CAV Styler++: Extending Neural Style Transfer

Traditional Neural Style Transfer (NST) applies fixed styles with limited control over degree or progression. Recent techniques introduce spatial attention, entropy regularization, and latent manipulation, but still lack intuitive, human-understandable controls, leaving fine-grained intermediate stylization dependent on tuning loss weights or abstract parameters that require domain expertise. Building on {\em CAV Styler} \cite{cavstyler2025}, which learns Concept Activation Vectors for human-interpretable concepts such as texture, blur, and edge strength, in this paper, we extend it to photorealistic stylization via WCT2 \cite{WCT2}, a wavelet-based backbone whose transformation is spread across several encoder, decoder, and skip-connection levels rather than confined to the single fixed layer of the earlier setting. We find that naively placing the control mechanism at the deepest, most semantically analogous level has no effect on the output, and introduce a lightweight diagnostic to locate the level that is actually controllable. Using it, we construct and validate an various concepts that orders held-out examples monotonically and generalizes across diverse content images without per-image retraining. We show results on various photorealistic concepts such as weather, day-night, and photographic styles.