Introspective Attention Modulation for Safe Text-to-Image Generation

As text-to-image (T2I) models become more powerful, they remain susceptible to generating harmful, explicit, or violent content, particularly when users employ parameter-efficient fine-tuning techniques like LoRA to bypass built-in safety filters. The authors investigate whether diffusion models can be made safer by introspectively monitoring and modulating their own internal attention mechanisms during the image generation process, rather than relying on post-hoc filters or permanent weight modifications.

Instead of retraining model weights or filtering prompts, the authors propose Introspective Attention Modulation (IAM). This method operates during the inference-time denoising process. At each timestep, the model extracts its cross-attention and self-attention maps. A pre-trained, sparse logistic regressor identifies attention patterns associated with unsafe concepts. The system then solves a constrained optimization problem to modulate these attention maps—suppressing the 'unsafe' signals while maintaining semantic alignment with the user's prompt. This process is integrated directly into the flow-matching ODE, allowing the model to self-correct its trajectory toward safe image synthesis without requiring backpropagation through the model parameters.

Experimental results demonstrate that IAM significantly outperforms existing concept-erasure techniques (such as UCE, ESD, and MCE) across standard safety benchmarks like I2P and SneakyPrompts. While baseline methods often suffer from visual artifacts, incomplete suppression, or collateral damage to image quality, IAM achieves high safety scores while preserving photorealism and semantic fidelity. The authors show that attention-space regulation is a more robust and flexible path for modern diffusion transformers than traditional weight-editing approaches, as it effectively neutralizes the influence of malicious adapters without needing to know the specific nature of the fine-tuning.