We propose Scaled Autoregressive Projection Steering (SAPS) for generating incongruity-based figurative language from literal inputs, extending contrastive paraphrasing-based steering vectors by autoregressively recomputing projection coefficients during decoding. SAPS comprises two novel components: (A) Autoregressive Projection Steering (APS), which recomputes the projection coefficient (c_t) at each decoding step t, adapting the steering signal to the evolving token distribution, and (B) a scaling coefficient 𝛼 whose effective range is layer-agnostic, enabling compact tuning of figurative intensity. Experiments across three instruction-tuned language models show that APS improves grammatical stability over fixed sentence-level steering while 𝛼 enables fine-grained figurative control, together yielding gains over prior steering methods and significantly increasing robustness to layer selection and steering strength. Cross-dataset experiments further confirm that steering directions generalize across domains. This is the first work to apply activation steering to incongruity-based figurative language generation. Our code is available at: https://github.com/SharathHN/saps-steering.