Koharu YOLO26s
Koharu YOLO26s is a unified comic-page instance-segmentation model. It detects and segments four classes:
| ID | Class |
|---|---|
| 0 | frame |
| 1 | dialogue_text |
| 2 | balloon |
| 3 | onomatopoeia_text |
The model was fine-tuned from
ShadowB/Manga109-panel-balloon-text-yolov26-segmentation
using MangaSeg annotations and locally licensed Manga109 page images. The
published repository contains no dataset images or annotations.
Files
model.safetensors: inference state dictionary; this is the only weight file.yolo26s-seg.yaml: YOLO26s segmentation architecture.config.json: class names and recommended image size.export-manifest.json: weight checksum and conversion validation.
The SafeTensors export contains 844 tensors for 11,437,172 parameters. It was
validated with strict state-dictionary loading and an exact forward comparison
against the training checkpoint (max_abs_diff = 0).
Loading
import json
from pathlib import Path
from huggingface_hub import snapshot_download
from safetensors.torch import load_file
from ultralytics import YOLO
root = Path(
snapshot_download(
"mayocream/koharu-yolo26s",
allow_patterns=["*.safetensors", "*.yaml", "*.json"],
)
)
config = json.loads((root / "config.json").read_text())
model = YOLO(root / "yolo26s-seg.yaml", task="segment")
model.model.load_state_dict(load_file(root / "model.safetensors"), strict=True)
model.model.names = {int(key): value for key, value in config["names"].items()}
results = model.predict("page.jpg", imgsz=1280, conf=0.25)
The checkpoint intentionally is not distributed as a PyTorch pickle (.pt).
Evaluation
Evaluation used a held-out 1,539-page, book-disjoint test split.
| Evaluation | Box mAP50 | Box mAP50-95 | Mask mAP50 | Mask mAP50-95 |
|---|---|---|---|---|
| Four classes | 0.879 | 0.795 | 0.830 | 0.422 |
| Shared three classes | 0.975 | 0.923 | 0.946 | 0.492 |
| Upstream checkpoint, shared classes | 0.932 | 0.877 | 0.741 | 0.380 |
The largest shared-class improvement is dialogue-text segmentation. Balloon mask AP remained approximately equal to the upstream checkpoint. The new onomatopoeia class reached box mAP50-95 0.414 and mask mAP50-95 0.236.
Training
- Input size: 1280 pixels
- Epochs: 41
- Training images: 7,532
- Validation images: 1,531
- Test images: 1,539
- Optimizer: MuSGD
- Best epoch: 36
- Ultralytics: 8.4.43
Terms and attribution
This checkpoint is a derived research artifact. Use and redistribution must comply with the terms of Manga109, MangaSeg, COO, the upstream checkpoint, and Ultralytics. Review those terms before commercial use or redistribution.