Rytability Magic Handwriting: handwriting synthesiser
The rewriting half of Rytability's Magic Handwriting ability. When a student taps Magic, Google ML Kit Digital Ink reads their pen strokes, the app's built-in Qwen 3.5 4B works out what they meant (spelling, their/there), and this network writes the corrected words back in the student's own handwriting style.
hand_synth.bin: float32 weights (little endian), 3.6M parameters.hand_synth.json: where each tensor starts in the .bin, plus the 13 built-in writing styles (pen-offset samples with their transcripts).
The network is Graves (2013) handwriting synthesis: three stacked 400-unit
LSTMs, a 10-component Gaussian attention window over the characters, and a
20-component bivariate mixture for each pen movement. Rytability runs it on the
CPU with Accelerate (HandwritingSynth.swift); writing 16 candidate lines takes
a fraction of a second.
Weights converted from the pretrained checkpoint of sjvasquez/handwriting-synthesis, which was trained on IAM-OnDB.
Google's ML Kit handwriting model is NOT in this repo: the app asks ML Kit to download it from Google, as ML Kit's terms intend.