AesCAST-An Automated Framework for Aesthetically-Guided Style Transfer
Abstract
Creative-media workflows such as visual-asset authoring and previsualization require a suitable reference style, yet existing systems usually leave that selection to the user and optimize primarily for feature correspondence. We introduce AesCAST, an automated framework that couples style recommendation with aesthetic-aware stylization. First, a dual-retrieval module uses CLIP similarity to shortlist semantically compatible WikiArt references and a frozen ArtCLIP-based scorer to select a high-scoring candidate. Second, an aesthetic loss derived from the same multi-attribute scorer augments the CAST backbone during training. On 1,000 held-out content-style pairs, AesCAST obtains the highest reported ArtCLIP and AANSPS scores among the evaluated CAST and AesUST baselines. In a 200-vote preference study, it is selected most frequently. These results support the proposed pipeline as an upstream authoring component under the reported protocol, rather than a general claim about aesthetic quality or downstream deployment.