Validated FP32 ONNX export of mayocream/koharu-layout-rfdetr-seg-2xl-1152. It predicts bounding boxes and instance masks for manga page layout analysis.

ID	Class
0	`text`
1	`onomatopoeia`
2	`bubble`
3	`panel`

The model does not perform OCR or reading-order prediction.

Files

rfdetr-seg-2xlarge.onnx — static-batch FP32 ONNX graph, opset 17
onnx_config.json — tensor, preprocessing, class, and threshold contract
onnx_validation.json — PyTorch/ONNX Runtime parity report

ONNX Runtime

pip install huggingface_hub numpy onnxruntime pillow

python

from huggingface_hub import hf_hub_download
from PIL import Image
import numpy as np
import onnxruntime as ort

repo_id = "ShiniShiho/koharu-layout-rfdetr-seg-2xl-1152-onnx"
model_path = hf_hub_download(repo_id=repo_id, filename="rfdetr-seg-2xlarge.onnx")
session = ort.InferenceSession(model_path)

image = Image.open("page.jpg").convert("RGB")
array = np.asarray(
    image.resize((1152, 1152), Image.Resampling.BILINEAR),
    dtype=np.float32,
) / 255.0
array = (array - np.asarray([0.485, 0.456, 0.406], dtype=np.float32)) / np.asarray(
    [0.229, 0.224, 0.225], dtype=np.float32
)
input_tensor = array.transpose(2, 0, 1)[None].astype(np.float32)

boxes_cxcywh, class_logits, mask_logits = session.run(
    ["dets", "labels", "masks"],
    {"input": input_tensor},
)
class_scores = 1.0 / (1.0 + np.exp(-class_logits[..., :4]))
class_ids = class_scores.argmax(axis=-1)
confidences = class_scores.max(axis=-1)

The graph has a fixed float32 NCHW input shape of [1, 3, 1152, 1152]. dets contains 300 normalized cx, cy, width, height boxes. labels contains four class logits plus a final no-object channel; apply sigmoid to the first four channels. masks contains 300 masks at 288 × 288; resize mask logits bilinearly to the target image size and threshold at zero.

Recommended confidence thresholds are 0.25 for text, 0.20 for onomatopoeia, and 0.50 for bubbles and panels.

Validation

The graph passed ONNX's full model checker and was compared with the source PyTorch graph on a deterministic input:

top-class agreement: 100%
maximum confidence drift: 2.25e-5
global thresholded-mask IoU: 0.99979
source weights SHA-256: 9bf6d2cbd7793c956d8c857bb1672a396eb7f100eb0682f86830d05e31168efb
ONNX SHA-256: 7cc10d4316371946b8441da3512261a8e148b129abcdb0ea6235ed1d1d06d351

The export uses rfdetr==1.7.0, ONNX opset 17, and a static batch size of 1. See onnx_validation.json for environment and raw-output error statistics.

License and data terms

RF-DETR software is Apache-2.0. The source model was fine-tuned using Manga109 images, which are distributed separately under Manga109's academic-use terms. This repository does not redistribute Manga109 images. The model therefore retains the source repository's custom/other license designation; users are responsible for complying with Manga109 and third-party content terms.