DiffusionPen generates Ukrainian handwriting from 308 writer styles

Handwritten text generation (HTG) has mostly focused on Latin scripts, leaving low-resource writing systems like Cyrillic underserved. To close this gap, Andrii Ahitoliev and Pavlo Berezin built a comprehensive Ukrainian handwritten word dataset—126,177 images from 308 distinct writers—using connected-component segmentation, quality filtering, and targeted oversampling for underrepresented Ukrainian characters. They then retrained DiffusionPen, a state-of-the-art latent diffusion model originally designed for Latin HTG, on this new dataset without any architectural modifications. DiffusionPen uses a MobileNetV2-based triplet-loss style encoder to capture writer-specific traits and a CANINE-conditioned U-Net for character-level text guidance.

The results are promising: the model generates legible, style-consistent word images in three cross-domain settings: cross-lingual transfer from English IAM samples, zero-shot imitation of an early 20th-century Ukrainian manuscript, and few-shot adaptation to contemporary writers. The work demonstrates that few-shot latent diffusion models can generalize beyond the Latin-script domain, offering a reproducible benchmark for writer-aware Cyrillic HTG. The authors have released the dataset, trained models, and evaluation protocol on arXiv, providing a foundation for extending stylized text generation to other underrepresented writing systems.