The paper presents HiCAST, a new approach to Arbitrary Style Transfer (AST). AST aims to inject the artistic features of a style reference into an image or video. Unlike existing methods, HiCAST allows for more flexibility and customization in the stylization results. The model is built on the Latent Diffusion Model (LDM) and introduces a Style Adapter to manipulate output results. HiCAST also extends to video AST, improving cross-frame temporal consistency while maintaining stylization strength.
Publication date: 11 Jan 2022
Project Page: https://arxiv.org/abs/2401.05870
Paper: https://arxiv.org/pdf/2401.05870
