AI/ ai · diffusion-models · safety · benchmarks

Safety Filters in Image AI Are Hiding a Fidelity Problem

A new paper finds that standard benchmarks mask how much safety alignment degrades the semantic accuracy of text-to-image models.

The metrics used to evaluate AI image generators after safety tuning may be flattering a problem that structured testing exposes.

Researchers studying text-to-image diffusion models found that common evaluation scores — FID and CLIPScore — are too blunt to catch the damage that safety alignment does to a model's ability to correctly render object counts, attributes, and relationships. When they applied TIFA, a question-answering-based faithfulness benchmark, safety-aligned models showed substantial drops in semantic accuracy. The culprit, they argue, is "semantic collapse": safety training contracts the spread of text-encoder embeddings and warps how similar prompts relate to each other in that space. The team then built a regularization method called SAGE that explicitly protects embedding structure during safety adaptation, recovering a 5-percentage-point TIFA gain over prior best results without sacrificing safety scores.

The implication cuts across the whole field: if the standard yardsticks don't catch semantic degradation, then every published claim of "high utility with high safety" is suspect. That matters because text-to-image models are increasingly deployed in products where prompt fidelity — not just aesthetic plausibility — is the actual requirement.

Safety and capability have long been framed as a trade-off in language models; this paper suggests image generators have been quietly making the same trade-off while the benchmarks looked the other way.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →