Datasetpublic

w46881/coloringzap-lineart-parameters

Photo-to-line-art parameter sweep (coloringzap) Review date: 2026-09-15 — every number here was measured by scripts/export-dataset.mjs, not estimated. Which settings turn a photo into a clean, printable line drawing? This dataset is the measured answer: 128 rows covering 32 parameter combinations across 4 synthetic image classes, plus the 8 test images that went in and came out. These measurements were produced by the same open-source pipeline that runs in the browser at… See the full description on the dataset page: https://huggingface.co/datasets/w46881/coloringzap-lineart-parameters.

sourceHugging Facecc0-1.0updated 11d agoView on Hugging Face

0likes57downloads

Dataset Card

Photo-to-line-art parameter sweep (coloringzap)

Review date: 2026-09-15 — every number here was measured by scripts/export-dataset.mjs, not estimated.

Which settings turn a photo into a clean, printable line drawing? This dataset is the measured answer: 128 rows covering 32 parameter combinations across 4 synthetic image classes, plus the 8 test images that went in and came out.

These measurements were produced by the same open-source pipeline that runs in the browser at https://coloringzap.com, published in full at https://github.com/w46881/coloringzap-algorithms.

Headline result

Best parameter set per image class:

image class	blur radius	sobel threshold	dilate	line coverage %
flat-vector (bold shapes, hard edges)	0	20	1	2.52
smooth-portrait (soft gradients)	0	20	0	0.00
detailed-texture (fine repeat)	3	90	0	23.55
noisy-lowlight (grainy)	1	90	0	4.02

Files

file	rows / count	what it is
`data/parameter-sweep-full.csv`	128 rows + header	every combination tested, with its score
`data/parameter-best-per-class.csv`	4 rows + header	the winning setting per image class
`data/REVIEW-DATE.txt`	1	the date these numbers were measured
`data/samples/sample-0X-input.png`	4 files	the synthetic test image, 400x300 grayscale
`data/samples/sample-0X-lineart.png`	4 files	what the winning parameters produce
`scripts/export-dataset.mjs`	—	regenerates all of the above (MIT)
`src/edge-detection.js`	—	the pipeline being measured (MIT)
`LICENSE-DATA.txt`	—	CC0 1.0 Universal, covers `data/`
`LICENSE-CODE.txt`	—	MIT, covers `scripts/` and `src/`

Columns in `parameter-sweep-full.csv`

column	meaning
`image_class`	short slug: flat-vector / smooth-portrait / detailed-texture / noisy-lowlight
`image_class_description`	what that class looks like
`blur_radius`	Gaussian blur radius applied before edge detection (0–3)
`sobel_threshold`	gradient magnitude needed to call a pixel a line (20/40/60/90)
`dilate_iterations`	how many times the edges were thickened (0 or 1)
`line_coverage_pct`	share of the image that became black line
`lone_speck_pct`	share of line pixels with no neighbouring line pixel (isolated dots)
`score`	`2 x lone_speck_pct` plus a penalty for coverage outside 1.5–12%; lower is better
`is_best_for_class`	`yes` on the winning row for that class

How "clean" was measured

—line coverage % — below ~1.5% the page is nearly blank; above ~12% it turns into a muddy grey blob.
—lone-speck % — isolated dots are what make a page look dirty.
—Winner = lowest 2 x lone-speck% + penalty(coverage outside 1.5–12%).

Known limits

1.Smooth gradients produce nothing. The smooth-portrait class returned 0.00% line coverage at all 32 combinations — soft shading has no local step for the Sobel operator to catch. No setting fixes this; it needs a different algorithm.
2.Busy photos stay muddy. detailed-texture only reached a printable state at blur 3 / threshold 90, and even then coverage was 23.55%, well above the ~12% ceiling.
3.The test images are synthetic, generated procedurally by the script — not real photographs, and not photos of any person. Treat the numbers as a starting point and re-run the script on your own images.
4.`lone-speck %` is 0.00 in every row. Do not read that as "the denoise step is perfect"; read it as "speckle was not the failure mode on these four images".

Reproduce

bash

node scripts/export-dataset.mjs

No dependencies, no network. Node 18 or newer.

License — two scopes, deliberately not mixed

—Data (data/*.csv, data/**/*.png) — CC0 1.0 Universal (public domain). Copy it, republish it, build on it, no attribution required.
—Code (scripts/, src/) — MIT.

Citation

coloringzap.com (2026). Photo-to-line-art parameter sweep (coloringzap).
Measured 2026-09-15. CC0 1.0 (data) / MIT (code).
https://coloringzap.com  |  https://github.com/w46881/coloringzap-algorithms