The Illusion of High Utility in Safety Alignment of Text-to-Image Diffusion Models

2026-07-01

AI Summary

Existing safety alignment methods for text-to-image diffusion models, while performing well on coarse-grained metrics, severely compromise fine-grained semantic fidelity. This work reveals for the first time that such apparent “high utility” is in fact illusory and introduces the concept of “semantic collapse” to characterize the loss of structural information in the prompt embedding space. To address this issue, we propose Structure-Aware Geometric Regularization (SAGE), a mechanism that explicitly preserves the geometric distribution and semantic relationships within the embedding space during safety fine-tuning. Experimental results demonstrate that SAGE maintains strong safety guarantees while achieving a 5.0% improvement in TIFA-structured utility, significantly outperforming existing approaches.

Abstract

Safety alignment of text-to-image (T2I) diffusion models aims to suppress harmful generations while preserving utility on benign prompts. Recent methods often appear to deliver high safety with high utility, but this conclusion rests largely on coarse global utility metrics (e.g., FID, CLIPScore) that are insensitive to fine-grained semantic correctness, creating an illusion of high utility. We show that when utility is measured with structured evaluation, this illusion breaks: on TIFA (Text-to-Image Faithfulness evaluation with Question Answering), safety-aligned models suffer substantial drops in semantic fidelity, including failures in object counts, attributes, and relationships. To diagnose the source of this gap, we analyze the text-encoder prompt embedding space and uncover semantic collapse, a contraction of embedding spread coupled with distortion of inter-prompt similarity structure, which strongly correlates with structured utility loss. Guided by this insight, we propose StructureAware Geometric Regularization (SAGE), a safety alignment objective that explicitly preserves embedding spread and inter-prompt relational structure during adaptation. Our method restores structured utility (TIFA +5.0% over prior state-of-the-art) while maintaining strong safety performance and competitive coarse-grained utility scores. Our source code and trained models are available at https://adeelyousaf.github.io/SAGE_ECCV26_Project_Page/.

Problem

• safety alignment
• text-to-image diffusion models
• semantic fidelity
• utility illusion
• structured evaluation

Innovation

• semantic collapse
• structured utility
• safety alignment
• text-to-image diffusion models
• embedding geometry