Show HN: Happy Hangul Day! An RNN for Generating Korean Handwriting Strokes

Commenters reacted positively to the RNN handwriting project and focused on practical and curious angles. One asked whether the model produced any weird or funny outputs during training, recalling past failures of machine translation and readers inventing content, and suggested an artistic use case: driving a real mechanical pen with the model to produce physical samples and photos. That commenter framed a tangible application and encouraged showing how generated strokes look when rendered by actual ink, implying such examples would make the work more compelling.

The author replied with concrete technical lessons from training: starting with a naive model that predicted the next x,y pen direction caused trouble at corners, so they adopted techniques from Alex Graves (2013). Specifically, they switched to a Mixture Density Network that predicts multiple (about ten) candidate directions and outputs gaussian distribution parameters, then samples to produce strokes; this change improved handling of turns. The author emphasized how much randomness factors into the model’s behavior and invited collaboration in Seoul, offering an email for anyone interested in helping produce physical pen samples.