sebekblessing/manga-tl-models

Reading manga pages on a phone

The five models the Mihon translate feature uses to find and read Japanese on a page, converted to run on a Snapdragon's neural unit. They are only useful together, and only to that app — it looks for these exact filenames.

file	what it does
bubbles.tflite	finds the speech balloons
comictextdetector.tflite	finds the ink, and gives the mask used to wipe it
detector.tflite	finds the text regions
ocr_encoder.tflite	looks at one balloon
ocr_decoder.tflite	writes out what it says, a character at a time
ocr_vocab.txt	the characters it can write

Why these exist

The originals are ONNX and public. Getting them onto the neural unit took a conversion pipeline: simplify the graph, pin the input shapes, narrow 64-bit integers to 32 (the unit has no 64-bit), and rewrite the batched lookups it will not accept. These are the output of that, and they were checked against the originals page by page.

Built for Hexagon v73/v75/v79/v81. On anything else they fall back to the processor, which works but is roughly six times slower.