lookx2/manga-colorization

Manga colorizer INT8 v1

A single ONNX model for automatic coloring of black-and-white manga. This is an INT8 conversion of the pretrained qweasdd/manga-colorization-v2 generator, without additional training. The reader predicts color at reduced resolution and restores original ink detail at the source resolution.

App handoff: CBOOK_READER_INTEGRATION.md. Machine-readable contract: model-manifest.json. Executable desktop reference: reference_colorize.py.

Model

Property	Value
File	manga-colorizer-int8.onnx
Size	32,536,990 bytes (32.54 MB / 31.03 MiB)
SHA-256	cbf5f19c289107f37de2add52b03c6ed713800a33ac2baf80ad3188c22c152f3
Format	ONNX IR 8, opset 17; self-contained weights
Quantization	Static QDQ, signed INT8 Conv weights/activations, per-channel weights
Input	sketch_and_hints, float32 [1,5,H,W]; grayscale [0,1] plus four zero hint channels
Output	rgb, float32 [1,3,H,W], RGB in [-1,1]
Mobile starting point	CPU, one page at a time; longest side 384, optional 576

INT8 describes internal quantization. The caller supplies and receives float32 tensors. No tokenizer, OCR model, denoiser, auxiliary checkpoint or server is required.

Measured behavior

On a Galaxy S25 Ultra (SM-S9380 / SM8750, Android 16), ONNX Runtime Android 1.29.0 CPU with four threads measured 0.52–0.64 seconds at a 384-pixel longest side and 1.09–1.15 seconds at 576 on two demonstration pages. These are warmed inference-only timings in a standalone native process, excluding the reader pipeline. Peak process RSS was approximately 482–513 MiB at 384 and 943–988 MiB at 576. A 30-run sequence slowed from a 1.04-second first-five median to 1.75 seconds for the last five. Midrange phones and the complete Flutter pipeline remain unmeasured.

See phone-benchmark-summary.json for the selected raw measurements. The model guesses colors; consistency between pages and canonical character colors are not guaranteed. Quantization errors and page-color quality require visual review. Calibration used four upstream demonstration pages at 384/576; two evaluation pages were excluded from calibration, but overlap with upstream training is unknown. There is no artist-colored ground-truth evaluation here.

Provenance and license status

• Upstream: qweasdd/manga-colorization-v2 at a0d0e4482e5e86ddbd49958475f3e95f282a1915.
• Original checkpoint: networks/generator.zip, downloaded from the author's linked Google Drive checkpoint.
• Checkpoint SHA-256: 087e6a0bc02770e732a52f33878b71a272a6123c9ac649e9b5bfb75e39e5c1d5.
• Main generator RGB output exported to ONNX and quantized with ONNX Runtime 1.29.0. No fine-tuning. The unused torchvision import was omitted in a separate inference copy of the upstream source; original source and checkpoint were unchanged.
• The inspected upstream revision has no top-level LICENSE file. Rights to redistribute the source/weights have not been established. No MIT or Apache license is asserted for this model. The existing OCR repository's MIT metadata does not establish the license of these third-party weights. Resolve the model's terms with its author before public redistribution; conversion does not supply a new license.

Upload layout

This folder is prepared for lookx2/cbook-manga-ocr; it has not been uploaded. Add the entire manga-colorization/ folder at the repository root. Keep the folder's README beside the model and keep the existing root OCR README. The app's intended URL after upload is:

https://huggingface.co/lookx2/cbook-manga-ocr/resolve/main/manga-colorization/manga-colorizer-int8.onnx

For an app release, replace main with the resulting commit SHA and validate both byte count and SHA-256 after download. The ZIP is a transfer bundle for the maintainer; the app downloads the .onnx file directly.

Add a short section to the existing root README so its MIT badge is not mistaken for the new model's license:

## Manga colorization

The optional INT8 colorizer is documented in [manga-colorization/README.md](manga-colorization/README.md).
Its upstream weight license is unestablished; the OCR assets' MIT declaration does not apply to these weights.