This paper addresses the problem of content-style decomposition (CSD), which separates content and style from a single image. Unlike existing diffusion model-based personalization methods, in this paper we propose a novel method, CSD-VAR, which performs CSD by utilizing visual autoregressive modeling (VAR). CSD-VAR introduces three key innovations to enhance the separation of content and style by leveraging the size-dependent generation process. First, we use a size-aware cross-optimization strategy to align content and style representations to their respective sizes. Second, we mitigate content leakage into style representations by using an SVD-based correction method. Third, we improve content identity preservation by using an augmented key-value (KV) memory. In addition, we introduce a new benchmark dataset, CSD-100, for CSD tasks. Experimental results show that CSD-VAR achieves better content preservation and style fidelity than existing methods.
