DTVI: Dual-Stage Textual and Visual Intervention for Safe Text-to-Image Generation
This work addresses safety concerns in text-to-image generation for users and platforms, offering a robust defense against malicious prompts, though it is incremental as it builds on existing inference-time methods.
The paper tackles the problem of unsafe content generation in text-to-image diffusion models by proposing DTVI, a dual-stage inference-time defense framework that achieves an average Defense Success Rate of 94.43% on sexual-category benchmarks and 88.56% across seven unsafe categories while preserving generation quality on benign prompts.
Text-to-Image (T2I) diffusion models have demonstrated strong generation ability, but their potential to generate unsafe content raises significant safety concerns. Existing inference-time defense methods typically perform category-agnostic token-level intervention in the text embedding space, which fails to capture malicious semantics distributed across the full token sequence and remains vulnerable to adversarial prompts. In this paper, we propose DTVI, a dual-stage inference-time defense framework for safe T2I generation. Unlike existing methods that intervene on specific token embeddings, our method introduces category-aware sequence-level intervention on the full prompt embedding to better capture distributed malicious semantics, and further attenuates the remaining unsafe influences during the visual generation stage. Experimental results on real-world unsafe prompts, adversarial prompts, and multiple harmful categories show that our method achieves effective and robust defense while preserving reasonable generation quality on benign prompts, obtaining an average Defense Success Rate (DSR) of 94.43% across sexual-category benchmarks and 88.56 across seven unsafe categories, while maintaining generation quality on benign prompts.