A Local Perceptual Approach for Few-Shot Text Effect Transfer

Text effect transfer (TET) aims to preserve the content of character images while rendering their style into various forms, including colors, outlines, shadows, textures, and glyphs. However, manually designing a complete font library is a labor-intensive task, making few-shot text effect transfer an increasingly important research focus. Existing methods often suffer from poor generalization, as their models are limited to a small range of text effects. Some approaches attempt to address this issue, but due to the scarcity of reference-style images, they tend to overfit or lack fine details, leading to failures when handling unseen text effects. To overcome these challenges, we propose a novel fine-tuning strategy that integrates Local Perceptual Fusion and Discrimination to enhance few-shot text effect transfer. Specifically, our fine-tuning strategy allows the model to adapt its parameters based on a small set of reference images from previously unseen styles, enabling the generation of realistic text effects. Additionally, we introduce a structure-level fusion mechanism in the style encoder to improve detail fidelity. To mitigate overfitting, we design a global discriminator and a local discriminator: the global discriminator assesses the overall realism of the generated styles, while the local discriminator performs fine-grained evaluation based on localized observations, ensuring both global consistency and fine-detail preservation. Experimental results demonstrate that our approach achieves advanced performance in few-shot text effect transfer, generating high-quality and highly faithful text effects.