A multi-constraint handwritten generation model for ancient Yi script digitization engineering

Ancient Yi books are preserved in handwritten copies, lacking unified standard encoding and frequently suffering from incompleteness and blurring. Their digitization relies on manual transcription by cultural experts, which suffers from low efficiency and poor standard consistency. Handwritten text generation offers a promising solution for such issues. Nevertheless, most existing handwritten generation algorithms are designed for high-resource languages such as English and Chinese, and struggle with the generation and restoration tasks for ancient Yi scripts. To address this gap, this paper proposes YiGen, a multi-constraint handwritten generation model for ancient Yi script, which adopts five customized loss functions and only requires one stylized ancient Yi script reference image and one content-specified ancient Yi script image. Specifically, the Content Extracter and Style Extracter separately extract valid features from the two images. Then the Diffusion Guidance Fusioner integrates the content and style information, and the fused features are fed into the Conditional Diffusion Module as guidance to generate the target handwritten ancient Yi scripts. In addition, we propose a two-stage training approach to enhance YiGen’s generation and generalization capabilities. We conduct comprehensive experiments on our self-built dataset AncientYi and public dataset Yi Dataset . The results indicate that YiGen outperforms the state-of-the-art handwritten text generation methods in both ancient Yi script generation and restoration of tasks. With these advantages, the model facilitates dataset augmentation, damaged manuscript restoration and standardized digital archiving, offering a feasible technical solution for the digitization engineering of ancient Yi books.