Personalized image generation has mostly been studied in diffusion models, where methods often fine-tune weights, embeddings, or adapters for each subject. Kingfisher takes a different route: it works with visual autoregressive models and avoids model tuning entirely.

The method builds on the observation that early autoregressive scale codes largely determine object structure and global appearance, while later scales only refine the result. Kingfisher decomposes those early codes into content- and style-related components using prompt-conditioned subspaces. Given two images, it recombines the content of one with the style of the other into a compact semantic prefix, then lets a frozen generator predict the remaining scales with bitwise self-correction.

On the CSD-100 benchmark, the authors report strong content preservation and the highest style and text alignment among evaluated methods. Pair-specific preparation time drops by about 54x compared with diffusion-based approaches and 7.5x compared with other autoregressive baselines. The results suggest that direct manipulation of a frozen model's scale-space representation is a viable path to fast personalization.