SPQR: A multidimensional benchmark for safety alignment of text-to-image models under benign adaptation

The researchers propose the SPQR benchmark to evaluate whether the safety alignment of text-to-image diffusion models can be maintained under benign post-deployment fine-tuning (such as LoRA personalization, style and domain adapters). The authors observe that existing safety methods frequently fail under benign fine-tuning. SPQR covers four dimensions—Safety, Prompt adherence, Quality and Robustness—and provides a unified, reproducible evaluation framework and leaderboard score as a single score, facilitating comparison across methods. The benchmark also includes multilingual, domain-specific and out-of-distribution analyses and per-category breakdowns, used to locate when safety alignment fails after benign fine-tuning.