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
TL;DR
This paper introduces EMoE, a novel framework that efficiently estimates epistemic uncertainty in text-to-image diffusion models, revealing biases and improving understanding of model confidence without additional training.
Contribution
EMoE leverages pre-trained networks and a latent space to estimate uncertainty directly from prompts, uncovering biases and enhancing fairness in AI-generated images.
Findings
EMoE correlates uncertainty with image quality on COCO.
It identifies under-sampled languages and regions with higher uncertainty.
Demonstrates effectiveness in revealing hidden biases.
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…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
- Zero shot uncertainty estimation. - Consistent on in-distribution data shows reliability. - Effective bias and OOD detection that can be used to enhance training. - Detailed ablation studies and qualitative examples.
- The experimental section's findings are largely correlational and serve only to re-confirm established phenomena in generative modeling (e.g., uncertainty's inverse correlation with image quality, OOD inputs increase uncertainty, descriptive prompts lower uncertainty). The paper successfully proves EMoE works, but fails to provide new scientific insight. - "halting the denoising process immediately for uncertain prompts". This is a strong case for the impact of the paper yet the paper contains
1. EMoE provides a novel approach to estimating epistemic uncertainty in diffusion models without additional training. 2. EMoE addresses transparency and bias detection in black-box generative models, a timely and socially relevant problem. 3. EMoE avoids training overhead (unlike DECU), showing computational savings and environmental benefits. 2. EMoE demonstrates superior performance on the COCO dataset compared to existing methods. 3. EMoE provides ablation studies that analyze design cho
1.The paper lacks a detailed comparison with other uncertainty estimation techniques beyond the COCO dataset, since this paper is mainly on DECU. 2.The scalability of EMoE to larger and more complex diffusion models is not thoroughly explored. 3.This paper only focus on COCO and CC3M (in appendix), which should be verified as evidence by evaluating other datasets. 4. Heavy reliance on CLIP-based metrics (known to favor English-centric datasets) may reinforce the very bias the paper aims to me
1. Their EMoE design does not require further training and estimates uncertainty in the latent space of the denoiser, enabling early detection of undersampled prompts before image generation is complete which makes it more applicable to real-world scenarios. 2. They conduct thorough experiments over different language prompts to prove their framework.
1. A potential weakness of this method is that the variance across experts may not purely reflect epistemic uncertainty. In DMs, cross-attention modules inherently promote generation diversity, which can inflate variance even for well-understood inputs. As a result, the estimated uncertainty may conflate true model epistemic uncertainty with normal semantic diversity. 2. The output differences between different expert modules may do not necessarily indicate model uncertainty, rather, to some ext
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsGenerative Adversarial Networks and Image Synthesis · Computational and Text Analysis Methods · Misinformation and Its Impacts
MethodsDiffusion