depreciation84/comic-text-and-bubble-detector-int8

comic-text-and-bubble-detector (int8, Conv-only static quantization)

This repository contains an int8 version of the speech bubble and text detector ogkalu/comic-text-and-bubble-detector. The ToonLens Android app (comic translation) downloads this file for its "precise detection" mode.

Notice of modification (Apache-2.0, Section 4(b))

This file is a modified version of the original work.

• Original work: detector.onnx from ogkalu/comic-text-and-bubble-detector, revision 16e8a622f91fabc6b5b65c96d32d1183f8843546 (168,481,531 bytes, SHA256 065744e91c0594ad8663aa8b870ce3fb27222942eded5a3cc388ce23421bd195).
• Original author: ogkalu. Original license: Apache-2.0.
• Change: Team.R quantized the Conv operators to int8 (2026-10-26). All other operators (including the MatMul operators of the transformer decoder) stay in fp32. The model architecture, the inputs and the outputs did not change.
• Team.R did not train or fine-tune the model.

File

File	Size	SHA256
detector_rtdetr_int8c.onnx	67,055,821 bytes	d501da38f5faeac2c4fbfd5ea33f9f0b24471b22400d4c9f6bae307e7733d5a4

Quantization method

Item	Value
Tool	ONNX Runtime 1.30, onnxruntime.quantization
Pre-processing	quant_pre_process(..., skip_symbolic_shape=True)
Method	quantize_static, QuantFormat.QDQ
Operators	op_types_to_quantize=["Conv"] only
Types	activations QUInt8, weights QInt8, per_channel=True
Calibration data	21 synthetic comic images made by Team.R (the first 3 images of each sample set). No third-party images.

When Team.R also quantized the MatMul operators, the scores of small speech bubbles became lower. For this reason, the MatMul operators stay in fp32.

Inputs and outputs

The inputs and outputs are the same as in the original model.

Name	Type	Shape	Meaning
input images	float32	N×3×640×640	RGB, values 0–1 (pixel / 255). The app resizes the full page to 640×640 (bilinear, no letterbox).
input orig_target_sizes	int64	N×2	Original [width, height] of each image
output labels	int64	N×300	Class ID: 0 bubble, 1 text_bubble, 2 text_free
output boxes	float32	N×300×4	[x0, y0, x1, y1] in pixels of the original image
output scores	float32	N×300	Confidence, 0–1

The ToonLens app keeps the boxes with a score of 0.3 or more and class 1 or 2. Then the app removes overlapping boxes.

Measured results (Team.R, for reference)

Model	Size	PC time (i5-1335U, 4 threads)	Galaxy Z Fold7 detection time	Dialogue boxes found (65 test images)
Original fp32	168.5 MB	837 ms	1944 ms	369 of 373
This int8 file	67.1 MB	455 ms	802 ms	369 of 373

The test images are 50 synthetic images and 15 real comic pages. Team.R did not use these 15 pages for calibration.

Use in the ToonLens app

• The app downloads this file only when the user turns on "precise detection".
• The app checks the size and the SHA256 of the file before it uses the file.
• The app shows the Apache-2.0 notice and this modification notice on its information screen.

License

Apache License 2.0. See LICENSE. The original work is also under Apache-2.0: ogkalu/comic-text-and-bubble-detector.