Ukrainian Handwritten Text Dataset and Cross-Domain Style Transfer Model

A new diffusion-based model has been created for producing handwritten Ukrainian text, filling a void in the generation of non-Latin script handwriting. The researchers compiled a dataset consisting of 126,177 images from 308 different writers, employing connected-component segmentation and quality filtering, while also focusing on oversampling less common Ukrainian characters. Named DiffusionPen, the model incorporates a MobileNetV2 triplet-loss style encoder along with a CANINE-conditioned latent diffusion U-Net, retrained on this dataset without any changes to its architecture. The research examines cross-domain style transfer across three scenarios: cross-lingual transfer from IAM English samples, zero-shot transfer, and fine-tuning, assessing the generalization capabilities of existing models beyond Latin scripts. The findings are available on arXiv under ID 2605.27487.