Disentangling Writer and Character Styles for Handwriting Generation

We organize our supplementary material as follows. • In Sec. A.1, we provide more implementation details. • In Sec. A.2, we provide more visualization examples for spectrum analysis of two style representations. • In Sec. A.3, we describe additional related works about handwriting generation and review the works in font generation. • In Sec. A.4, we provide qualitative results of offline Chinese handwriting generation with a comparison to previous state-of-the-art works. • In Sec. A.5, we study the effect of the sampling ratio α and compare different combination strategies in decoder based on online Chinese handwriting dataset. • In Sec. A.6, we provide the discussions on the format of style inputs. • In Sec. A.7, we conduct failure case analysis. • In Sec. A.8, we report more evaluation metrics on Japanese dataset. • In Sec. A.9, we conudct more experiments on Indic dataset. • In Sec. A.10, we give detailed data representations of online characters. • In Sec. A.11, we describe more details about the pen moving prediction and pen state classification losses. • In Sec. A.12, we show a large number of generated online samples, covering Chinese, Japanese, Indic and English scripts.