InvDiff: Invariant Guidance for Bias Mitigation in Diffusion Models

TL;DR

InvDiff addresses unknown biases in text-to-image diffusion models without relying on bias annotations. It introduces invariant guidance by adding a mean-shift term derived from invariant semantic information, and a two-phase learning framework (bias-annotation inference and invariant regularization) implemented through a lightweight module $G_\psi$ and an encoder $\Phi(y)$. The authors prove a generalization-bound-based guarantee and demonstrate reduced bias across Waterbirds, CelebA, and FairFace while preserving image quality, with extensions to data augmentation and time-series forecasting. The approach offers a practical, parameter-efficient route to debias diffusion models and holds promise for broader robust generative applications.

Abstract

As one of the most successful generative models, diffusion models have demonstrated remarkable efficacy in synthesizing high-quality images. These models learn the underlying high-dimensional data distribution in an unsupervised manner. Despite their success, diffusion models are highly data-driven and prone to inheriting the imbalances and biases present in real-world data. Some studies have attempted to address these issues by designing text prompts for known biases or using bias labels to construct unbiased data. While these methods have shown improved results, real-world scenarios often contain various unknown biases, and obtaining bias labels is particularly challenging. In this paper, we emphasize the necessity of mitigating bias in pre-trained diffusion models without relying on auxiliary bias annotations. To tackle this problem, we propose a framework, InvDiff, which aims to learn invariant semantic information for diffusion guidance. Specifically, we propose identifying underlying biases in the training data and designing a novel debiasing training objective. Then, we employ a lightweight trainable module that automatically preserves invariant semantic information and uses it to guide the diffusion model's sampling process toward unbiased outcomes simultaneously. Notably, we only need to learn a small number of parameters in the lightweight learnable module without altering the pre-trained diffusion model. Furthermore, we provide a theoretical guarantee that the implementation of InvDiff is equivalent to reducing the error upper bound of generalization. Extensive experimental results on three publicly available benchmarks demonstrate that InvDiff effectively reduces biases while maintaining the quality of image generation. Our code is available at https://github.com/Hundredl/InvDiff.

Figures (13)

• Figure 1: Proportion and samples of water/land bird in terrestrial/aquatic backgrounds.
• Figure 2: An overview of InvDiff. We first design a novel debiasing objective $\mathcal{L}_e$ for diffusion models. Then we propose a max-min game with the debiasing objective. We first infer potential bias annotations by maximizing the objective. Given the annotations, we finetune the biased model to unbiased by minimizing the proposed objective.
• Figure 3: Images sampled from our unbiased model.
• Figure 4: InvDiff's results on multiple unknown biases.
• Figure 5: Impact of $\boldsymbol{G}_\psi$ Parameters on InvDiff's Performance.