kaonashi-tyc/zi2zi

Extends pix2pix with category embeddings and dual losses (from AC-GAN and DTN) to enable multi-font style transfer across CJK characters. Implements label shuffling during training to prevent discriminator collapse and improve generalization on unseen characters with minimal examples. Built on TensorFlow, it preprocesses font files into pickled binaries and supports inference with style interpolation and animation generation across font pairs.

2,717 stars. No commits in the last 6 months.