evaluations on seven contemporary T2I models reveal persistent racial biases and reveal distillation as a potential driver of bias, with human evaluation corroborating automated findings. FAIntbench provides a transparent, reproducible framework and open-source tools to advance debiasing research in AIGC, guiding safer deployment of T2I systems.

Abstract

The rapid development and reduced barriers to entry for Text-to-Image (T2I) models have raised concerns about the biases in their outputs, but existing research lacks a holistic definition and evaluation framework of biases, limiting the enhancement of debiasing techniques. To address this issue, we introduce FAIntbench, a holistic and precise benchmark for biases in T2I models. In contrast to existing benchmarks that evaluate bias in limited aspects, FAIntbench evaluate biases from four dimensions: manifestation of bias, visibility of bias, acquired attributes, and protected attributes. We applied FAIntbench to evaluate seven recent large-scale T2I models and conducted human evaluation, whose results demonstrated the effectiveness of FAIntbench in identifying various biases. Our study also revealed new research questions about biases, including the side-effect of distillation. The findings presented here are preliminary, highlighting the potential of FAIntbench to advance future research aimed at mitigating the biases in T2I models. Our benchmark is publicly available to ensure the reproducibility.

Figures (11)

• Figure 1: Pipeline for the prompt set. The grey rectangle represents an identity prompt, the orange rectangle represents an photorealism prompt, the green rectangle represents the protected attributes, and yellow rectangle represents the acquired attributes.
• Figure 2: Pipeline for the evaluation. The yellow rectangle represents generated images, the black rectangle represents the meta data from alignment, the green box represents selected prompts for manifestation factor, and the red box represents the ground truth.
• Figure 3: Axis for the manifestation factor $\eta$. The red line shows the value of $\eta$ and the dashed line shows the initial value $\eta_0$, which is 0.5.
• Figure 4: Examples of prompts and images. The blue rectangle means an implicit prompt with only acquired attributes, yellow for explicit with both acquired attributes and protected attributes, and orange for photorealism. Green and red boxes show success and failure, respectively. Images are generated by SDXL Turbo.
• Figure 5: Quantitative results for implicit bias scores of the protected-attribute level.