Graph-Driven Uncertainty Quantification in Text-to-Image Diffusion Models

ICLR 2026 Conference Submission17097 Authors

Keywords: Agent diffusers, Prompt Embeddings, Contrastive Clustering

Abstract: In this paper, we explore the problem of uncertainty quantification (UQ) in text-to-image generation models, focusing on the propagation of uncertainty through a graph-based structure of diffusion models. We propose three novel strategies to quantify and propagate uncertainty: Intrinsic and Propagated Uncertainty Coupling, Spectral Graph Uncertainty Propagation, and Path-Specific Uncertainty Influence. Each strategy leverages different aspects of graph theory to capture both local and global uncertainties in the generated images. We demonstrate how these methods provide insights into model reliability and robustness, and present experiments on several state-of-the-art text-to-image generation models. The results show that incorporating uncertainty information enhances model performance, guides further refinement, and improves reliability in real-world applications.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 17097