From isolated glyphs to realistic handwriting: diffusion-refined synthetic data generation for Ukrainian cyrillic text recognition

Abstract

Recognition of handwritten Cyrillic, and Ukrainian in particular, remains constrained by the lack of richly annotated handwriting corpora. This study treats the shortage of training data as the central bottleneck and investigates synthetic data generation as a response. A comparative evaluation of established optical and handwritten text-recognition systems on several Cyrillic handwriting benchmarks confirms that this task remains difficult and motivates a data-centric methodology. The main contribution is an end-to-end pipeline for generating synthetic Ukrainian handwriting. It constructs an isolated glyph bank from a purpose-designed full-alphabet collection form, filters anomalous samples in a learned representation space, extracts cleaned phrase content from a modern Ukrainian corpus, composes multi-word line images, applies conventional visual augmentation, and refines the resulting composites with a prompt-controlled general-purpose image-editing model to obtain more natural handwriting. The work further proposes an automatic verification stage to reduce the risk that visually plausible but textually incorrect samples enter a future training corpus. The architecture can combine glyphs from different writers within a line while preserving the target transcription and allowing controlled refinement of writing appearance. In the current study, the diffusion-refined extension is demonstrated qualitatively rather than through a full downstream training evaluation. The paper therefore presents a controllable architecture for Ukrainian handwriting synthesis, clarifies its limitations, and specifies the validation protocol required to establish its recognition value.