Waheed786dar/Comiman-Dataset

Comiman Dataset

Attribution is required for every use: Comiman Dataset by Waheed (huggingface.co/Waheed786dar) - https://huggingface.co/datasets/Waheed786dar/Comiman-Dataset

A license-gated comics and manga page corpus built for training a model that can plan and draw full comic/manga series (the planned model: Waheed786dar/Comiman). Every book passed an automatic license gate (Creative Commons / CC0 / Public Domain Mark metadata, or a public-domain claim limited to works published up to 1963) and every page passed quality, NSFW, blank-page and duplicate filters. Every accept/reject decision is public in ledger/.

Dataset summary

Books	193
Pages	50,467
Panels (heuristic)	93,816
Image size on disk (WebP)	5.58 GB
Mean book rating	73.51 / 100
Splits	train / validation / test (deterministic, by book, 96/2/2)
Last build	2026-10-03

By license type

license_type	books
public_domain_claimed	193

By language (as tagged by the source)

language	books
eng	89
fre	74
ger	13
spa	9
rus	2
dut	2
English	1
por	1
dan	1
ita	1

By decade

decade	books
1900	41
1890	27
1880	26
1910	26
1850	24
1870	21
1840	10
1860	6
1920	5
1830	2
1790	1
1720	1
1700	1
1950	1
1780	1

By genre tag (keyword-derived, multi-label)

genre	books
humor	188
war	15
unclassified	4
children	2
funny_animal	1
romance	1

By quality tier

tier	books
B	127
A	43
C	23

By reading direction

direction	books
ltr	193

Curation decisions (ledger)

decision	books
accepted	193
rejected_year	23
rejected_low_quality	2
error_process	1

Supported tasks

Comic page / panel generation, layout and panel-order modeling, page-type and quality classification, retrieval with the included CLIP embeddings, panel detection (COCO json), and as the visual stage of a story-to-comic pipeline. Not included yet: OCR text, captions, speaker labels (planned GPU stage 2).

Dataset structure

data/{train,validation,test}-NNNNN.parquet        page rows (images as WebP bytes + annotations)
webdataset/{train,validation,test}-NNNNN.tar      same pages as WebDataset tar (key.webp + key.json)
coco/{split}-<tag>.json                           panel boxes in COCO format (absolute pixels)
manifest/books-<tag>.parquet                      one row per book
ledger/ledger-<tag>.parquet                       every accept/reject decision with reason
state/page_hashes-<tag>.parquet                   perceptual hashes (duplicate detection / resume)
integrity/files-<tag>.json                        sha256 + row counts of every shard
exports/books.jsonl, books.csv, ledger.csv        convenience exports

Page fields

field	type	meaning
book_id	string	stable id cmn-<sha1[:12]> of the source id
page_no	int32	reading page index inside the book (after filtering)
split	string	train / validation / test (by book, never split inside a book)
width, height	int32	stored image size (longest side <= 1600)
orig_width, orig_height	int32	size of the source scan before crop/resize
aspect_ratio, is_spread	float32, bool	width/height; spread = aspect > 1.15
cropped	bool	scan borders/margins auto-cropped
color_mode	string	color / grayscale / bw
reading_direction	string	ltr or rtl (manga, Arabic, Hebrew, Persian, Urdu)
image_webp	binary	WebP-encoded page
webp_sha256	string	checksum of image_webp
dhash	uint64	64-bit perceptual hash
panel_count, panel_boxes	int16, list	heuristic panel boxes [x0,y0,x1,y1] normalized 0-1, reading order
sharpness, contrast, brightness, ink_ratio	float32	classical image metrics
aesthetic	float32	LAION aesthetic head on CLIP ViT-L/14 (about 1-10; null if GPU model unavailable)
nsfw_score	float32	probability from an NSFW classifier (pages above 0.85 were dropped)
page_type, page_type_conf	string, float32	zero-shot CLIP: cover / story / ad / text_page / blank / back_cover
quality_score, quality_star	float32, int8	page rating 0-100 and 1-5 stars
clip_l14_fp16	binary	L2-normalized CLIP ViT-L/14 embedding, 768 x float16

Book fields (manifest)

book_id, source_id, split, title, creator, year, decade, language, subjects, genre_tags, reading_direction, license_type, license_confidence, license_url, rights_text, copyright_status, attribution_required, share_alike, attribution_text, source, source_url, file_used, page_count, pages_dropped, story_pages, webp_bytes, avg_panels, mean_quality, mean_aesthetic, mean_sharpness, mean_nsfw, book_rating, rating_stars, quality_tier, added_at

Quality ratings (how they are computed)

Page quality (0-100) = weighted mean of: resolution of the source scan (25%), sharpness via Laplacian variance (25%), contrast (15%), ink coverage sanity (10%), panel structure found (10%), aesthetic head (15%, when available; weights are renormalized otherwise). Stars: >=85 five, >=70 four, >=55 three, >=40 two, else one.

Book rating (0-100) = 0.60 x mean page quality + 0.15 x completeness (kept / kept+dropped) + 0.10 x license confidence (high 100, medium 60) + 0.15 x share of pages with 2+ panels. Tier A >= 80, B >= 65, C >= 50, D otherwise. Books below 35.0 are not published.

How the data was built

1. Discovery: Internet Archive search by comics/manga subjects plus license/copyright metadata; optional user-supplied archives listed in licenses.json.
2. License gate: CC0/PDM/CC BY/CC BY-SA accepted; NC/ND rejected; "public domain" claims accepted only with a publication year <= 1963; everything else is rejected and logged.
3. Extraction: CBZ, CBR (via libarchive) and PDF (embedded scan or rendered at 130 dpi), natural page order.
4. Cleaning: junk/tiny images removed, scan-border auto-crop, resize to 1600 px, WebP q82, blank-page removal, perceptual-hash duplicate detection (book dropped when >= 0.8 of pages already exist).
5. Annotation: panel boxes (OpenCV heuristic), image metrics, CLIP ViT-L/14 embeddings, aesthetic score, page-type, NSFW score on GPU (T4 x2).
6. Validation before every upload: row counts, schema, image checksum + decode sample, tar member count, upload path whitelist; after upload remote file sizes are compared with local sizes.

Usage

from datasets import load_dataset
import io, numpy as np
from PIL import Image

ds = load_dataset("Waheed786dar/Comiman-Dataset", split="train", streaming=True)
row = next(iter(ds))
img = Image.open(io.BytesIO(row["image_webp"]))
emb = np.frombuffer(row["clip_l14_fp16"], dtype=np.float16)       # (768,)

# only high quality story pages
good = ds.filter(lambda r: r["quality_star"] >= 4 and r["page_type"] in (None, "story"))

WebDataset: load_dataset("webdataset", data_files="hf://datasets/Waheed786dar/Comiman-Dataset/webdataset/train-*.tar", split="train", streaming=True).

Book-level filtering: read manifest/ (or exports/books.csv), keep quality_tier in ("A","B"), then select pages by book_id.

Licensing and attribution

Compilation layer (metadata, ratings, boxes, embeddings, docs): CC BY 4.0 - attribution to Waheed is required (see LICENSE).

Underlying page images keep their original terms (license_type per book). Public-domain works: no copyright is claimed by the compiler.

CC BY books: keep attribution_text. CC BY-SA books (share_alike = true): share-alike applies.

Trained-model credit is requested ("Trained on the Comiman Dataset by Waheed"); strict enforcement against model weights is legally unsettled, so treat this as a condition of use for the data and a request for the model.

Considerations

License verification is metadata-level. Uploaders on source sites can be wrong. Public-domain status of old comics is a United States judgment (for example non-renewal) and may differ in your country. Report problems for takedown.

Historical content can contain outdated and offensive depictions. Review before training or serving.

Bias/coverage: strongly skewed to mid-century Western comics; little modern manga because modern manga is almost always copyrighted.

Heuristics: panel boxes, genre tags and reading direction are heuristic, not human-verified.

NSFW filter is a classifier and will miss some content and wrongly flag some art.

No personal data is collected; creators' names come from public archive metadata.

Citation

@misc{comiman_dataset,
  title  = {Comiman Dataset},
  author = {Waheed},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Waheed786dar/Comiman-Dataset}
}

Contact / takedown

GitHub issues: https://github.com/uzairlovesM/comiman-dataset-reports