2605.30049v1 May 28, 2026 cs.AI

Robust and Generalizable Safety Steering for Text-to-Image Diffusion Transformers

Zhen Bi
Zhen Bi
Citations: 15
h-index: 2
Jungang Lou
Jungang Lou
Citations: 44
h-index: 3
Zihao Xue
Zihao Xue
Citations: 6
h-index: 2
Longtao Huang
Longtao Huang
Citations: 31
h-index: 2
Long Ma
Long Ma
Citations: 5
h-index: 2
Zeyu Yang
Zeyu Yang
Citations: 3
h-index: 1
Bin Zhu
Bin Zhu
Citations: 132
h-index: 5
Jie Xiao
Jie Xiao
Citations: 17
h-index: 2
Yan Wang
Yan Wang
Citations: 12
h-index: 2
Zhonglong Zheng
Zhonglong Zheng
Citations: 493
h-index: 12

Diffusion Transformers have become a powerful backbone for text-to-image generation, but their layered and cross-modal generation process makes safety control fundamentally different from prompt-level filtering or output-level detection. Harmful semantics may be weakly expressed in text representations, progressively bound to visual latents, and finally entangled with rendering dynamics. As a result, safety steering at a fixed layer can be unstable, and a steering mechanism learned from known risks may not transfer reliably to a shifted target risk domain. We propose SafeDIG, a safety steering framework that formulates DiT safety adaptation as position-aware sparse feature transfer. SafeDIG first constructs Sparse Autoencoders over functionally distinct DiT intervention positions and uses robustness-aware pre-training routing to prioritize intervention sites that are expected to remain stable under source-target risk shift. It then separates transferable safety features from domain-specific activation geometry by freezing the SAE encoder as a reusable sparse safety dictionary and adapting only the decoder to the target-domain activation manifold. During inference, SafeDIG combines Blend and Repel operations to steer unsafe activations toward transferred safety manifolds or away from harmful sparse directions. Experiments on FLUX.1 Dev and Stable Diffusion 3.5 Large show that SafeDIG consistently reduces target-domain and overall unsafe generation rates while preserving source-domain safety and image quality.

0 Citations
0 Influential
6 Altmetric
30.0 Score
Original PDF

No Analysis Report Yet

This paper hasn't been analyzed by Gemini yet.

Log in to request an AI analysis.

댓글

댓글을 작성하려면 로그인하세요.

아직 댓글이 없습니다. 첫 번째 댓글을 남겨보세요!