FontTransformer: Few-shot high-resolution Chinese glyph image synthesis via stacked transformers

Automatic generation of high-quality Chinese fonts from a few online training samples is challenging task, especially when the amount very small. Existing few-shot font methods can only synthesize low-resolution glyph images that often possess incorrect topological structures or/and incomplete strokes. To address problem, this paper proposes FontTransformer, novel learning model, for high-resolution image synthesis by using stacked Transformers. The key idea to apply parallel Transformer avoid accumulation prediction errors and utilize serial enhance quality synthesized Meanwhile, we also design encoding scheme feed more information prior knowledge our which further enables visually-pleasing images. Both qualitative quantitative experimental results demonstrate superiority method compared other existing approaches in task.