Safety alignment for text-to-image diffusion models typically judges completed images, but that compresses many generation factors into one coarse signal. The authors of SaTPO argue this can push the model to change more than necessary. Their method, Safety Transition Preference Optimization, instead applies alignment incrementally during the denoising process, using the insight that an unsafe prompt and its safe rewrite form a natural preference pair when conditioned on the same parent state and noise.
By anchoring optimization on identical parent states, SaTPO avoids historical trajectory divergence and confines safety adaptation to a small step. Multi-scale drift constraints and a benign replay mechanism further limit behavioral deviation, helping preserve ordinary generation. The team also built CoProPV, a refined dataset with visually verified contrasts, to support training.
Experiments on Stable Diffusion v1.5 and SDXL show improved safety-quality trade-offs compared with AlignGuard. Unsafe-detection rates on Inappropriate Image Prompts drop from 12.70% to 6.00% and from 15.04% to 9.30%, respectively, while FID improves from 32.28 to 25.95 and from 36.50 to 27.51. The results suggest that step-wise alignment can be a more precise way to suppress harmful content without degrading useful generation.