Steering Fields generalize classical steering vectors to be adaptive along the generation trajectory. Instead of a fixed shift, the steering direction is re-estimated at each step based on the current noisy state, which better preserves local structure and avoids unintended global changes.
The method achieves state-of-the-art results on safety steering benchmarks, effectively suppressing harmful content while maintaining prompt fidelity. It also naturally extends to image editing without requiring inversion or spatial masks, setting new best semantic adherence scores on CLIP and VQAScore.
Steering Fields are model-agnostic, work across different flow models, and offer a single parameter to balance steering strength and content preservation. They support simultaneous induction and inhibition of concepts, making them a flexible tool for controlled generation.