Seeing the Unseen: How EMoE Unveils Bias in Text-to-Image Diffusion Models

Lucas Berry, Axel Brando, Wei-Di Chang, Juan Camilo Gamboa Higuera, David Meger

Keywords: Diffusion Models, Uncertainty Quantification, Text-to-Image

Abstract: Estimating uncertainty in text-to-image diffusion models is challenging due to their massive parameter counts (often exceeding 100M) and operation in complex, high-dimensional spaces with virtually unbounded input domains. We introduce Epistemic Mixture of Experts (EMoE), a framework for efficient estimation of epistemic uncertainty in diffusion models. EMoE leverages pre-trained networks without requiring additional training, enabling direct uncertainty estimation from a prompt. By probing a latent space within the diffusion process, EMoE captures epistemic uncertainty more effectively than existing approaches. Experiments on the COCO dataset demonstrate EMoE’s superior performance. Beyond benchmark gains, EMoE highlights under-sampled languages and geographic regions associated with elevated uncertainty, uncovering hidden biases in training data. Since training data for online diffusion models is rarely made public, this bias-detection capability is especially valuable. Together, these contributions position EMoE as a practical tool for addressing data imbalance and improving inclusivity in AI-generated content.

Primary Area: generative models

Submission Number: 6336