Self-attention Handwriting Generative Model

Recent advancements in font generation models include GANs, diffusion models, and transformers. This study introduces a GAN-based model, zi2zi self-attention, which enhances the zi2zi model by incorporating Residual Blocks (ResNet Block) and Self-Attention Layers in the encoder and decoder. These enhancements improve the ability to capture complex font details, mimicking the writer's style while preserving the source image's content. The study also introduces Perceptual Loss, using high-level features from the VGG-19 network to improve visual consistency and quality. The Adam optimizer with appropriate learning rates and Beta parameters is employed to enhance convergence speed and stability. For comparison, four other mainstream models are evaluated: MX-Font, CF-Font, SDT, and Font-Diffuser. These models differ from zi2zi in their focus on extracting content characters and writing style features rather than solely using adversarial networks. Experimental results show the potential and effectiveness of zi2zi-Self-Attention in generating high-quality, visually appealing handwritten