Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2603.03163 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866914366122622976 |
|---|---|
| author | Chrabąszcz, Maciej Szymczyk, Aleksander Dubiński, Jan Trzciński, Tomasz Boenisch, Franziska Dziedzic, Adam |
| author_facet | Chrabąszcz, Maciej Szymczyk, Aleksander Dubiński, Jan Trzciński, Tomasz Boenisch, Franziska Dziedzic, Adam |
| contents | Despite their impressive capabilities, current Text-to-Image (T2I) models remain prone to generating unsafe and toxic content. While activation steering offers a promising inference-time intervention, we observe that linear activation steering frequently degrades image quality when applied to benign prompts. To address this trade-off, we first construct SafeSteerDataset, a contrastive dataset containing 2300 safe and unsafe prompt pairs with high cosine similarity. Leveraging this data, we propose Conditioned Activation Transport (CAT), a framework that employs a geometry-based conditioning mechanism and nonlinear transport maps. By conditioning transport maps to activate only within unsafe activation regions, we minimize interference with benign queries. We validate our approach on two state-of-the-art architectures: Z-Image and Infinity. Experiments demonstrate that CAT generalizes effectively across these backbones, significantly reducing Attack Success Rate while maintaining image fidelity compared to unsteered generations. Warning: This paper contains potentially offensive text and images. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_03163 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Conditioned Activation Transport for T2I Safety Steering Chrabąszcz, Maciej Szymczyk, Aleksander Dubiński, Jan Trzciński, Tomasz Boenisch, Franziska Dziedzic, Adam Computer Vision and Pattern Recognition Artificial Intelligence Despite their impressive capabilities, current Text-to-Image (T2I) models remain prone to generating unsafe and toxic content. While activation steering offers a promising inference-time intervention, we observe that linear activation steering frequently degrades image quality when applied to benign prompts. To address this trade-off, we first construct SafeSteerDataset, a contrastive dataset containing 2300 safe and unsafe prompt pairs with high cosine similarity. Leveraging this data, we propose Conditioned Activation Transport (CAT), a framework that employs a geometry-based conditioning mechanism and nonlinear transport maps. By conditioning transport maps to activate only within unsafe activation regions, we minimize interference with benign queries. We validate our approach on two state-of-the-art architectures: Z-Image and Infinity. Experiments demonstrate that CAT generalizes effectively across these backbones, significantly reducing Attack Success Rate while maintaining image fidelity compared to unsteered generations. Warning: This paper contains potentially offensive text and images. |
| title | Conditioned Activation Transport for T2I Safety Steering |
| topic | Computer Vision and Pattern Recognition Artificial Intelligence |
| url | https://arxiv.org/abs/2603.03163 |