CAV Styler++: Locating Interpretable Control in Photorealistic Style Transfer

Building on CAV Styler's interpretable concept directions, we extend this mechanism to photorealistic stylization and show that \emph{where a concept is injected determines whether it controls the output}. In WCT2, the deepest stage produces no measurable response in our diagnostic. A lightweight sensitivity test identifies a shallow stage with $36\times$ the response of the next encoder level. There, a day$\leftrightarrow$night CAV learned from extreme grades of 15 images correctly orders three held-out intermediate grades and transfers across 12 content images without relearning the direction. Learning takes under a minute. We characterize the content-dependent range and directional asymmetry, establishing a working extension beyond artistic stylization and an empirical procedure for locating effective concept control.