This repository hosts the compiled ONNX Runtime (.onnx) export of KoharuLayout-RFDETR-Seg-2XL scaled to 768x768 resolution for CPU inference and native embedded applications, such as XianScan.

Credits and Acknowledgments

All credit for the original model architecture, training, and supervision belongs to:

Model Creator: mayocream
Original Model Weights (SafeTensors): mayocream/koharu-layout-rfdetr-seg-2xl-1152
Original Project: Koharu (Manga Translation Studio)
Dataset: Manga109 Segmentation v2.0.0 (Manga109 academic-use terms apply)

We express our gratitude to mayocream for their open-source contributions to comic AI layout analysis.

Target Classes

The model predicts bounding boxes, class logits, and instance segmentation masks across 4 classes:

ID	Class	Description
0	`text`	Dialogue, captions, titles, credits, and general text
1	`onomatopoeia`	Comic sound effects and stylized text (SFX)
2	`bubble`	Speech and dialogue balloons
3	`panel`	Manga panel frames and layout boundaries

Export Details

File: rfdetr-seg-2xlarge.onnx (~134 MB)
Input Resolution: 768x768 (RGB, ImageNet normalized: mean [0.485, 0.456, 0.406], std [0.229, 0.224, 0.225])
Input Tensor: input [1, 3, 768, 768] float32
Output Tensors:
- dets [1, 300, 4] float32 (normalized [cx, cy, w, h] in [0, 1])
- labels [1, 300, 5] float32 (logits for [text, onomatopoeia, bubble, panel, background])
- masks [1, 300, 192, 192] float32
Positional Embeddings: 2D bicubic interpolated from 9,217 (96x96 patch grid) to 4,097 (64x64 patch grid)
Inference Latency: Approximately 1.7s to 3.5s on multi-threaded consumer x86_64 CPUs (roughly 3x lower latency than the 1152px baseline) while maintaining identical bubble and text detection counts
Runtime: ONNX Runtime (CPU, DirectML, CUDA, CoreML)

DevilishDaoSaint/koharu-layout-rfdetr

KoharuLayout-RFDETR-Seg-2XL (768px CPU-Optimized ONNX Runtime Export)

Credits and Acknowledgments

Target Classes

Export Details

DevilishDaoSaint/
koharu-layout-rfdetr