The Safety-Aware Denoiser for Text Diffusion Models

The article introduces the Safety-Aware Denoiser (SAD), a framework designed to enhance safety in text diffusion models by guiding the denoising process towards safe text regions without requiring retraining of the underlying model. Experimental evaluations indicate that SAD significantly reduces unsafe outputs while maintaining generation quality, showing its effectiveness over existing safety methods. This approach is crucial for practitioners as it offers a lightweight and scalable solution to integrate safety measures directly into the inference process of text diffusion models.