Diffusion models have become powerful tools for text-to-image generation, but they are not reliably safe. A new paper on arXiv proposes Decoupled Safety Control, a safety-control algorithm for training-free safety guidance. The approach is designed to intervene during sampling rather than requiring a separate safety-tuned model.
Training-free safety guidance is an inference-time strategy: it adjusts the generation process as it happens. The authors position this as a practical alternative for cases where retraining or fine-tuning is undesirable. Decoupled Safety Control is the specific mechanism they introduce for this purpose.
Because the abstract is the only available detail, the paper's exact formulation and evaluation results are not described here. The contribution, as stated, is a safety-control algorithm that works within the training-free paradigm to reduce unsafe content while preserving the generation strengths of diffusion models.