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

TL;DR

EMoE tackles the challenge of estimating epistemic uncertainty in large text-to-image diffusion models by leveraging pre-trained, sparse MoE components and performing uncertainty estimation directly in a diffusion latent space. By disentangling expert outputs at the first cross-attention layer, EMoE computes a latency-based uncertainty measure EU(y) that captures disagreement among experts, enabling early halting of expensive denoising for high-uncertainty prompts. Across COCO, multilingual prompts, and ablations, EMoE reveals that uncertainty correlates with image quality and exposes linguistic biases, particularly favoring high-resource or European languages. The approach offers zero-training uncertainty estimation with substantial computational savings and provides a practical tool for fairness, accountability, and bias assessment in AI-generated content.

Abstract

Estimating uncertainty in text-to-image diffusion models is challenging because of their large parameter counts (often exceeding 100 million) and operation in complex, high-dimensional spaces with virtually infinite input possibilities. In this paper, we propose Epistemic Mixture of Experts (EMoE), a novel framework for efficiently estimating epistemic uncertainty in diffusion models. EMoE leverages pre-trained networks without requiring additional training, enabling direct uncertainty estimation from a prompt. We leverage a latent space within the diffusion process that captures epistemic uncertainty better than existing methods. Experimental results on the COCO dataset demonstrate EMoE's effectiveness, showing a strong correlation between uncertainty and image quality. Additionally, EMoE identifies under-sampled languages and regions with higher uncertainty, revealing hidden biases in the training set. This capability demonstrates the relevance of EMoE as a tool for addressing fairness and accountability in AI-generated content.

Figures (16)

• Figure 1: This figure presents the model's uncertainty estimates when interpreting the same prompt in two languages. On the left, the English sentence "Two teddy bears are sitting together in the grass." yields a lower uncertainty score of 0.38, indicating greater model confidence. On the right, the corresponding Finnish translation "Kaksi nallekarhua istuu yhdessä nurmikolla." results in a higher uncertainty of 0.83, reflecting reduced confidence. This contrast highlights how the model responds differently to in-distribution (English) versus out-of-distribution (Finnish) language inputs.
• Figure 2: EMoE separates expert components in the first cross-attention layer in the first $\mathrm{down}$-block and processes each component separately as an independent computation path in the MoE pipeline. This results in $M$ distinct latent representations after the first denoising step. The figure illustrates an ensemble with two expert components, $\left(\,\text{and}\,\right)$.
• Figure 3: CLIP score across different uncertainty quartiles. EMoE accurately attributes prompts that produce images with high CLIP scores with low uncertainty unlike DECU. The red line indicates the average CLIP score across all quartiles.
• Figure 4: Uncertainty distribution for Finnish and English prompts, showing higher uncertainty for Finnish prompts compared to English.
• Figure 4: Mean Length of Finnish Prompts by Quartile of Uncertainty.