PRACTICAL GUIDE
A line-art model simplifies tones and textures into visible contours. The best result is not necessarily the one with the most lines. Useful line art preserves recognizable structure, separates important shapes and avoids turning noise, hair or foliage into an unreadable web.
Use a source with a distinct subject, readable silhouette and controlled background. Strong overlap, low contrast and heavy motion blur make it difficult to decide which boundaries matter. Cropping around the subject reduces irrelevant detail.
Fine fabric, skin texture, leaves and compressed JPEG blocks can produce excessive marks. A lightly cleaned or resized source may create a clearer drawing. Avoid aggressive sharpening before conversion because every sharpened artifact can become a line.
A tracing reference can retain internal construction lines, while a children's coloring page needs larger enclosed regions and fewer marks. Tattoo concepts, craft templates and technical references each require different levels of simplification. Print a small test or view the image at final size: if important shapes merge or texture dominates the subject, simplify the source and convert again rather than expecting every generated contour to remain useful.
For portraits, inspect the eyes, nose, mouth, jaw and hairline. For objects, check corners, openings and perspective lines. Remove or redraw misleading contours before using the output as a painting reference or educational worksheet.
A style conversion does not remove copyright or personality rights. Confirm permission before distributing line art derived from somebody else's photograph, illustration, character or product design.
LINE-ART LAB · SEPTEMBER 7, 2026
The earlier portrait comparison is a single source/output composite. These two new cases publish separate source photographs and untouched PNG exports, so readers can repeat the complete transformation. A coffee service tests manufactured contours against wood grain and reflections. A cat tests whether important facial landmarks survive when fur supplies far more edges than the eyes do.
Both images were uploaded through the visible Photo to Line Art interface in the Codex in-app browser on macOS arm64, using a local static preview at 127.0.0.1:4173/linechange/. A long-running busy-status fault had first been repaired: the old percentage display could crash while waiting. The repair changed the status interface, with the same sketch model, worker, normalization and image conversion retained.
Each run finished with a visible result and a PNG saved through the download button. No mask, density setting, recoloring or output cleanup was applied. These are local-preview observations; the record does not establish production-origin behavior. The session included other work, so its elapsed window is not presented as processing speed.
Photo: Rachel Michetti · Coffee photograph courtesy of Pikolo Espresso Bar · CC0, as documented by scikit-image. The linked source is unchanged; the output is the newly recorded browser export.
The cup opening, handle and elliptical saucer remain immediately recognizable. Large pale interiors separate them from much of the scene. This supports using the export as a first structural tracing reference: the main objects have not dissolved into unrelated marks.
The difficult part is deciding which lines deserve to remain. Diagonal wood grain becomes a dense competing field, particularly at the left and lower-right edges. The saucer has several adjacent rim contours. Bright reflections on the spoon become nested fragments inside its outline, while spots on the handle become isolated loops. Those marks follow visible contrast, but they are not necessarily construction lines. Reject the raw export as a clean coloring page; select the main silhouette and remove background and reflection-derived clutter before that use.
Photo: Stefan van der Walt · Chelsea the cat · CC0, as documented by scikit-image. The linked source is unchanged; the output is the newly recorded browser export.
The nose, eye surrounds and broad face structure survive. Forehead stripes and fine fur become strong, closely packed marks. Several long whiskers extending toward image-left also remain visible, but smaller strands break into fragments. The source already crops the ears and body, so incomplete framing is not a model reconstruction failure.
The pupils expose a more consequential loss. The image-left pupil becomes a thin internal contour instead of a dark filled shape; the image-right pupil largely disappears into the white eye interior. The cat is still recognizable, yet its gaze changes. For a portrait brief, that is a rejection condition even if the fur looks detailed. Restore the pupils from the original and choose meaningful fur lines before accepting the drawing. We did not perform that repair on the downloadable result.
| Case | Source pixels | Output pixels | Near-white pixels | Non-gray pixels |
|---|---|---|---|---|
| Coffee | 600 × 400 | 1152 × 768 | 82.90% | 3.17% |
| Chelsea the cat | 451 × 300 | 1152 × 768 | 76.08% | 6.78% |
Both exports are 1152 × 768 pixels. The reviewed preprocessing sets the short edge to 768 and rounds the scaled long edge down to a multiple of eight; these files match that rule. Their larger dimensions do not establish recovered detail. Source and output sizes differ, so a direct pixel-error score would require additional alignment choices that this review does not make.
“Near-white” means every RGB channel is at least 250. “Non-gray” means the channels are not identical; the dark red marks visible in the exports make clear that these are not purely monochrome drawings. All alpha values are 255, so the white background is opaque. These counts describe file distribution, not contour accuracy, preserved identity or readiness for print. The cat has fewer near-white pixels while still losing important pupil information.
Download either original, open Photo to Line Art, upload it unchanged and wait for a real result before downloading. Compare the cup opening, saucer, spoon, eye interiors and whiskers with the source. Inspect dense marks at full size, then view the drawing at its intended print or display size. Reject confusing edge hierarchies, missing landmarks or open regions that undermine the intended use. This raster export contains no editable vector paths.
Place the measurement script, coffee.png, chelsea.png and both named output PNGs in one folder. With Python, Pillow and NumPy installed, run python measure-artifacts.py to reproduce dimensions, hashes, opacity and threshold counts without rewriting images.
Actual run and patched-runtime fingerprints · Measurements and definitions · Artifact hashes · Source provenance and permissions
REFERENCE
The source may contain strong texture, noise or a complex background.
Yes, after checking that regions are clear and sufficiently enclosed.
The current result is an image; vector tracing is a separate workflow.
Usually yes when the subject is the main purpose of the drawing.
OBSERVED OUTPUT / LINE EXTRACTION
This composite places the photograph on the left and its line interpretation on the right. Because the center split aligns the same face and street scene, it makes preserved contours, omitted color information and unwanted background detail directly inspectable.
EVIDENCE RECORD
REPRODUCE THE CHECK
The glasses, eye, face contour, cat ear and long hair remain recognizable in the line half. Major street shapes also survive, showing that the output is a scene-level edge interpretation rather than a manually isolated portrait drawing.
Highlights, hair texture, signs, pedestrians and blurred street lights create competing marks. Some contours are doubled or broken, and the red-brown line color is part of this output rather than proof of a clean monochrome drawing.
The published 1266 × 1314 artifact is a single aligned RGB composite, and 36.77 percent of its pixels are near-white. That figure describes the large blank regions in this specific output; it is not a universal cleanliness score or a substitute for checking whether the remaining lines are meaningful.
A usable line drawing needs hierarchy. If background lights create darker or denser marks than the face and glasses, the extraction has preserved contrast without preserving intent. Crop, simplify the source or remove background clutter before asking the model for a second pass.
For a coloring page, remove stray background marks and preserve larger closed regions. For tracing reference, retain structural lines but do not assume geometric accuracy. Compare the final drawing at normal size because an image can look detailed while zoomed in and still have weak visual hierarchy.
