Vasyanator2/cleaned_webtoon_dataset

Licensing and source material

This repository contains derived/processed images from publicly accessible web comics/manga sources. The original works are copyrighted by their respective rights holders. No ownership of the original artwork is claimed.

The CC BY-NC 4.0 license applies only to the annotations and dataset metadata created by the dataset authors, not to the underlying original artwork.

Unpacking the dataset

The dataset ships as a zstd-compressed tarball split into parts named dataset.tar.zst.*.part. You just need to concatenate the parts back in order, decompress, and extract. (Parts use zero-padded suffixes, so the default sort order is the correct order.)

Linux / macOS

One clean pipe — reassemble, decompress, and extract without writing the combined archive to disk:

LC_ALL=C cat dataset.tar.zst.*.part | tar --zstd -xf -

Requires zstd and a tar with zstd support (GNU tar ≥ 1.31 or bsdtar).

Windows

Windows doesn't really have a cat that streams binary into a pipe. copy ... CON dumps to the console instead of the pipeline, type helpfully smears filename headers all over your data the moment it sees more than one file, and if you run the pipe through PowerShell it'll cheerfully re-encode your binary stream into confetti. So we skip the clever one-liner, glue the parts into a single file the boring way, and let tar handle the rest:

copy /b dataset.tar.zst.*.part dataset.tar.zst
tar --zstd -xf dataset.tar.zst
del dataset.tar.zst

Run this in cmd.exe, not PowerShell. tar is built into Windows 10/11. And yes, you'll need the temporary disk space for the reassembled archive — consider it the Windows convenience fee.

Cleaned Webtoon Dataset

A dataset of full webtoon episodes ("strips") prepared for training text-removal / inpainting models (e.g. ZITS++) and bubble/text detection models.

Each episode is a single very tall vertical strip (full reading width, tens of thousands of pixels high). For every strip we provide:

• the original image,
• a version with on-background text automatically erased (cleaned),
• the exact mask of what was erased,
• per-pixel confidence maps from speech-bubble and text detectors,
• bounding boxes for bubbles and text, and the list of clean, usable crops ("approved regions").

Episodes (chapters)	~1,011
Source titles	168
Approved regions (clean training crops)	~75,260
Strip width	full reading width (typically ~690–800 px)
Strip height	up to ~130,000 px

---

What's in each chapter folder

A chapter lives in one directory named <title_id>_<episode_no> (e.g. 724815_1). It contains 8 files:

file	mode	description
base.png	RGB	The stitched original strip, with text. Source image for inpainting.
cleaned.png	RGB	Same strip after autoclean: text sitting on a homogeneous background is erased (filled with the surrounding solid colour). Text inside bubbles / over complex art is left intact.
clean_mask.png	L (8-bit)	Mask of every pixel that autoclean filled (255 = erased, 0 = untouched). base and cleaned differ only where this mask is 255.
conf_bubble_seg.png	L	Per-pixel speech-bubble confidence map (non-binary, 0–255), full strip.
conf_text_seg.png	L	Per-pixel text confidence map from the segmentation model (0–255).
conf_text_ctd.png	L	Per-pixel text confidence map from the comic-text-detector (0–255).
info.json	—	Geometry + all boxes + approved regions (schema below).
visualization.png	RGBA	Human-readable overlay rendered over cleaned.png (see below). Provided for the example chapters; can be regenerated for any chapter.

All PNGs in one chapter share the same width and height, so masks, confidence maps and overlays are pixel-aligned with base.png / cleaned.png.

info.json schema

{
  "title_id": 724815,
  "episode_no": 1,
  "strip_width": 690,
  "strip_height": 62133,

  // Clean rectangular crops worth training on. Coordinates are strip-absolute.
  "approved_regions": [
    { "index": 1, "x0": 0, "y0": 1038, "x1": 690, "y1": 1710 }
  ],

  // Speech-bubble detections (both bubble models), strip-absolute boxes.
  "bubble_boxes": [
    { "model": "bubble_seg", "label": "speech_bubble", "score": 0.82,
      "piece_index": 1, "x1": 68, "y1": 1102, "x2": 141, "y2": 1211 }
  ],

  // Text detections (both text models), strip-absolute boxes.
  "text_boxes": [
    { "model": "text_seg", "label": "text", "score": 0.92,
      "piece_index": 1, "x1": 156, "y1": 1102, "x2": 428, "y2": 1148 }
  ],

  // Text the autoclean stage chose NOT to erase (non-homogeneous background:
  // inside a bubble, over art, etc.). Useful as "keep this text" supervision.
  "uncleaned_boxes": [
    { "x1": 163, "y1": 31000, "x2": 264, "y2": 31046 }
  ]
}

• approved_regions are the curated, content-bearing slices of the strip (dead margins, ad banners and blank gaps are excluded). index is the planning order top-to-bottom; (x0,y0)–(x1,y1) is the box. These ~75k regions are the intended training samples.
• boxes use (x1,y1)–(x2,y2); model tells you which detector produced it (bubble_seg/bubble_det for bubbles, text_seg/text_ctd for text); score is detector confidence; piece_index links a box to the approved region it fell in.

visualization.png legend

RGBA overlay rendered over cleaned.png; transparent outside approved regions.

• green — speech bubbles
• yellow — pixels autoclean erased (matches clean_mask.png)
• red — text that was not cleaned (the uncleaned_boxes, over the text masks)
• blue line — boundary where two approved regions touch

---

How to use it

Text-removal / inpainting (ZITS++ style)

For each chapter, the pair (base.png, cleaned.png) plus clean_mask.png is a ready (degraded → target, mask) triple:

• input = base.png (has text), target = cleaned.png (text gone),
• the inpainting mask is clean_mask.png (where text was removed).

Train on crops rather than full strips. Iterate approved_regions and crop each from base/cleaned/clean_mask:

import json, os
from PIL import Image

ch = "examples/724815_1"
info = json.load(open(os.path.join(ch, "info.json")))
base    = Image.open(os.path.join(ch, "base.png"))
cleaned = Image.open(os.path.join(ch, "cleaned.png"))
mask    = Image.open(os.path.join(ch, "clean_mask.png"))   # 'L', 255 = erased

for r in info["approved_regions"]:
    box = (r["x0"], r["y0"], r["x1"], r["y1"])
    inp  = base.crop(box)       # with text
    tgt  = cleaned.crop(box)    # text removed
    m    = mask.crop(box)       # inpainting mask for this crop
    # ... feed (inp, m) -> tgt

If you want to inpaint all text (not only the auto-removed part), build the mask from the text confidence maps / text_boxes instead of clean_mask.png.

Detection / segmentation

• Boxes: bubble_boxes and text_boxes give detector outputs per crop.
• Soft masks: conf_* maps are full-strip per-pixel confidences; threshold them (e.g. > 64/255) for binary masks, or use them as soft targets.
• uncleaned_boxes marks text that should be preserved (in-bubble / over art) — useful for a "removable vs. keep" text classifier.

Reassembling the full strip

The PNGs are aligned, so any per-pixel layer (mask, confidence) can be composited directly onto base/cleaned at native resolution. Strips are extremely tall; decode lazily / in bands if memory is a concern, and set PIL.Image.MAX_IMAGE_PIXELS = None.

---

examples/

Ten chapters from ten different titles, each including visualization.png, are provided under examples/ as a quick look at the format before downloading the full set.

---

How the data was produced

Strips were downloaded, stitched, and segmented into approved regions; an ensemble of open-weight detectors (two speech-bubble models + two comic-text models) produced the boxes and confidence maps. An autoclean stage then erased text only where the local background is homogeneous (solid-colour fill, never a guessed/hallucinated inpaint), leaving in-bubble and over-art text for the model to learn. Everything is pixel-aligned and deterministic; the cleaned image differs from the original only inside clean_mask.png.

Licensing & intended use

Source artwork is the property of the respective rights holders. This dataset is released for non-commercial research (text-removal, inpainting, and detection model training) only. It contains derived artefacts (masks, confidence maps, boxes) alongside the imagery; do not redistribute the imagery for commercial use. If you are a rights holder and want content removed, please open an issue.