| Humans can quickly learn handwriting styles from only a few examples and naturally apply them to new words while keeping both readability and stylistic consistency. Achieving a similar capability in generative models remains challenging, especially when synthesizing variable-length handwritten text with realistic stroke structure and natural style variations. Inspired by this, we investigate the integration of attention mechanisms into a style-controlled handwritten text generation network. Our approach focuses on improving feature interaction inside the generator by inserting different attention modules at various depths of the architecture. This allows the model to better capture both global text structure and fine local stroke details. Experiments on the IAM Handwriting Database show improved visual realism, distribution similarity and text readability compared to the baseline model. Qualitative evaluation further confirms more stable ligatures, consistent stroke thickness and more natural handwriting variability. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.