Vasyanator2/cleaned_manga_dataset

Licensing and source material

This repository contains derived/processed images from publicly accessible manga sources (GANMA! and Comic Walker / カドコミ). 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 single zstd-compressed tarball split into parts named dataset.tar.zst.part-NNN (part-000, part-001, …). You just need to concatenate the parts back in order, decompress, and extract. The suffixes are zero-padded, so the default lexical 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). If your tar lacks --zstd, pipe it yourself:

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

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.

Extraction yields a dataset/ directory with one folder per chapter (see below).

Cleaned Manga Dataset

A dataset of manga chapters prepared for training text-removal / inpainting models (e.g. ZITS++) and bubble/text detection models. It is the page-oriented sibling of the Cleaned Webtoon Dataset: the unit here is a single manga page, not a tall stitched webtoon strip. Every page is downloaded, analysed and saved on its own.

Pages come from two sources on equal footing:

• GANMA! (ganma.jp) — free stories, standard page images.
• Comic Walker (comic-walker.com / カドコミ) — free episodes; pages are XOR-scrambled WEBP and are descrambled on download.

For every page we provide:

• the original page image,
• a version with on-background text automatically erased (cleaned),
• the exact mask of what was erased,
• a binary mask of "uninteresting" (flat gutter / empty margin) areas,
• bounding boxes for speech bubbles and text, and the boxes for text that was deliberately left in place.

Chapters	~2,335
Source titles	~533
Pages	~43,217
Sources	Comic Walker ~2,133 · GANMA ~179 chapters
Text boxes	~537,000
Speech-bubble boxes	~488,000
Page size	full page resolution, typically ~1,000–2,100 px wide, up to ~7,300 px tall (double-page spreads)

---

Dataset layout

dataset/
  <key>/                     one directory per chapter
    info.json                chapter metadata + per-page geometry & boxes
    page_0.png               page 0, original (with text)
    page_0_cleaned.png       page 0, autocleaned (text on flat bg erased)
    page_0_clean_mask.png    page 0, mask of erased pixels
    page_0_uninteresting.png page 0, binary mask of uninteresting areas
    page_1.png
    page_1_cleaned.png
    ...

The chapter directory <key> encodes the source:

• GANMA: <alias>_<storyId> (e.g. akagaminohanayome_2ca52150-68ee-11f0-8e03-96e8f2a966e0)
• Comic Walker: cw_<workCode>_<episodeId> (e.g. cw_KC_000040_S_018d6a9e-4072-73a5-812f-3244507f14dd)

What's in each chapter folder

A chapter contains one info.json plus 4 PNGs per page N (0-based, matching the page_N file prefix):

file	mode	description
page_N.png	RGB	The original page, with text (alpha composited over white). Source image for inpainting.
page_N_cleaned.png	RGB	Same page 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.
page_N_clean_mask.png	L (8-bit)	Mask of every pixel that autoclean filled (255 = erased, 0 = untouched). page_N and page_N_cleaned differ only where this mask is 255.
page_N_uninteresting.png	L (8-bit)	Binary (0/255) mask of uninteresting areas — flat gutters, empty margins, solid blank regions (255 = uninteresting). Useful to skip dead space when sampling training crops.
info.json	—	Chapter metadata + per-page geometry and all boxes (schema below).

All four PNGs for one page share the same width and height as that page's page_N.png, so masks are pixel-aligned with the original and the cleaned image. Different pages within a chapter can have different dimensions.

Optional debug artifacts. By default only the four files above are written. Re-running the builder with --debug (ganma_rs --reprocess <chapter_dir> --debug --ai-detect-socket …) additionally emits per-model confidence maps page_N_conf_<model>.png (bubble/text detector soft masks, 0–255) and an page_N_uninteresting_overlay.png preview. They are not shipped here to keep the archive small; they can be regenerated for any chapter.

info.json schema

{
  "source": "ganma",                 // "ganma" or "comicwalker"
  "magazine_id": "e9fa8129-…",        // GANMA magazine / Comic Walker work id
  "alias": "akagaminohanayome",       // GANMA alias (Comic Walker: work code)
  "magazine_title": "赤髪の花嫁は…",    // series title
  "story_id": "2ca52150-…",           // chapter / episode id
  "story_title": "第3話",             // chapter title
  "subtitle": "後編",

  "pages": [
    {
      "index": 0,
      "page_number": 1,               // 1-based reading order
      "width": 1975,
      "height": 2783,

      "base": "page_0.png",
      "cleaned": "page_0_cleaned.png",
      "clean_mask": "page_0_clean_mask.png",
      "uninteresting_mask": "page_0_uninteresting.png",

      // Speech-bubble detections (both bubble models), page-absolute boxes.
      "bubble_boxes": [
        { "model": "bubble_seg", "label": "speech_bubble", "score": 0.96,
          "piece_index": 0, "x1": 312, "y1": 1321, "x2": 718, "y2": 1751 }
      ],

      // Text detections (both text models), page-absolute boxes.
      "text_boxes": [
        { "model": "text_seg", "label": "text", "score": 0.97,
          "piece_index": 0, "x1": 361, "y1": 1342, "x2": 659, "y2": 1707 }
      ],

      // 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": 361, "y1": 1342, "x2": 659, "y2": 1707 }
      ]
    }
  ]
}

• boxes use (x1,y1)–(x2,y2) in page-absolute pixels; 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 is always 0: a manga page is a single piece (the field exists for schema parity with the strip-based webtoon dataset, where a strip is cut into several pieces).
• uncleaned_boxes mark text that autoclean intentionally left in place (in-bubble text, text over busy art) — good "removable vs. keep" supervision.

---

How to use it

Text-removal / inpainting (ZITS++ style)

For each page, the pair (page_N.png, page_N_cleaned.png) plus page_N_clean_mask.png is a ready (degraded → target, mask) triple:

• input = page_N.png (has text), target = page_N_cleaned.png (text gone),
• the inpainting mask is page_N_clean_mask.png (where text was removed).

import json, os
from PIL import Image
Image.MAX_IMAGE_PIXELS = None

ch = "examples/akagaminohanayome_2ca52150-68ee-11f0-8e03-96e8f2a966e0"
info = json.load(open(os.path.join(ch, "info.json")))

for pg in info["pages"]:
    base    = Image.open(os.path.join(ch, pg["base"]))            # with text
    cleaned = Image.open(os.path.join(ch, pg["cleaned"]))         # text removed
    mask    = Image.open(os.path.join(ch, pg["clean_mask"]))      # 'L', 255 = erased
    dead    = Image.open(os.path.join(ch, pg["uninteresting_mask"]))  # skip these areas
    # ... crop content-bearing regions (avoid `dead`) and feed (base, mask) -> cleaned

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

Detection / segmentation

• Boxes: bubble_boxes and text_boxes give per-page detector outputs.
• uncleaned_boxes marks text that should be preserved (in-bubble / over art) — useful for a "removable vs. keep" text classifier.
• Soft masks (per-model confidence maps) are not shipped by default; regenerate them with the builder's --debug flag if you need soft targets.

Skip the page_N_uninteresting.png == 255 areas (flat gutters / empty margins) when sampling crops so training does not waste capacity on blank space.

---

examples/

Ten chapters from ten different titles — six from GANMA and four from Comic Walker — are provided under examples/ as a quick look at the format before downloading the full set. They are the real default artifacts (no debug confidence maps).

---

How the data was produced

Pages were downloaded (Comic Walker pages XOR-descrambled after download) and each page was segmented by an ensemble of open-weight detectors (two speech-bubble models + two comic-text models), producing the boxes. 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 (recorded in uncleaned_boxes). A separate pass detected "uninteresting" flat/empty areas. Everything is pixel-aligned and deterministic; the cleaned image differs from the original only inside clean_mask.

Pages are processed independently — there is no strip stitching or dead-gutter band planning (those are webtoon-strip concepts). See ../README.md and ../src/ for the builder (ganma_rs).

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, 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.