On Fairness of Unified Multimodal Large Language Model for Image Generation

TL;DR

The paper addresses demographic bias in unified multimodal large language models (U-MLLMs) used for text-to-image generation. It introduces a locate-then-fix framework that first audits bias sources by testing whether bias originates from the vision encoder or the language model, finding that the language model is the primary source and that a partial alignment between understanding and generation exists. A data-synthesis workflow plus a balanced preference loss are proposed to debias outputs, implemented via a two-stage training pipeline (supervised finetuning followed by balanced preference optimization). Empirical results show substantial reductions in gender and race bias across multiple U-MLLMs (e.g., gender bias reductions up to about 0.72 of prior levels; specific instances include a 71.9% reduction for VILA-U) while preserving semantic fidelity and image quality, indicating the approach is effective and generalizable. The work highlights the need for holistic fairness strategies that address both perceptual understanding and generative tendencies in UMLLMs, offering practical guidance for deploying fair multimodal generators.

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.

Figures (13)

• Figure 1: U-MLLMs are capable of generating images with high quality, but the generation lacks diversity. The model has a bias for some occupations. In this example, given the prompt "construction worker", the model generate most of images with demographic attribute as "male" and "white".
• Figure 2: The optimization objective is to minimize the deviation of preference between different demographic group.
• Figure 3: Given the test prompt: “a photo of the face of a electrical and electronics repairer, a person”. Images generated by: VILA-U (bottom-left), our method reduced bias in gender(bottom-right), race(top-left), gender-race(top-right). For each image, the first color-coded bar represents the predicted gender: male or female. The second bar represents the predicted race: While, Asian, Black, or Indian. The height of each bar indicates the prediction confidence.
• Figure 4: Compare the understanding and generation bias.
• Figure 5: Overview of framework’s multi-modal training and inference processwu2024vila Visual inputs are converted into discrete tokens and merged with textual tokens to create a unified multi-modal token sequence. This sequence is used in next-token prediction process, which supports a unified training objective. During inference, output tokens are processed through either text detokenizer or vision tower decoder, generating multi-modal content outputswu2024vila.