Koharu Layout HighRes
An RF-DETR Seg 2XL model for manga page layout, made for Koharu. It predicts boxes and instance masks for four classes.
| ID | Class | Meaning |
|---|---|---|
| 0 | text |
Dialogue, captions, titles, and other text that isn't a sound effect |
| 1 | onomatopoeia |
Sound effects (SFX) |
| 2 | bubble |
Speech and text balloons |
| 3 | panel |
Panels |
It succeeds koharu-layout-rfdetr-seg-2xl-1152. It does no OCR or reading order.
Files
model.safetensors: the weights.config.json: Koharu's configuration for the model.LICENSENOTICE
Results
COCO AP 50:95, scored against manga109-segmentation v3.0.0 at page resolution, with up to 160 detections per image. Each model was run at its own input. The 1152 model's numbers here differ from its own card, which scored it against v2.0.0's labels.
Test split, used for no training or selection decision (1,908 single pages and 969 spreads):
| View | Model | Box AP | Mask AP | Text box | SFX box | Text mask | SFX mask |
|---|---|---|---|---|---|---|---|
| Pages | 1152 model | 0.653 | 0.631 | 0.502 | 0.226 | 0.348 | 0.277 |
| Pages | this model | 0.824 | 0.701 | 0.852 | 0.556 | 0.536 | 0.365 |
| Spreads | 1152 model | 0.649 | 0.590 | 0.509 | 0.236 | 0.265 | 0.219 |
| Spreads | this model | 0.803 | 0.660 | 0.840 | 0.511 | 0.443 | 0.313 |
Bubble and panel AP are about the same for both models: box 0.92–0.95, mask 0.92–0.97.
Validation split (1,964 pages and 1,001 spreads):
| View | Model | Box AP | Mask AP | SFX box | Text mask | SFX mask |
|---|---|---|---|---|---|---|
| Pages | 1152 model | 0.669 | 0.626 | 0.220 | 0.354 | 0.227 |
| Pages | this model | 0.814 | 0.691 | 0.496 | 0.542 | 0.294 |
| Spreads | 1152 model | 0.659 | 0.584 | 0.206 | 0.267 | 0.174 |
| Spreads | this model | 0.790 | 0.651 | 0.437 | 0.447 | 0.250 |
Training
- Data: manga109-segmentation v3.0.0. Training images come from the Manga109 release and are not distributed here.
- Starting point: the 1152 model, fine-tuned in two stages at increasing input resolution.
- Hardware: 4 × NVIDIA H200.
Limitations
- Short lines in bubbles (「!」, 「で?」, a single word) are sometimes called onomatopoeia instead of text. The training labels put some sound effects inside bubbles, so the boundary there is genuinely fuzzy.
- The same lettering can be detected twice: as both text and onomatopoeia, or as pieces nested inside a whole sound effect.
- Character name tags and similar labels in colour web comics are mostly missed. Manga109 has none.
- Page numbers, running headers and copyright lines may be detected as text.
- In colour comics, decorations (sparkles, sweat drops, hearts) can be detected as onomatopoeia.
- Japanese manga only: the training data is Japanese manga. Other comic styles and scripts may behave differently.
License
Copyright © 2026 Mayo Takanashi. Licensed under the Apache License, Version 2.0.
NOTICE carries the notice for RF-DETR (Apache-2.0), from whose weights this model is fine-tuned. Manga109's images are not distributed here and remain under Manga109's terms.
Weight integrity
SHA-256 of model.safetensors:
03de2f77281974f5084cdd8652ace2b55dc15ac0378aed66654e62537cdda255