Imagine A Reference: MLLM-Augmented Versatile Image Stylization
Abstract
Image stylization under multi-modal conditions faces challenges in semantic-style alignment due to limited cross-modal understanding. While Multimodal Large Language Model (MLLM) offers a solution, it often fails to ensure consistent semantic-style alignment. We propose MAIST, an MLLM-Augmented Image Stylization framework based on an Imagine-Infuse paradigm. The Imagine stage uses an MLLM to generate a text description from input conditions (text/image), which guides the creation of a stylized exemplar via a pre-trained diffusion model. In the Infuse stage, this text provides rich semantic and stylistic guidance, while the exemplar offers strong spatial constraints to enhance fidelity and controllability. This dual conditioning grounds structure while allowing flexible input control for content and style. Extensive experiments show MAIST achieves superior performance across diverse style transfer tasks. Project Page: https://github.com/Vincotto/MAIST