zi2zi

Why we included this project

Designing a full typeface is slow work, and for script styles like calligraphy or cursive it is effectively impossible to do by hand. zi2zi treats the problem as a conditional GAN: you give it a character in a standard font, and it draws that same character in the target style. A category embedding lets a single trained model switch between several styles, and beyond the baseline pix2pix setup it borrows AC-GAN's category loss and DTN's constant loss. The label-shuffling trick, enabled after the discriminator loss flattens, forces the model to keep generalizing and handles characters it never saw in training noticeably better. The author's write-up includes network diagrams and side-by-side gallery comparisons, which makes it a useful reference even though the codebase dates from 2017 and was built for research rather than ongoing production use.