On Fairness of Unified Multimodal Large Language Model for Image Generation

TL;DR

This paper investigates demographic biases in unified multimodal large language models for image generation, identifies bias sources, and proposes a balanced preference model to mitigate bias while maintaining output quality.

Contribution

It benchmarks U-MLLMs for bias, introduces a locate-then-fix debiasing strategy, and proposes a balanced preference model to reduce demographic bias effectively.

Findings

Most U-MLLMs exhibit significant gender and race bias.

Bias mainly originates from the language model component.

The proposed balanced preference model reduces bias while preserving semantic fidelity.

Abstract

Unified multimodal large language models (U-MLLMs) have demonstrated impressive performance in visual understanding and generation in an end-to-end pipeline. Compared with generation-only models (e.g., Stable Diffusion), U-MLLMs may raise new questions about bias in their outputs, which can be affected by their unified capabilities. This gap is particularly concerning given the under-explored risk of propagating harmful stereotypes. In this paper, we benchmark the latest U-MLLMs and find that most exhibit significant demographic biases, such as gender and race bias. To better understand and mitigate this issue, we propose a locate-then-fix strategy, where we audit and show how the individual model component is affected by bias. Our analysis shows that bias originates primarily from the language model. More interestingly, we observe a "partial alignment" phenomenon in U-MLLMs, where understanding bias appears minimal, but generation bias remains substantial. Thus, we propose a novel balanced preference model to balance the demographic distribution with synthetic data. Experiments demonstrate that our approach reduces demographic bias while preserving semantic fidelity. We hope our findings underscore the need for more holistic interpretation and debiasing strategies of U-MLLMs in the future.