manga-translator-models

The model weights the Rust Engine of manga-translator downloads on first run. The Engine pins every file by commit-pinned URL, sha256 and size in its compiled manifest; files here are immutable, and a new export gets a new filename.

File	What	sha256	Size (bytes)	Licence
comictextdetector.pt.onnx	Text detector, unchanged mirror	1a86ace74961413cbd650002e7bb4dcec4980ffa21b2f19b86933372071d718f	94669756	GPL-3.0
encoder.onnx	manga-ocr ViT encoder	73f84b2b826dea748c23453a56578667ab15f218a2eb68b52f28d5ad851a1779	344310126	Apache-2.0
cross.onnx	manga-ocr decoder cross-attention K/V	99a3528c4fc3f254eb35251f3eeab00f80ea91122b4e3abcf3673b3d657f25ec	9456788	Apache-2.0
step_reorder.onnx	manga-ocr decoder step with KV cache and beam reorder	7163328b45dadb80fdb9fb21d747d074a88bc989b1f3c35fecdd904e62783bb7	107928324	Apache-2.0
vocab.txt	manga-ocr tokenizer vocabulary, verbatim	344fbb6b8bf18c57839e924e2c9365434697e0227fac00b88bb4899b78aa594d	24072	Apache-2.0
lama_mat.onnx	LaMa (big-lama), dynamic H/W, matmul DFT	c170bd504ee71711ebc774816aa616ab9076b9f38230e460aeecd1590733ae82	208171227	Apache-2.0
cross_fp16.onnx	manga-ocr decoder cross-attention K/V, fp16 (fp32 I/O)	d827f51247b19e94dbe8a2b6e462af39636b901f66547170a2fbb562e0a7d4bf	4736642	Apache-2.0
step_reorder_fp16.onnx	manga-ocr decoder step, fp16 (fp32 I/O)	27e1c5e3b9787e7054cd825dca05bba0e04768ed89e5c3be484c85dfb7dd5eb4	54001635	Apache-2.0
lama_mat_fp16.onnx	LaMa (big-lama), fp16 body, fp32 I/O, sigmoid and composite	fa58a3b84b55445f06152935b94bc82c53ce7136bb609f6b7ddccf39a4a2a8a9	105530715	Apache-2.0
NotoSansCJK-Bold.ttc	Noto Sans CJK Bold, the typesetting fallback for CJK text, verbatim	faa5f3656a78b2e2d450d27fe8382c778bc2b6bb5ea29c986664a6a435056ceb	20050760	OFL-1.1
OFL-NotoSansCJK.txt	The SIL Open Font License text shipped with Noto Sans CJK

Since manga-translator #90 the Engine downloads the _fp16 files in place of cross.onnx, step_reorder.onnx and lama_mat.onnx. Those three fp32 files stay here unchanged for Engines pinned to commit 5b2793b.

Provenance

• Detector: comictextdetector.pt.onnx from zyddnys/manga-image-translator release beta-0.3, itself the ONNX export of dmMaze/comic-text-detector. Mirrored byte for byte. Both projects are GPL-3.0; their source is at the links above.

• manga-ocr: exported from kha-white/manga-ocr-base at revision aa6573bd10b0d446cbf622e29c3e084914df9741 (pytorch_model.bin sha256 c63e0bb5b3ff798c5991de18a8e0956c7ee6d1563aca6729029815eda6f5c2eb) by scripts/export/export_kv.py at manga-translator commit 497e3d8. fp32, a custom cross + step KV-cache split. manga-ocr is by Maciej Budyś (Apache-2.0).

• LaMa: exported from big-lama.pt (sha256 7ba7aa7ac37a4d41fdbbeba3a2af7ead18058552997e3a3cd1a3b2210c9e6b4c, as distributed by simple-lama-inpainting 0.1.0) by scripts/export/export_mat.py at manga-translator commit 497e3d8. The weights load strict into advimman/lama's FFCResNetGenerator, with FourierUnit replaced by a matmul DFT. fp32, dynamic height and width in multiples of 8. LaMa is by Suvorov et al. (Apache-2.0).

• fp16 (manga-translator #90): cross_fp16.onnx and step_reorder_fp16.onnx are cross.onnx / step_reorder.onnx (re-exported byte-identically) converted by onnxconverter-common 1.16.0 convert_float_to_float16(keep_io_types=True) in scripts/export/export_kv.py. lama_mat_fp16.onnx is exported by scripts/export/export_mat.py (default --precision fp16, with transforms in scripts/export/fp16.py) from the same big-lama.pt. BatchNorm statistics are folded into scale and shift. The hidden activations are scaled by 2^-10, which is exact. The generator body runs in fp16 up to the last conv; the sigmoid and the composite run in fp32. Script commits: LaMa ab82e44, decoder 563ae4f (manga-translator #90). The manga-ocr encoder stays fp32: in fp16 it changed 2 of 733 corpus texts.

• Noto Sans CJK (manga-translator #91): NotoSansCJK-Bold.ttc is copied byte for byte from notofonts/noto-cjk release Sans2.004, asset 03_NotoSansCJK-OTC.zip. OFL-NotoSansCJK.txt is that archive's LICENSE. The Engine uses face 0 as the fallback when the lettering font (embedded in the Engine binary) does not cover a block's text.

Every export matches the original PyTorch pipeline in manga-translator's parity suite (OCR text exact on all 94 corpus pages; fp32 LaMa within the R4 pixel bar on the CUDA and CPU execution providers). With the fp16 files, the OCR text is identical on all 733 corpus blocks. fp16 LaMa has no overflow on any of the 442 corpus crops, and every crop is at least 44 dB PSNR against fp32.