SDT

Why we included this project

Handwriting generation is niche enough that most general text and image models ignore it, so a clean research implementation like SDT earns its place in any list. Built for the CVPR 2023 paper on the topic, the model separates a writer's overall hand from the way individual characters vary within it, and can produce online stroke sequences or offline rendered pages that actually follow the source style. The repo ships pre-trained weights, training and evaluation scripts, and datasets for Chinese, Japanese, and English, plus a tutorial for synthesizing text in a style you pick yourself. Later additions include Content Score and Style Score, two evaluation metrics that go beyond the usual visual checks. Researchers comparing style-disentanglement approaches and anyone prototyping personalized handwriting will both find a usable baseline here, and the MIT license keeps downstream reuse simple.