String Art Portrait Optimization

A string art portrait asks thousands of straight segments to perform two jobs at once: accumulate tone and preserve a face. A lower global pixel error is not enough if an eye vanishes, a pin overloads, or the simulated darkness cannot be reproduced by actual thread.

This field note builds a feature-aware residual, deterministic chord search, physical pin constraints, human-readable sequence, and thread calibration frame. Likeness, optimization, and labor remain visible as separate parts of the object.

A string art portrait starts with a target in line space

A string portrait approximates tone with many straight thread segments stretched between fixed pins. Dark regions accumulate crossings; light regions receive fewer passes. A string art portrait should begin with a measured grayscale target, not a line-rendered preview that already flatters the method. Crop, lighting, background, facial expression, and edge treatment define what likeness the optimizer is asked to preserve.

Use an image you have rights and consent to transform. Convert it through a declared luminance method, mask the intended circular or polygonal domain, and create separate evaluation crops for eyes, nose, mouth, hair, silhouette, and background. A face-wide mean error can improve while an eye disappears. Preserve the original, normalized target, and every later residual.

The OpenCV color conversion documentation gives official formulas and implementation context for common color transforms. Pin the conversion and working space. Do not let a photo editor's hidden tone curve become an unrecorded part of the optimization recipe.

Lay pins with physical clearance and identity

Place pins around a measured frame with unique integer IDs and known coordinates. Account for pin diameter, frame edge, thread thickness, knot space, and the fact that many passes occupy circumference. Equal angular spacing is simple, but a noncircular frame may need arc-length spacing. Print a full-size drilling template and verify the first, quarter, half, and final pins against the physical boundary.

A pin circle geometry receipt includes pin count, center, radius, start angle, winding order, coordinate units, drill diameter, hardware, and measured placement error. Exclude chords between nearby pins to prevent a dark perimeter and impractical shallow wraps. Also limit repeated use of one pin so thread does not climb off its head.

The SVG 2 coordinate systems specification defines transformations and units for portable templates. Flatten transforms before fabrication, include a dimensional calibration mark, and never assume CSS pixels equal millimeters. The generated guide can be visually correct yet physically wrong if a printer scales to fit.

Runnable artifact: The fixture uses greedy chord selection and proves that a bounded sequence reduces the target residual.

Save this proof as string-art-greedy.test.mjs and run node string-art-greedy.test.mjs. Expected final line: PASS: thread reduces residual.

import assert from "node:assert/strict";
const target=[1,.8,.2,.9], lines=[[.7,.5,0,0],[0,.4,.2,.8],[.4,0,0,.5]];
const error=x=>x.reduce((s,v,i)=>s+(target[i]-v)**2,0); let canvas=[0,0,0,0],before=error(canvas);
for(let k=0;k<3;k++){const choices=lines.map(line=>canvas.map((v,i)=>Math.min(1,v+line[i]*.35)));canvas=choices.sort((a,b)=>error(a)-error(b))[0]}
assert.ok(error(canvas)<before); assert.ok(canvas.every(v=>v>=0&&v<=1)); console.log("PASS: thread reduces residual");

Rasterize every candidate chord consistently

Precompute the pixels or supersampled coverage of every allowed chord. A line contributes opacity based on thread color, thickness, and a simple accumulation model. Store endpoints and sparse coverage so candidate scoring avoids rerasterizing thousands of segments. Validate horizontal, vertical, diagonal, boundary-touching, and reversed chords; endpoint order should not change coverage.

Use a thread line rasterizer tied to output millimeters and evaluation resolution. Anti-aliased screen lines can make one thread appear softer than its material shadow. Compare binary, area-coverage, and measured-thread profiles on a small physical sample. The chosen model does not need perfect optics, but it should rank candidate chords in a way that survives construction.

Canvas dithering also distributes limited marks to approximate tone. String constrains every mark to a full chord and accumulates darkness nonlocally, so local error has global consequences. Keep a uniform-grid control and a blurred target at the expected viewing distance to judge whether the portrait uses crossings as tone rather than incidental line noise.

Choose the next thread by residual reduction

At each pin, score allowed destination pins by how much their chord reduces weighted residual error. Penalize overshooting already dark pixels, repeatedly traversing the same chord, long jumps if material tension demands it, and feature damage in protected light regions. Choose the best admissible move, update the canvas, and continue until improvement falls below a threshold or the thread budget is exhausted.

The runnable fixture uses a tiny greedy solver and proves three chosen lines reduce squared error. A full greedy chord optimization stores every candidate score and selected reduction for replay. Add deterministic tie breaking. Periodically allow a short lookahead or local rerouting, but compare its gain with runtime and construction complexity rather than assuming more search always gives more likeness.

The NetworkX approximation documentation for traveling-salesperson problems is an official reference for route approximation, but this objective differs from shortest tour: repeated pins and tonal residual drive selection. TSP art single-line portraits makes the contrast explicit—there the point placement carries tone and routing seeks continuity; here every chord changes tone.

Weight likeness features without tracing them

Create soft spatial weights around eyes, brows, nostrils, mouth boundary, jaw, and distinctive hairline, then inspect them as a visible map. High weight means errors matter more, not that the optimizer may draw an outline. Keep weights broad enough for tone and crossings to form features. A narrow edge mask can produce a diagrammatic face with harsh threads and little photographic volume.

A portrait residual map should show target minus current darkness, clipped highlights, overshot shadows, and feature-weight contours. Capture it every 250 lines. If global error falls while a landmark region worsens, change the objective or stop. The optimizer's numerical progress must correspond to the reason the image is recognizable.

Run an ablation: uniform weights, face mask only, and feature-aware weights at the same thread count. Ask blinded reviewers to match each output to one of several source portraits and rate naturalness. Archive confusion and uncertainty rather than selecting only flattering comments. Measurement supports art direction when it reveals which abstraction retains identity.

Plan a thread sequence a person can build

The optimized sequence must become instructions. Emit current pin, next pin, cumulative segment, estimated thread length, repeated chord warning, and checkpoints. Mark direction around the frame and provide large, unambiguous pin labels. Break a very long edition into color or spool phases with secure knots located away from key facial highlights.

Model thread tension planning through maximum consecutive span, acute wrap angle, per-pin load, total path length, and local pass count. These are proxies, so build a small frame before scaling. A visually optimal sequence can pull lightweight pins inward or create a bundle too thick to seat. Hardware limits belong in candidate constraints, not in a repair note after construction.

Bézier plotter art with curvature control treats machine motion and physical tolerance as part of image generation. The same principle applies without a plotter: hands, spool, tension, and frame are the renderer. Time a 200-segment trial, note error-prone label transitions, and revise instruction grouping before committing to thousands of wraps.

Compare simulation with measured thread

Photograph a calibration frame containing single, double, fourfold, and eightfold thread crossings under the planned light. Measure their luminance profile and shadow spread. Fit the simple accumulation model or at least choose opacity values that order these cases plausibly. Thread color, sheen, background, and lighting change the apparent density dramatically.

A material thread calibration strip anchors the simulation. Re-render the portrait with the measured line profile and check whether highlight bridges and dark knots move. Then build a cropped region around an eye or mouth at final scale. Compare the physical photograph to the predicted crop with alignment marks, not by memory.

Pen plotter hatching offers useful language for density and line accumulation, although thread floats above a surface and casts shadows. Decide whether the final work is evaluated under fixed exhibition light or expected to change with viewing angle. Document that condition beside the scan or photo used for the edition record.

Publish the portrait as sequence and object

The receipt includes source rights and consent, crop, luminance conversion, masks, feature weights, frame dimensions, pin coordinates and errors, hardware, allowed chord rules, rasterizer, thread profile, objective terms, tie policy, stopping condition, selected sequence, error history, feature metrics, instruction files, build time, thread length, tension observations, lighting, final photograph, accessibility description, and hashes. A string art portrait is both algorithm and labor.

Reject a string art portrait with a broken landmark despite improved average error, pin loads beyond the prototype, instructions a builder cannot follow, hidden manual sequence edits, thread density that closes hardware, or a final photograph whose lighting was tuned to hide structural weakness. Preserve a same-count baseline and the residual map so likeness is not asserted by the solver alone.

Measured thread does not remove intuition. The artist chooses the source, crop, weights, budget, frame, material, and stopping moment; the greedy search reveals how those choices interact over thousands of spans. A credible edition lets the viewer see a face, approach the object and see its straight segments, then inspect a receipt that connects both scales.

String art portrait work must preserve likeness under the constraints of straight thread. Compare each string art portrait at a fixed segment count before praising an optimizer. Calibrate the string art portrait with measured line darkness and pin load. Publish the string art portrait as both an instruction sequence and photographed object.