Models v3 — DBNet detector

偵測器換 DBNet（m-i-t default detector, ResNet34 + DB head）取代 comic-text-detector：真機讀對率 1.6-2.5×、密集/手寫文字涵蓋更全。OCR（ocr_int8）與去字（mit_aot）沿用 models-v2。

Models v2 — NCNN + int8

Yakuyomi engine model set v2 — NCNN detector + int8 OCR + NCNN AOT-GAN inpaint.

Replaces v1 (ORT detector + fp32 OCR + LaMa): ~2.9x faster detection, ~3.6x faster OCR on ARM, ~470MB → ~92MB total, GPU/NPU-ready (fixed-shape NCNN, not yet enabled).

NCNN roles ship as .param + .bin pairs (both required). All weights GPL-3.0. Consumed via models.json auto-download; see docs/MODELS.md.

Models v1

The three ONNX models Yakuyomi needs (detection / OCR / text removal). The reader can auto-download these and verifies each file's sha256 against models.json. See docs/MODELS.md for details.

File	bytes	sha256
comictextdetector.pt.onnx	94669756	1a86ace74961413cbd650002e7bb4dcec4980ffa21b2f19b86933372071d718f
ocr_48px_ctc.onnx	164974063	3019b406dfc7b3dedf3d0a17a8b5f78c4b483712e3acaa3987db078d9b2c35d8
lama-manga.onnx	207482644	4512adab295ee5a5e02ccd1bdf8d45dccbac88309d9cff1532ffd5de876f02a4

Weights are GPL-3.0 (base LaMa architecture Apache-2.0), redistributed with attribution:

• comic-text-detector (dmMaze) — GPL-3.0
• 48px CTC OCR — exported here from manga-image-translator weights — GPL-3.0
• lama-manga (Koharu / mayocream) — GPL-3.0