mayocream/comic-layout-yolo26s

Comic Layout YOLO26s

SafeTensors distribution of a YOLO26s instance-segmentation model for comic-page layout analysis. It detects panels, visible text regions, and speech or narration balloons for OCR and translation workflows.

Labels

ID	Label	Meaning
0	frame	Panel or frame region
1	text	Visible text region
2	balloon	Speech, thought, or narration container

The class order is checkpoint-sensitive and must not be changed.

Files

• model.safetensors: model state dictionary
• config.json: Koharu model and processor configuration
• yolo26s-seg.yaml: authoritative Ultralytics architecture configuration extracted from the checkpoint
• export-manifest.json: conversion hashes and validation results

The model uses a 1280-pixel inference size. Single-image prediction follows Ultralytics letterboxing with stride 32, confidence threshold 0.25, IoU threshold 0.7, and original-resolution masks.

Conversion

The MIT-licensed source checkpoint was published by ShadowB and trained with Ultralytics 8.4.43. This repository converts the original PyTorch archive to SafeTensors; it does not contain the pickle-based source checkpoint.

Property	Value
Source revision	3a860269ee0beb43ce9f31d82c7851441eb178ae
Source checkpoint SHA-256	0b4376e426fa96af3976afa6a2602421dacf2dec96ef87b4a44f5e8d4971cb6f
SafeTensors SHA-256	d577a9e66bcf4e3445c0ee799d8407b7f095cb94010745f8eca763c20583de66
Parameters	11,436,269
Tensors	844

Conversion validation performs all of the following:

1. Reloads every serialized tensor and checks exact equality.
2. Loads the state dictionary into a freshly constructed YOLO26s segmentation model with strict key and shape checking.
3. Compares all forward-output tensors on a deterministic input.

The recorded maximum absolute forward-output difference is 0.0.

Koharu

Koharu resolves config.json and model.safetensors directly from this repository. The public result contains the original image dimensions and a label, confidence, bounding box, area, and original-resolution binary mask for each instance.

License and provenance

The converted model is distributed under the upstream MIT license. Dataset terms remain separate from the model license. Validation metrics above are copied from the upstream checkpoint and have not been independently reproduced during conversion.