Abstract
Recently, attentional arbitrary style transfer methods have been proposed toachieve fine-grained results, which manipulates the point-wise similaritybetween content and style features for stylization. However, the attentionmechanism based on feature points ignores the feature multi-manifolddistribution, where each feature manifold corresponds to a semantic region inthe image. Consequently, a uniform content semantic region is rendered byhighly different patterns from various style semantic regions, producinginconsistent stylization results with visual artifacts. We proposed theprogressive attentional manifold alignment (PAMA) to alleviate this problem,which repeatedly applies attention operations and space-aware interpolations.The attention operation rearranges style features dynamically according to thespatial distribution of content features. This makes the content and stylemanifolds correspond on the feature map. Then the space-aware interpolationadaptively interpolates between the corresponding content and style manifoldsto increase their similarity. By gradually aligning the content manifolds tostyle manifolds, the proposed PAMA achieves state-of-the-art performance whileavoiding the inconsistency of semantic regions. Codes are available athttps://github.com/computer-vision2022/PAMA.