The Safety-Aware Denoiser for Text Diffusion Models

Short summary

Researchers propose Safety-Aware Denoiser (SAD), an inference-time framework that guides text diffusion models toward safe outputs during denoising without retraining. SAD reduces unsafe generations while maintaining quality and diversity, outperforming existing safety methods. The approach is computationally efficient and scalable for enforcing safety constraints.

• •Introduces SAD framework for safety guidance in text diffusion models
• •Operates at inference time, avoiding expensive model retraining
• •Reduces unsafe outputs while preserving quality, diversity, and fluency

Generated with AI, which can make mistakes.